The talk first introduces the theoretical foundation and technical requirements of immersive communications, which is one of the six major scenarios of 6G. We then analyze the technical challenges in supporting full-sensing holographic immersive communications, including transmission and networking, content capture and reconstruction. Finally, we summarize the research progress in full-sensing holographic immersive communications.

HAPS is a technology that can deliver communications directly from the stratosphere to smartphones on the ground, extending coverage. This presentation introduces our initiatives for early commercialization of HAPS in Japan and our R&D activities for future advancement.

In this paper, we propose an ensemble learning framework for unmanned aerial vehicle(UAV) identification based on radio frequency(RF) signal signatures. While RF signals provide a promising modality for UAV recognition, their reliability significantly degrades in low signal-to-noise ratio(SNR) conditions. To address this challenge, we introduce an SNR-aware feature grouping strategy based on a genetic algorithm, which clusters a set of 20 statistical RF features according to their distributional similarity using Kullback-Leibler divergence. Each feature group is processed by a dedicated one-dimensional convolutional neural network base learner, and the outputs of these learners are combined through a weighted ensemble mechanism, where the weights are optimized during training. Through experimental simulations using real-world RF data, it is demonstrated that the proposed framework significantly outperforms conventional deep learning approaches in terms of accuracy, F1-score and confusion matrix, particularly under both low SNR conditions and realistic environments.

Signal detection under low signal-to-noise ratio (SNR) conditions is a persistent limitation in wireless communication systems. Cyclostationary analysis provides a way to detect modulated signals by extracting periodic spectral features that are distinguishable from stationary noise. The Spectral Correlation Function (SCF) represents these periodicities as two-dimensional matrices indexed by spectral and cyclic frequencies. Conventional SCF-based methods rely on peak detection and require parameter tuning, which leads to unstable performance and high computational cost under noise. This study presents a convolutional neural network (CNN) model named Cyclo-CNN, which treats the SCF matrix as a spatial pattern and learns modulation-specific features without manual thresholding. The Cyclo-CNN model was trained and evaluated on SCF matrices constructed from real-world cellular signal measurements. Results show that the model maintained high classification accuracy across a range of SNR levels. At 18 dB, accuracy reached 94.7%, while at 10 dB and 0 dB, the values were 92.3% and 82.1%, respectively. Under low-SNR conditions, including −10 dB and −20 dB, the model continued to produce meaningful outputs with accuracies of 57.6% and 41.2%. Compared to a traditional SCF-based approach, Cyclo-CNN demonstrated consistently higher accuracy across all tested noise conditions. This approach eliminates the need for manual SCF tuning and performs classification directly from the signal's spectral structure. The method supports application to signal detection tasks under realistic interference and may serve as a foundation for future adaptive classification systems with expanded input representations and improved scalability.

Orthogonal frequency division multiplexing (OFDM) has been widely adopted in modern communication systems. Then, multi-cooperative OFDM employs several relay nodes and improves the system performance by using the packet splitting method. The conventional packet splitting method requires the optimum route weight. However, since the value of the weight is fixed, the optimum results may not be obtained under a varying channel condition. As one approach to solve this problem, the nonlinear processing with deep neural networks (DNN) has attracted attention. Therefore, in this paper, we propose the packet splitting transmission with the DNN-based nonlinear weight for multi-cooperative OFDM systems.

As a new chirp-based multicarrier scheme, affine frequency division multiplexing (AFDM) can fully compensate for performance degradation caused by both propagation delay and the Doppler effect in wireless communications with high mobility. Furthermore, AFDM can be considered more flexible and adaptive than orthogonal time frequency space (OTFS) modulation. This paper considers integrated sensing and communication (ISAC) based on AFDM that can adapt to users' mobility. To realize such ISAC, a channel estimation scheme to cope with fractional Doppler shifts is proposed. After transforming received signals in the discrete affine Fourier transform (DAFT) domain into a pseudo delay-Doppler (DD) domain, the proposed scheme searches for peaks of correlation between received pilot signals and a checking function in the pseudo DD domain, which is referred to as pilot-based peak searching correlation (PSC). In contrast with conventional channel estimation methods for AFDM, the proposed algorithm does not require prior information on the number of propagation paths, which can increase the feasibility of the proposed channel estimation scheme in real-world scenarios. AFDM along with the PSC channel estimation can adaptively control parameters of AFDM, thereby achieving better trade-off between spectral efficiency and bit error rate (BER) than that of OTFS. Computer simulations demonstrate that AFDM employing the proposed channel estimation scheme can adjust the parameters to channel conditions, thereby overcoming OTFS in terms of both adaptability and flexibility.

This paper proposes a novel channel estimation algorithm, termed gradient projection with sequential one-order negative exponential (GP-SOONE), for orthogonal time frequency space (OTFS) systems. While traditional sparse recovery methods often require prior knowledge of channel sparsity, suffer from high computational complexity, or degrade under noise, GP-SOONE overcomes these limitations. By replacing the Gaussian smoothing kernel in the SL0 algorithm with a one-order negative exponential function, GP-SOONE eliminates the need for known sparsity, enhances sparse support recovery and estimation accuracy, and maintains low complexity. Simulations demonstrate that GP-SOONE consistently outperforms baseline methods under integer Doppler scenarios, making it well-suited for real-time applications.

This paper proposes a linear overloaded MIMO system that can transmit more signal streams than the conventional overloaded MIMO spatial multiplexing. In the proposed overloaded MIMO, the number of spatially multiplexed signal streams exceeds the number of transmit and receive antennas. The proposed overloaded MIMO consists of only a nonlinear precoder at the transmitter and a linear spatial filter at the receiver. In particular, the nonlinear filter can be implemented with the same amount of computation as the linear filter, and the system as a whole can be implemented with low computational complexity. We analyze the transmission performance, which is highly dependent on the configuration of the linear spatial filter. Based on the analysis, some examples of the linear spatial filter are picked up, which are expected to achieve excellent transmission performance. We verify the transmission performance of the proposed overloaded MIMO by computer simulation. The computer simulation proves the analysis that the best spatial filter enables the proposed overloaded MIMO to achieve the best transmission performance when four signals are spatially multiplexed in a 2×2 MIMO channel.

The evolution to sixth-generation (6G) wireless systems operating in the centimeter-wave (cmWave) frequency range introduces large-scale multiple-input multiple-output (MIMO) architectures with over 128 antenna ports, resulting in increased energy consumption and substantial challenges for efficient channel state information (CSI) feedback. In the Third Generation Partnership Project (3GPP) New Radio (NR) Release 18, spatial domain network energy saving (NES) is achieved through multiple CSI sub-configurations, each requiring independent feedback. This design incurs considerable signaling overhead and limits flexibility in antenna port pattern adaptation, especially with the increasing number of antenna ports. To address these challenges, a CSI feedback framework based on Kronecker-structured nested codebook is proposed. In this framework, the user equipment (UE) reports a single nested precoding matrix indicator (PMI) composed of modular sub-PMIs, which are combined at the next-generation Node B (gNB) using Kronecker products to reconstruct precoding matrices for various antenna port patterns, such as full, half, and quarter arrays. A simulated annealing (SA) algorithm is introduced for the determination of sub-PMIs, which considers the weighted sum spectral efficiency (SE) across all supported port patterns with the objective of enhancing overall energy efficiency (EE). Compared to the Release 18 baseline with four CSI sub-configurations, the proposed method reduces CSI overhead by approximately 75% and supports more than four antenna port pattern combinations. Simulation results confirm the effectiveness of the proposed framework.

This paper proposes a novel unsourced random access scheme that can consider relative movement between devices and a single base station, which is essentially different from conventional ones. Such movement causes the Doppler effect, which makes it very hard to track fast varying channel state information (CSI) and degrades accuracy of the channel estimation. Thus, the outdated estimate of CSI can heavily damage reliability of the symbol detection. For the purpose of solving this problem, a DopplerFormer model is proposed to infer the average Doppler shift for each user device. Therefore, the time-variation of the CSI can be compensated for, thereby improving the reliability of the symbol detection. Computer simulations demonstrate that the proposed scheme can effectively track time-variant CSI for a large number of mobile users.

This paper proposes a clustering strategy to meet the demands of high throughput in multi-user collaborative uplink MIMO. In the collaborative MIMO systems, coordinated multiple user terminals (UTs) perform spatial multi-stream transmission toward a base station (BS) to enhance the throughput. Although collaborative MIMO transmission provides significant gain over single-user transmission, its performance depends on the configuration of coordinated UTs due to the signaling overhead required for information sharing. Optimal user clustering is essential in this system. This paper proposes a modified K-means based clustering scheme, called collaborative K-means (CKM). K-means generates an arbitrary number of clusters from input data using the centroids of clusters. The proposed K-means algorithm is modified to ensure that each cluster contains at least two UTs, thereby avoiding isolated clusters. This guarantees the acquisition of collaborative MIMO transmission gains. The effectiveness of the proposed scheme is evaluated through computer simulations.

To support systems that require high throughput and low latency, high-frequency bands such as millimeter-wave and terahertz (THz) must be utilized. The use of directional beams is essential when operating in these high-frequency bands. However, beam-based communication requires precise scheduling of beam alignment to prevent path loss, collisions, and deafness issues. Therefore, in this study, we propose a directional Medium Access Control (MAC) protocol based on a Deep Q-Network (DQN) framework for THz-band communication in indoor environments.

Integrated sensing and communication (ISAC) is one of the important capabilities as well as technologies for the sixth generation (6G) mobile communication system. Based on whether base station (BS) or user equipment (UE) acts as sensing transmitter and receiver, six sensing modes could be defined, i.e., BS/UE monostatic sensing and BS-BS/UE-UE/BS-UE/UE-BS bistatic sensing. In 3GPP Release 20 study on ISAC for fifth generation advanced (5G-A) systems, BS monostatic sensing is only supported for unmanned aerial vehicle (UAV) use case due to no impacts on UE. However, towards 6G ISAC with diverse sensing use cases, six sensing modes should all be discussed to determine the suitable scenarios of each sensing mode. In this paper, we analyze and compare the performance of six sensing modes both quantitatively and qualitatively. Firstly, we propose a new metric for system-level analysis of sensing performance. Then, we evaluate the performance of six sensing modes via both link-level and system-level simulations. Finally, we provide our observations and conclusions of six sensing modes based on the performance analysis and comparison, which could provide insights for ISAC system design and deployment.

Open radio access network (O-RAN) introduces a transformative architecture that disaggregates base station functions into three components, namely the radio unit (RU), distributed unit (DU), and centralized unit (CU), enabling flexibility, vendor diversity, and improved interoperability. A core element, the RAN Intelligent Controller (RIC), facilitates intelligent radio resource management through standardized open interfaces. Beyond disaggregation, O-RAN's architecture can also allow efficient spectrum sharing across heterogeneous systems. Despite persistent challenges like service interruption, adjacent-channel interference, and fragmented spectrum, O-RAN's centralized control and programmability can provide robust solutions. This paper introduces a novel spectrum sharing framework utilizing effective bandwidth, accounting for traffic variability and QoS requirements, to ensure mutually beneficial and efficient spectrum access among operators.

High frequency bands such as millimeter and sub-terahertz waves offer wide bandwidths but suffer from severe path-loss and limited delay spread, reducing coverage and degrading cell-edge performance. To address this, we propose a delay diversity enhancement scheme that introduces intentional delay offsets in mobile fronthaul multiplexing (FHM) to improve frequency diversity benefit. Unlike previous studies limited to downlink analysis, our scheme applies to both downlink and uplink and is evaluated using link-level simulations. These simulations reflect practical signal processing, including channel coding, modulation, and frequency-domain equalization, under 3GPP channel models. Numerical results show that the proposed scheme improves throughput by up to 10% in small delay-spread environments by enhancing frequency diversity benefit.

This paper presents TELEscope, an automated security evaluation framework designed for comprehensive assessment of the 5G core network control plane. The transition from traditional telecom protocols to RESTful APIs in Service-Based Architecture (SBA) with HTTP/2-based Service-Based Interfaces (SBI) has introduced new security challenges that existing User Plane-oriented approaches fail to address. TELEscope adopts a discovery-driven methodology to automatically map network topology, analyze SBI vulnerabilities, and execute template-based security scenarios, enabling systematic evaluation of network function management flaws and operational misconfigurations across various deployment environments. Applying TELEscope to Open5GS v2.7.1, one of the most widely used open-source 5G core implementations, exposes critical vulnerabilities in all seven tested scenarios under default configurations, including unauthorized network function registration, exposure of subscriber authentication data, and multiple denial-of-service attack vectors.

Understanding scenes and human states is essential for immersive technologies such as user-centric services and human-computer interaction. Although recent advances in machine learning have greatly improved the robustness of such estimation tasks, many conventional approaches still rely heavily on visual input. This dependence causes a significant drop in performance in environments with poor visibility, such as those affected by darkness, occlusion, or sensor limitations. In this talk, I will introduce approaches that use a variety of sensing methods, including acoustic signals, millimeter-wave radar, and event-based cameras. These techniques enable the estimation of scenes and human states beyond what standard video cameras can perceive. I will also discuss the current challenges and future prospects of developing technologies that overcome the limitations of visual perception.

Smart Cities are powered by millions of interconnected IoT devices, generating vast streams of real-time data that promise unprecedented insights into urban life and resource management. However, this flood of information brings formidable challenges: ensuring data integrity, security, and trust across heterogeneous networks. Traditional centralized systems struggle to meet these demands, exposing critical vulnerabilities and single points of failure. Decentralized Ledger Technology (DLT), or blockchain, offers a transformative approach by providing a shared, immutable, and transparent record of every transaction and data exchange. This talk will share our recent works in integrating IoT networks with diverse blockchain architectures, including practical implementations and rigorous performance evaluations. We will highlight innovative approaches to scalability, one of the key obstacles for IoT-Blockchain systems. Techniques such as intelligent peer selection and advanced sharding provide promising pathways to reduce latency and improve efficiency while maintaining security and transparency at scale.

The rapid proliferation of unmanned aerial vehicles (UAVs) in civilian and industrial sectors has led to an increasing demand for real-time and reliable terminal authentication. Conventional physical-layer authentication techniques often require additional processing steps. Such methods-location-based, hardware-specific, or hybrid-typically incur high communication overhead. Since the physical random access channel (PRACH) signal is the first uplink transmission in the random access procedure between a terminal and the base station (BS), it enables early detection of unauthorized access attempts. Building on this, we propose an innovative authentication framework that leverages the PRACH signal to embed physical-layer fingerprint information into air-to-ground communication. The proposed approach employs an angle estimation algorithm to extract key spatial parameters from the PRACH signal-precisely, elevation angle, azimuth, and signal power-enabling accurate and real-time identification of UAV terminals within designated airspace boundaries. Simulations show that under standard PRACH conditions, the classification accuracy of the proposed machine learning algorithm exceeds 95%. Furthermore, experiments using a standard fifth-generation (5G) communication environment-implemented via OpenAirInterface (OAI) and Universal Software Radio Peripheral (USRP)-confirm the BS's ability to reject UAVs from unauthorized regions during the random access phase.

In the 5G Cloud-RAN (C-RAN), efficient allocation of computing resources across data centers and edge computing environments has become a critical challenge for optimizing network performance. Traditional strategies fail to adapt to the dynamic resource demands of the wireless network nodes, resulting in low resource utilization, load imbalance, and energy waste. In this paper, we propose a dynamic resource prediction and allocation framework based on a deep integration of Heterogeneous Graph Attention Network (HGAT) and Long Short-Term Memory (LSTM) networks. The framework first employs an HGAT to model the topological dependencies among heterogeneous nodes in C-RAN, capturing spatial structural characteristics. Subsequently, we integrate node-type-specific LSTM modules to model long-term temporal evolution trends. Experimental results demonstrate that, across local cloud, edge cloud, and central cloud C-RAN deployment scenarios, the proposed framework significantly outperforms traditional static allocation strategies and existing AI-based approaches. Specifically, it achieves higher resource prediction accuracy across various test scenarios; enhanced load balancing, evidenced by a 6.8%-57.8% reduction in CPU utilization variance; high training efficiency, converging in as few as 5 iterations with fewer than 100 samples; superior inference efficiency relative to other composite models, maintaining stable latency across different network scales; and lower energy consumption, with reductions ranging from 19% to 25%. Moreover, by leveraging these accurate predictions, our framework further enhances system throughput by 7.5%-14.5%. This study may provide a feasible, accurate, and efficient framework for resource prediction and balanced scheduling in dynamic 5G C-RAN environments.

This paper proposes deep learning aided estimation scheme for the polarity of the signal-to-interference power ratio (SIR) using spectrogram images of the received signal. The objective is to enable spectrogram superposition in scenarios where different wireless communication systems, such as LTE and Wireless LAN, share the same spectrum resources. Inter- system interference can be suppressed by blind adaptive array (BAA) known to constant modulus algorithms (CMA) and power inversion (PI), etc. Robust interference suppression is achieved by selecting the algorithm based on the pre-estimated SIR. Com- pared to previously studied IQ-constellation based approaches, spectrogram data provide richer time-frequency representations, which are especially effective for identifying interference from different signal waveforms using convolutional neural network (CNN). Simulation results under coexistence of LTE and Wireless LAN systems show that the proposed approach achieves high binary classification accuracy of SIR, reaching up to 95%. This approach enables the possibility of selecting the most appropriate BAA algorithm according to the interference condition.

In this paper, a cache-enabled device-to-device (D2D) network is investigated, in which an optimization problem on minimizing the long-term energy consumption in content provision is formulated. To alleviate the single-point-of-failure in conventional federated learning (FL) that relies on a single central server, a novel decentralized federated deep reinforcement learning (DFDRL) framework is proposed to coordinate the collaborative content caching and provisioning among proximal devices. By aggregating the local models among distributed devices, higher robustness is achieved in making caching decisions and lower energy is consumed in provisioning contents. Compared to the state-of-the-art methods, the energy consumption is reduced by up to 8.78%.

Millimeter-wave (mmWave) and sub-terahertz (sub-THz) communication systems are promising candidates for enabling high-speed wireless connectivity in unmanned aerial vehicle (UAV) networks. However, the rapid mobility of UAVs leads to frequent beam misalignment. To address this challenge, predicting future beam directions becomes a critical solution. In this paper, we propose a vision-aided beam prediction framework that leverages camera images captured at the base station (BS) to predict future beam indices. Our approach combines a Squeeze-and-Excitation (SE) network with a Long Short-Term Memory (LSTM) network to effectively capture both spatial and temporal dependencies in the data. We evaluate our method using the real-world DeepSense 6G dataset and compare its performance against a GPS-based model. The results demonstrate that the image-based model achieves significantly higher top-k prediction accuracy, validating the benefit of vision-based sensing in dynamic UAV scenarios. This work underscores the potential of predictive, sensor-driven beam tracking to improve reliability and reduce latency in future 6G UAV communication systems.

A neural-assisted decoding framework is proposed to enhance the performance of coded compressed sensing (CCS) in unsourced random access (URA) systems. The method focuses on improving the tree decoding phase by incorporating a machine-learned beam scoring network that selects promising decoding paths based on learned relevance. Message reconstruction is performed iteratively via residual-based peeling, with sparse decoding at each step followed by neural-guided tree search. The beam scorer, trained offline on synthetic data, significantly reduces the number of explored paths by filtering out invalid combinations early in the process. Simulation results demonstrate that the proposed approach achieves near-zero per-user probability of error (PUPE) with up to 4 dB lower signal-to-noise ratio per bit compared to traditional CCS decoding, particularly under high user densities and weak parity structures. The reduction in computational complexity and improved energy efficiency highlight the potential of neural guidance in structured random access decoding.

Considering the satellite Internet of Things (IoT) downlink communication, an efficient multiuser transmission scheme is proposed based on the code-division long range (LoRa) technique. First, a multiuser downlink transmission system model is provided, followed by a pilot-aided carrier synchronization algorithm and a quasi-coherent detection scheme exploring the target performance under extremely low signal-to-noise ratio (SNR), large Doppler shift and time delay. The former can easily estimate the joint offset with respect to the time delay and Doppler shift, and the latter can be implemented by double process of the soft-output noncoherent demodulation and the code correlation decoding. Simulation results show that the proposed multiuser transmission scheme can achieve good performance near to the singleuser transmission.

With the rapid deployment of low Earth orbit (LEO) satellite communication networks, the availability of microwave signals from space has significantly increased. While recent studies have demonstrated the feasibility of opportunistically utilizing these signals to detect and locate a flying object by traditional signal processing methods, research on how to classify it has not been investigated. In this paper, we propose a neural network-based classification method, CNN-BiLSTM-Att, which combines convolutional neural networks (CNNs), bidirectional long-short-term-memory networks (BiLSTM), and attention mechanisms. The model is trained on datasets constructed using the Fresnel-Kirchhoff diffraction model, simulating received signals under various signal-to-noise ratios and object-to-link distances. Experimental results demonstrate that the proposed neural network can effectively distinguish between different categories of flying objects, achieving high classification accuracy with very low false alarm and miss detection rates.

The access demand from low earth orbit (LEO) satellite is surging in the near future, even reaching ten-thousand-level degree. To improve the access capacity of the system, this letter considers that geodesic dome base station (GDBS) accesses multiple LEO satellite networks at the same time. The required beam resources for GDBS full coverage of a certain constellation are first evaluated and a feasible access scheme is proposed. A layered power allocation scheme is proposed to address user decoding failures caused by km-level orbital differences within a single beam. An inter-layer adjustment mechanism is proposed to balance signal amplitude differences at the receiver. Meanwhile, an iterative closed-form solution for optimal power allocation is derived in each layer using slack variables and fractional programming. The proposed algorithm is deployed in a feasible scheme, and simulations indicate it effectively increases the number of successfully accessed LEO satellites compared to traditional methods.

High Altitude Platform Station as IMT Base Station (HIBS) is a key component of the 6G multi-domain fusion network, which can enhance and supplement terrestrial IMT networks, providing mobile communication services for remote and border areas. With the progress of HIBS standardization and the increasing investment from various countries, its application prospects are broad, but cross-border interference issues in border areas have become a critical challenge. This paper focuses on the frequency coordination of HIBS in border areas for spectrum sharing, analyzing the technical parameters of HIBS and terrestrial IMT components, and establishing coexistence scenarios between HIBS and terrestrial IMT systems in neighboring countries' border regions. Through interference coexistence analysis, the compatibility of HIBS and terrestrial cellular networks in candidate frequency bands below 2.7 GHz is evaluated, including interference probability, throughput degradation, and Power Flux Density (PFD) limits. Based on the PFD limits specified in relevant resolutions, the power thresholds of HIBS for bilateral coordination are calculated. The results show that HIBS can meet the corresponding protection criteria in most cases, but mitigation measures are needed under certain distance conditions. This study provides a theoretical basis and practical guidance for HIBS frequency coordination and compatible coexistence in border areas, contributing to the promotion of global communication integration.

This study presents an air-ground integrated network (AGIN) architecture, integrating the free-space optics (FSO)-based high-altitude platform (HAP) for backhaul and unmanned aerial vehicles (UAVs) for last-mile radio frequency (RF) access. Specifically, the HAP equipped with an intelligent reflecting surface (IRS) provides the FSO-based backhaul to multiple UAV-mounted base stations (BSs). To offer high-speed data services over dynamic networks, the objective is to seek the optimal UAV-mounted BS placement to maximize the number of supported mobile users (MUs). To this end, we propose an optimization framework that integrates a convolutional neural network (CNN) with a proximal policy optimization (PPO)-based multi-agent deep reinforcement learning (MADRL), referred to as CP-PPO-based MADRL. Simulation results demonstrate the effectiveness of the proposed framework over the state-of-the-art approaches and support for the proper selection of IRS size.

Existing adaptive Hello messaging schemes dynamically adjust the sending interval of Hello messages based on local topology changes, enabling routing protocols to cope with the dynamic changes of network topology caused by high mobility of unmanned aerial vehicles. However, these schemes typically use fixed empirical thresholds to evaluate whether the local topology changes are drastic or not, which naturally make them not universally applicable to different node speeds. To improve the adaptability of Hello messaging schemes to different degrees of topology changes, this paper proposes a dynamic threshold based adaptive Hello messaging scheme. It dynamically sets the threshold based on the actual degree of local topology changes, so that it can reflect the differences in topology changes caused by different node speeds. The simulation results show that the proposed scheme can achieve higher packet delivery rate. At the same time, the Hello overhead required to increase 1% packet delivery rate is equivalent to those of existing schemes, effectively balancing the two goals of improving network performance and suppressing control overhead.

Intelligent Reflecting Surfaces (IRS) enable the smart configuration of wireless environments via software control of a larger number of programmable passive reflective elements. A multi-IRS system has been shown to exhibit superior performance when compared to a single-IRS system. However, challenges such as reflection optimization, channel estimation, and IRS deployment are becoming increasingly complex. In this paper, we investigate the beamforming optimization problem of a massive multiple-input multiple-output (MIMO) system aided by multi-IRS scenarios. In the scenario, a multi-antenna base station (BS) transmits orthogonal beams toward the user equipment (UE), which are received after being reflected through multiple different IRS along the Line-of-Sight (LoS) link. As the number of reflections increases, the power attenuation of the signal will be greater. To identify the optimal beamforming routing path so that maximize the received signal, an reasonable balance of the path has become the key to solving the problem. To address this issue, first, beamforming is performed for the base station IRS system, and then the problem is approximated as a shortest path problem by combining the obtained channel information and location information, and processed according to the principles of graph theory. Subsequently, apply the deep reinforcement learning algorithm for the IRS to the UE part to determine the optimal beamforming path. The simulation results demonstrate that the outcomes obtained through our algorithm for optimizing the cooperative beamforming of multiple IRS are significantly better than those without optimization or without considering multi-hop, thus confirming the effectiveness of this algorithm.

With the advances in flexibility, wide coverage, and cost-effectiveness, integrating unmanned aerial vehicles (UAVs) with wireless sensor networks (WSNs) has been regarded as a promising technology in data collection and transmission within complex environments. Due to the limited battery capacity of sensor nodes (SNs) and UAVs, UAV-assisted WSNs face challenges in achieving energy efficiency and optimizing resource allocation. To reduce the total energy consumption of UAV-assisted WSNs, we design a multifactor fitness function which includes the distance between SNs, energy levels of SNs, and UAV trajectory length. Based on that, we propose a balanced clustering and cluster head (CH) selection strategy. As for UAV trajectory design, we employ the branch and bound method according to the selection of CHs. Simulation results show that the proposed algorithm reduces the average energy consumption of all SNs and achieves short UAV trajectory length compared to Low-Energy Adaptive Clustering Hierarchy (LEACH), Hybrid Energy-Efficient Distributed clustering (HEED), Particle Swarm Optimization (PSO), and Stable Election Protocol (SEP).

Event cameras, with their high temporal resolution and high dynamic range, can function as visual sensors comparable to conventional image sensors. These characteristics enable event cameras to excel in detecting fast-moving objects and operating in environments with extreme lighting contrasts, such as tunnel exits. As a result, they are gaining attention as a next-generation sensing technology for autonomous vehicles.

In this study, we propose a novel self-localization system by integrating visible light communication (VLC) and visible light positioning (VLP) functionalities into a single event camera. The system uses VLC to obtain coordinate information from transmitters, while VLP estimates the distance to each transmitter. By combining these two types of information, the receiver can determine its own location, even in environments where GPS signals are unavailable, such as tunnels.

The most significant contribution of this study lies in the simultaneous realization of both VLC and VLP functions using a single event camera. Multiple LEDs are installed on the transmitter side, each assigned a unique pilot sequence based on Walsh-Hadamard codes. The event camera uses these codes to calculate the presence probability of each LED within its field of view, allowing for clear separation and identification. This enables high-capacity data transmission via MISO (Multi-Input Single-Output) communication and accurate distance estimation through triangulation using POC (Phase Only Correlation) between multiple LED pairs. To the best of our knowledge, this is the first vehicle-mounted system to achieve simultaneous VLC and VLP using an event camera.

We conducted performance evaluations in real-world conditions by mounting the system on a vehicle traveling at 30 km/h (8.3 m/s). The results showed that the root mean square error (RMSE) of distance estimation was within 0.75 m for ranges up to 100 meters, and the bit error rate (BER) for communication remained below 0.01 across the same range.

Sign language serves as a vital means of communication for individuals who are deaf or hard of hearing; however, Japanese Sign Language (JSL) research remains limited due to the absence of standardized datasets and accessible recognition systems. This study presents a wearable sign language recognition system integrating a custom-designed smart glove with a two-stage XGBoost framework, referred to as Double-XGBoost (DXGB). The glove incorporates five flex sensors and a 6-axis inertial measurement unit to capture eleven channels of fine-grained motion data in real time. A structured data collection framework was established, involving 26 participants-20 experienced signers and 6 beginners-who contributed gesture data for 30 Hiragana-based signs under consistent recording conditions. Experimental results demonstrate that the proposed DXGB model achieves a validation accuracy of 91% and a testing accuracy of 82.5%, outperforming Deep Neural Network (DNN) and Random Forest (RF) baselines. Furthermore, the hierarchical model effectively reduced confusion among visually similar gestures by 7.5%. These results demonstrate the effectiveness and robustness of the proposed wearable-based system for Japanese Sign Language recognition and its potential as a practical assistive communication tool.

In Spatial Crowdsourcing (SC) systems, maintaining data quality and model robustness is essential for reliable AI- driven decision-making in secure smart cities. However, spa- tial data from human contributors often contains biases and anomalies that can negatively impact performance. This paper presents an enhanced trustworthy AI model for SC to detect outliers, mitigate bias, and improve user confidence. This method designs Trust-DBSCAN, a trust-aware clustering algorithm that incorporates behavioural reliability for outlier detection and enables bias-aware training. Experiments with real datasets from Yelp and Foursquare demonstrate that our approach significantly improves outlier detection (F1-score of 0.90), reduces geographic disparity (urban-rural task response ratio from 3.2× to 1.5×), and improves demographic fairness, while maintaining a 94% task completion rate. Our approach surpasses current baselines in accuracy, fairness, and robustness, offering insights for the development of trustworthy AI in smart city applications such as SC platforms.

Problematic smartphone use (PSU) is increasingly recognized as a digital addiction associated with sleep disruption and mental health issues. While prior studies have utilized smartphone sensor logs to estimate short-term usage states, they lack mechanisms for capturing long-term behavioral shifts and explaining their underlying causes in a clinically actionable form. To address this limitation, we propose an automated method for detecting and interpreting weekly changes in smartphone usage trends. Our approach constructs multivariate time-series data from estimated usage states, detects behavioral changes using a lightweight online change point detection technique, identifies key contributing features through sparse regression, and characterizes how and when those features changed using trend segmentation and pattern classification with correction for detection delay. Evaluation on real-world PSU patient data shows that our method detects meaningful behavioral transitions, including those missed by self-reports, and attributes them to interpretable usage patterns. This enables scalable and clinically relevant monitoring of long-term smartphone behavior.

Locating error symbol positions is a necessary step for Vector Symbol Decoding (VSD). The existing method is done algebraically. Deep learning (DL) is proposed to locate the error symbol positions either more accurately (DL-assisted VSD) or more quickly with fixed time (DL-driven VSD). VSD can be used to provide more reliable data transmission in channels with burst errors such as wireless, mobile and power line channels. Using the feature that VSD knows when it fails to decode, DL is proposed to further decode after VSD fails. This leads to more successful decoding. Thus, this DL-assisted VSD always provides the same or better decoding performance than VSD. On the other hand, replacing the algebraic calculation of error-locating vector with DL-driven VSD provides lower decoding performance than VSD, but leads to faster decoding time. Thus, DL-driven VSD can be considered when time is more critical than performance.

Automatic Modulation Recognition (AMR) remains a challenging task in wireless communications. In this paper, we propose a novel heterogeneous dual-channel network with Gated Recurrent Units (GRU) and multi-scale linear attention (MLA) named DGMLA. We utilize GRU to capture the signal dynamic evolution and deep dependencies, the MLA efficiently extracts multi-level cross-scale temporal features from this processed signal representation. Deep Neural Networks (DNNs) are employed to deeply fuse the extracted expert features, enhancing the model's discriminative capability. DGMLA achieves state-of-the-art performance in the AMR task through dual-channel collaborative processing of two different modal inputs. Experiments on the RML2018.10a dataset show that DGMLA achieves an average recognition accuracy of 92.01% at 0 dB to 30 dB and significantly enhances the differentiation between higher-order modulations such as 64QAM, 128QAM, and 256QAM. Comparisons with state-of-the-art models including MCLDNN, PET-CGDNN, DAE, DMCNN, and GRU2 show impressive performance.

Multi-criteria decision-making (MCDM) methods commonly derive criterion weights from subjective preferences or objective data variance. Such approaches do not capture causal interactions among interrelated criteria in complex systems. Relational Impact Scoring (RIS) determines criterion weights by quantifying the functional impact each criterion has on others in achieving a specified outcome. RIS is applied to a cellular network performance dataset for demonstration and validation. Performance comparison against the Rank Order Centroid (ROCSAW) method and the Entropy Weight Method (EWM) shows that RIS generates a unique weighting scheme that separates causal drivers (Distance, Attenuation) from resultant indicators (SNR). Statistical analysis using Spearman's correlation and rank volatility confirms that the RIS ranking differs from benchmark methods while conforming to domain principles. It found that the RIS framework can be implemented as a tool for decision analysis in interconnected systems requiring causal understanding.

As wireless communication networks rapidly evolve, spectrum resources are increasingly scarce, making effective spectrum management critically important. Radio map is a spatial representation of signal characteristics across different locations in a given area, which serves as a key tool for enabling precise spectrum management. To generate accurate radio maps, extensive research efforts have been made. However, most existing studies are conducted on simulation data, which differs significantly from real-world data and cannot accurately reflect the spectrum characteristics of practical environments. To tackle this problem, we construct a dataset of real-world radio map with a self-developed measurement system. Due to the limited volume of real-world data and the distributional discrepancies between simulation and real-world data, we propose a Pixel Transformer (PiT)- based model enhanced with the test-time adaptation (TTA) strategy, named RadioPiT, for real-world radio map generation. Experimental results demonstrate that our proposed RadioPiT significantly outperforms baseline methods in real-world scenarios, yielding a 21.9% decrement in the root mean square error (RMSE) compared to RadioUNet.

In mobile networks, the handover mechanism is crucial for ensuring seamless connectivity as User Equipment (UE) moves across different cells. While the 3rd Generation Partnership Project (3GPP) has continuously improved handover mechanisms from Release 16 to Release 18, the increasing complexity of the wireless environment presents new challenges. From Release 19 onwards, AI/ML has emerged as a transformative technology for mobility management. By harnessing large-scale data analysis, AI/ML facilitates intelligent prediction and adjustment of handover procedure, which drives overall network efficiency. This paper explores the integration of AI/ML techniques in mobility management, with a focus on the requirements outlined in Release 19. The proposed AI/ML-based handover scheme not only enhances the accuracy of handover decision but also reduces measurement and signalling overhead. The key contribution of this research lies in its potential to connect theoretical advancements with practical implementation, paving the way for smarter and more adaptive mobility management, which contributes to the evolution of more resilient and reliable network.

Incentive mechanisms in cross-silo federated learning pose a potential privacy risk owing to the information disclosed by the participants. This study resolves this issue by introducing secure multiparty computation into the distributed algorithm of a previous study. The proposed method enables each organization to perform computations while keeping its bidding information private and cryptographically enhancing privacy, without compromising the economic efficiency of the original mechanism. Simulation evaluations demonstrate that the proposed method converges to an identical equilibrium solution as the original method, and that its computational cost remains within a practical range.

Recently, an increasing number of users are engaging in cryptocurrency transactions, a trend expected to continue. Consequently, major blockchain platforms, such as Bitcoin and Ethereum face scalability issues when the number of transactions exceeds the system's processing capacity. While rollups-Layer 2 solutions-offer a promising approach to address this problem, their modeling and performance evaluation remain insufficiently explored.
In this study, we propose a blockchain queueing model that incorporates rollups. The model can be analyzed using a GI/M/1-type Markov chain, with stability conditions and the stationary distribution through matrix-analytic methods. We calculated performance indicators and conducted numerical experiments to highlight the model's behavior. The model's validity was confirmed through comparison of theoretical and simulated values, while its practical utility was demonstrated using real blockchain data. Our findings suggest that increasing the adoption rate of rollups can significantly improve blockchain scalability and provide quantitative guidance for system design.

Digital identity verification has become crucial to every service in daily life. The privacy concerns associated with traditional Know Your Customer (KYC) systems have come to the forefront. These systems often require the sharing of personal information, which is stored in centralized databases, making them vulnerable to unauthorized access. To address these challenges, this work implements an electronic KYC system with selective disclosure using Merkle Tree and Zero-Knowledge Proofs (ZKP). Selective disclosure enables users to share only the necessary information, thereby reducing the exposure of sensitive data. ZKP enables the verification of this information without revealing the actual data, ensuring that privacy is preserved. The combination of selective disclosure and zkSNARKs in the proposed framework provides a solution for generating a single proof compared to multiple market proofs. This work demonstrates significant improvements in privacy protection compared to traditional identification systems. The implementation process and performance evaluation explore its potential impact on eKYC.

Cardano, a major Proof-of-Stake (PoS) blockchain, aims to achieve decentralization through its reward distribution mechanism governed by the protocol parameter k, which defines the ideal number of reward-receiving stake pools. While k plays a critical role in shaping network dynamics, its current value is statically determined and lacks a formal theoretical basis.

In this study, we conduct an empirical analysis of Cardano's on-chain data to evaluate the effects of changing k on decentralization metrics such as the number of active pools, stake distribution, and the Nakamoto coefficient. Our findings show that while increasing k initially enhances decentralization, the long-term effects are limited due to dynamic shifts in pool performance and stake concentration.

Based on these insights, we propose the foundation for a dynamic optimization framework that determines the appropriate value of k by integrating decentralization, operational cost, and behavioral factors. This work contributes to the development of sustainable and fair PoS governance through data-driven parameter tuning.

Smart contracts are integral to blockchain applications; yet, their immutability creates security vulnerabilities, such as reentrancy and overflow, which can be critically damaging. While detection tools exist, many rely on symbolic execution, graph preprocessing, or binary classification, limiting their efficiency and practicality. This study presents an optimized DeBERTa V3-based transformer model for multi-label detection of smart contract vulnerabilities. The proposed approach operates directly on tokenized Solidity code, leveraging disentangled attention and position embeddings to model semantic patterns. Evaluation on three public datasets achieves up to 100% F1-scores on key vulnerabilities and maintains an average inference latency below 58 ms per smart contract. These results demonstrate the feasibility of integrating high-accuracy, low-latency vulnerability detection into real-time auditing tools, thereby enhancing contract security before deployment.

This paper presents PureTransfer, a novel hash-lock mechanism engineered to solve the pervasive problem of irreversible cryptocurrency transfers to incorrect addresses, which frequently leads to substantial financial losses for users. PureTransfer functions by having the receiver generate a secret preimage (hash1) and its corresponding cryptographic hash (hash2). The sender then locks a specified amount of cryptocurrency on a smart contract using hash2. Only the legitimate receiver, by subsequently revealing hash1, can unlock and claim these funds. Leveraging established cryptographic primitives and smart contract capabilities, PureTransfer shares conceptual similarities with existing Hash Timelock Contracts (HTLCs). However, PureTransfer introduces a key innovation: network-adaptive hash lock mechanisms. This is achieved by embedding network-specific information and transaction metadata directly within hash1 and hash2 used in the HTLCs. This embedding facilitates dynamic adaptation across different networks, intelligent processing, and robust error prevention, significantly enhancing reliability and security. PureTransfer is designed for seamless integration with the PureWallet platform, which supports diverse cryptocurrency and blockchain networks.

Message spoofing and denial-of-service (DoS) attacks threaten vehicular network security by disrupting communication channels and falsifying safety-critical data. Traditional intrusion detection systems (IDS) exhibit high computational overhead and limited adaptability to evolving attack patterns. This paper presents a hybrid security framework integrating Language Agent Models (LAM) with a dual-layer blockchain architecture for real-time threat detection in Internet of Vehicles (IoV) networks. The LAM operates on edge devices to analyze heterogeneous data streams from CAN bus, V2X, and GPS sources. It identifies spoofing and DoS anomalies through transformer-based attention mechanisms with fewer than 1 billion parameters. The dual-layer blockchain combines Proof-of-Authority-and-Association (PoA2) consensus at layer 1 with zero-knowledge rollup (zk-Rollup) at layer 2. This architecture ensures tamper-proof alert logging while reducing on-chain storage overhead. The PoA2 mechanism employs pre-authenticated validators to achieve microsecond-scale transaction finality. The zk-Rollup layer aggregates alert transactions into cryptographic validity proofs, minimizing blockchain storage requirements. Performance evaluation demonstrates the framework's effectiveness across multiple metrics. Detection accuracy reaches 95.7% on the CICIoV2024 dataset and 96.9% on the Car-Hacking dataset. Precision exceeds 97% with F1-scores above 95% on both benchmarks. The system maintains false positive rates below 5.2%. End-to-end response latency remains under 5 milliseconds (ms), meeting real-time safety requirements. The blockchain layer processes over 2,187 transactions per second with 25 validator nodes. Storage optimization achieves a 92% reduction in on-chain data volume. Energy consumption decreases by 4.3 times compared to cloud-hosted language models. The proposed architecture provides deterministic threat detection with cryptographic auditability for large-scale IoV deployments.

As we rapidly move from the ideation phase to the standardization of 6G, it is time to reflect on how this new generation of mobile networks will be shaped by the current moment and how it will potentially influence how we live and work. Every mobile generation to date has brought performance gains, but not every generation has significantly altered the way the network has been used for communications, commerce, and entertainment. Whether 6G represents an evolution or a revolution will depend on how we leverage the key technologies that are being integrated into this generation. More than anything else, Artificial Intelligence (AI) holds enormous promise to revolutionize how the network is managed and used. There are also inherent dualities at play here, in the use of AI to run the network and in AI's dependence on robust network infrastructure, as well as in the potential of AI to improve network security and in the need for new cybersecurity mechanisms to prevent and mitigate attacks on AI. Similar dualities are present in other areas. For example, there is an urgent need to reduce the energy consumed by networks, while at the same time there is great potential for communication networks to increase the efficiency of smart energy grids, transportation, and other verticals. And integrated sensing and communication can potentially introduce new and innovative use cases for mobile networks while raising important privacy concerns. This talk will discuss these dual aspects of 6G technologies and how they can be navigated to make this a truly impactful generation.

Artificial intelligence (AI) will play a prominent role in the next generation of wireless networks, including within the 6G air interface. With the 6G standardization journey just starting, we will begin by encapsulating the lessons learned during the 5G Advanced work and the latest 3GPP updates. We will also provide clarity on the key potentials, achievements, and obstacles that need to be overcome to enable widespread adoption of AI for air interface functionalities. Throughout our presentation, we will share the latest insights and vision from the Ericsson labs. Finally, we will conclude by highlighting the key areas we believe the research community should focus on to make the effective integration of AI in 6G a reality.

This paper proposes a hybrid analog-digital beamforming scheme for cell-free massive multiple-input multipleoutput (CF-mMIMO) systems. Unlike fully-digital schemes, each AP uses a limited number of RF chains and applies analog beamforming under constant modulus constraints. The digital beamforming vectors are optimized via the weighted minimum mean square error (WMMSE) algorithm, while the analog phase shifters are updated using gradient descent. Simulation results show that the proposed scheme achieves sum-rate performance close to the fully-digital case as the number of RF chains increases, demonstrating its efficiency and hardware scalability.

Research on sixth-generation wireless communication systems has gained momentum from the large-scale deployment of 5G networks, thereby increasing the demand for advanced technologies capable of supporting diverse and high-performance application scenarios. Among them, reconfigurable intelligent surfaces (RIS) have emerged as a promising technology to enhance millimeter-wave (mmWave) MIMO systems by adaptively controlling wireless propagation environments. However, channel estimation in RIS-based systems faces challenges due to the passive nature of RIS elements, which cannot actively transmit or sense signals. Existing compressed sensing methods for RIS channel estimation suffer from quantization errors arising from discrete spatial frequency sampling, significantly degrading estimation accuracy. In this paper, we propose a bi-sparse robust simultaneous orthogonal matching pursuit (BiR-SOMP) algorithm that leverages the bi-sparse structure inherent in cascaded RIS channels. By employing Taylor series expansion to construct an extended virtual angular domain codebook with first-order derivatives, the proposed algorithm effectively compensates for quantization errors. Simulation results demonstrate that the BiR-SOMP algorithm significantly outperforms conventional methods in terms of normalized mean square error performance, particularly under limited training overhead and large-scale antenna configurations, highlighting its practical value for future RIS-assisted mmWave MIMO systems.

We present an experimental demonstration of a physical-layer encryption scheme for high-speed optical communication, focusing on cipher block chaining (CBC)-mode symbol-block phase encryption applied to 100 Gbaud dual-polarization quadrature phase-shift keying (DP-QPSK) signals within a 1.6 Tbps optical waveband. In the proposed approach, each of the two 100 Gbaud subchannels is independently encrypted using distinct 128-bit AES keys, yielding a combined key space of 2^256 and ensuring strong security without key interdependence. Unlike conventional bit-level operations, encryption is performed directly on symbol blocks, which increases confidentiality while remaining compatible with high-baud-rate DSP implementations. Experimental transmission over 560 km of standard single-mode fiber with multiple ROADM nodes confirms the feasibility of this scheme. The results show that, although CBC-mode encryption introduces additional phase variations and slightly degrades Q² performance due to stronger interactions with fiber nonlinearities, the Q² values consistently remained above the FEC threshold, demonstrating reliable transmission. Furthermore, the proposed method benefits from practical integration with forward error correction and exhibits tolerance to cascaded ROADM traversals, with only a modest Q² penalty of ~0.9 dB after 10 nodes. These findings indicate that CBC-mode symbol-block encryption can effectively enhance the confidentiality of high-capacity optical networks while maintaining transmission performance, thereby advancing the physical-layer security of next-generation waveband-based optical transport systems.

Uplink coverage enhancement is an important research topic for 6th-generation (6G) mobile communication systems. Discrete Fourier transform spreading orthogonal frequency division multiplexing (DFT-s-OFDM) enhancement with spectrum extension (SE) is one of the most promising research directions for uplink coverage enhancement. There are two types of SE methods: symmetric SE and asymmetric SE. Symmetric SE has low computational complexity but results in high peak-to-average power ratio (PAPR). Asymmetric SE achieves low PAPR but requires high computational complexity. In this paper, a new method to achieve asymmetric SE for DFT-s-OFDM based on frequency-domain cyclic extension (FD-CE) is proposed. The proposed method can achieve low PAPR while requiring low computational complexity simultaneously. Additionally, theoretical analysis has demonstrated the PAPR optimality of asymmetric SE for any SE factor. The enhanced demodulation reference signal (DMRS) sequences based on both enhanced Zadoff-Chu (ZC) sequences and binary phase-shift keying (BPSK) rotation modulated sequences for asymmetric SE are also designed. Compared to NR DMRS with standard ZC sequences, the proposed DMRS schemes achieve PAPR gains of 0.8 dB and 2.5 dB, respectively. The proposed scheme addresses the challenges of high computational complexity and PAPR reduction in DFT-s-OFDM for 6G systems, providing a practical and efficient solution for uplink coverage enhancement.

This paper investigates the performance of an autoencoder (AE)-based receiver for downlink satellite communication systems with power-domain non-orthogonal multiple access (NOMA). To be specific, we consider a scenario where two geostationary Earth orbit (GEO) satellites simultaneously transmit data to a single ground base station (GBS) using different transmission power levels. We further design an end-to-end AE-based receiver architecture to enhance the decoding performance of multi-satellite signals. We evaluate the symbol error rate (SER) performance of the proposed AE-based receiver with conventional decoding techniques such as successive interference cancellation (SIC) or joint maximum likelihood (JML) detection techniques through extensive simulations.

Quantum annealing offers a powerful means to solve combinatorial optimization problems formulated as Quadratic Unconstrained Binary Optimization (QUBO). In constrained problems, appropriate penalty weights are critical: too small yields infeasibility, too large harms solution quality. This paper proposes a rigorous method to determine lower bounds for penalty weights, ensuring constraint satisfaction without degrading optimization. Applied to the Dynamic Spectrum Allocation (DSA) problem with priority-aware extensions, the method leverages iterative quantum annealing to reduce tuning overhead. Experiments show improved efficiency and feasibility over conventional approaches.

We investigate a downlink communication system with multiple low Earth orbit (LEO) satellites, where intersatellite links (ISLs) enable cooperative transmission. To address the difficulty of acquiring instantaneous channel state information (CSI), we exploit statistical CSI and formulate a precoding optimization problem to maximize the ergodic sumrate. To overcome the complexity of ergodic rate computation, we approximate the ergodic rates and propose a weighted minimum mean squared error-based algorithm. Numerical results demonstrate that the proposed method achieves performance close to a benchmark scheme based on perfect CSI and significantly outperforms non-cooperative transmission.

This paper describes an IEEE 802.11ax wireless emulator designed, implemented, and evaluated to achieve both real-time, real-system, large-scale operation, and throughput reproduction accuracy. The wireless emulator was designed to improve the accuracy by operating each virtual wireless device as a pseudo-discrete event simulator, and to facilitate parallel processing by relaxing the consistency requirement between events to maintain them as stochastic events rather than causal events. Based on this design, the wireless emulator was built on Linux and achieved the throughput up to around 1 Gbps with high accuracy in real-time operation on a real Linux system.

This study examines improvements to the Asynchronous Pulse Code Multiple Access protocol designed to facilitate multi-frame message transmission. Additionally, it explores the implementation of a rudimentary error detection mechanism utilizing a dedicated parity-check frame. This design directly addresses the need to handle larger data values efficiently by distributing them across multiple, shorter frames. The inclusion of a parity-check frame provides an effective mechanism to identify and discard corrupted messages, thereby improving the overall message success probability when supporting massive Internet of Things scenarios.

In covert communication, adaptive power control can help balance reliability and covertness, particularly when working with a cooperative jammer. Recent studies considering adaptive power control in covert communications typically assume perfect channel estimation at the transmitter. However, channel estimation is imperfect in practice, resulting in inaccurate power allocation with degraded covertness performance. This paper investigates the effect of channel estimation error on jammer-assisted covert communication. We model channel estimation error in covert communication and analyze its impact on the covertness throughput. Simulation results show that estimation uncertainty leads to degraded throughput compared to perfect channel state information (CSI).

In IEEE 802.11-based wireless LANs, the Distributed Coordination Function (DCF) is employed as a medium access control protocol. Because DCF provides channel access to all contending APs and stations in a fully distributed manner, it is robust for wireless systems operating in unlicensed bands. However, due to its contention-based nature, DCF is not suitable for transmitting traffic with QoS requirements or for efficient use of radio resources. To address this issue in environments with overlapping Basic Service Sets (BSSs), we previously proposed a method that groups BSSs based on their attributes (e.g., supported standards, transmission performance, and required QoS), and assigns different transmission opportunity acquisition periods to each group. In this paper, we extend our proposed method to specifically enhance delay performance for real-time traffic and demonstrate its effectiveness through detailed numerical simulations.

The global internet traffic has kept increasing for decades. To support the realization of future applications such as ultra-immersive communication and remote-control applications in cyber-physical systems, it is necessary to develop wireless communication technologies including wireless LAN (WLAN) capable of transmitting a huge amount of data such as ultra-high-definition video and sensing information. Currently, we are studying new architecture and channel access protocol to realize multi-stream transmission (joint transmission: JT), enabling simultaneous transmission from multiple access points (APs) to a single station over wireless LAN. In order to make WLAN faster, it is better to operate in a high frequency band to use a wide bandwidth, whereas beamforming is necessary in such a band. Therefore, in this paper, we propose an architecture and a procedure for JT including beamforming framework. Through numerical analysis, the goodputs of JT and single-input single-output (SISO) are compared, and it is shown that JT can achieve a higher goodput than SISO.

This study considers video and audio transmission over IEEE 802.11be, which is also known as Wi-Fi 7. One of its distinctive features is MLO (Multi-Link Operation). This enables the wireless device to simultaneously handle two or more links. MLO has a technology called TID (Traffic Identifier)-to-link mapping, which statically bundles TID and a link. This paper investigates adequate mapping for QoE enhancement. We perform a computer simulation of video and audio transmission from terminals to a base station over IEEE 802.11be MLO and then carries out a subjective experiment. We assess the application-level QoS and QoE of upstream video and audio transmission with and without TID-to-link mapping for video and audio streams.

In a dense network where numerous stations (STAs) are connected to a wireless network within a designated area, simultaneous data transmissions by multiple STAs frequently cause packet collisions, leading to reduced communication efficiency and increased transmission delays. In wireless local area network (LAN), which is utilized as a wireless network, carrier sense multiple access with collision avoidance (CSMA/CA) is used to avoid packet collisions. In CSMA/CA, a STA observes whether the channel is idle before transmission; if the channel is idle, it waits for a random duration based on the contention window (CW) before transmitting data. However, if multiple STAs select the same waiting time, they transmit simultaneously, resulting in packet collisions. To reduce the probability of packet collisions, it is effective to set a CW that is appropriate for the environment. Nevertheless, in dense networks, packet collisions cannot be avoided even when the CW is set to its maximum value. Therefore, this paper explores the application of p-persistent CSMA, which introduces a transmission probability, and investigates the optimal control of both CW and transmission probability for the environment. The effectiveness of this approach is demonstrated through computer simulations.

The demand for large-capacity and latency-sensitive applications, such as ultra high-definition video transmission, is increasing in wireless communication systems. When private networks with heavy traffic are deployed in proximity, it becomes challenging to ensure sufficient bandwidth and communication opportunities to support both large-capacity and low-latency traffic. Wireless local area networks (LANs) are capable of operating across multiple frequency bands. Therefore, if each station selects an appropriate frequency band, the impact of interference can be mitigated. In previous work, the authors have focused on video transmission, a representative large-capacity application, and have proposed the concept of video throughput, which incorporates the specific requirements of video delivery. This paper presents results from indoor experiments demonstrating that deep reinforcement learning can effectively enable frequency band selection based on video transmission requirements.

GraphRAG is a cutting-edge technique that combines information retrieval with text generation, and its distinguishing feature lies in the application of graph theory. Capitalizing on this graph-centric approach, our work focuses on graph databases (GDBs) for managing real-world, schema-less data. As a prominent NoSQL database type, GDBs can effectively serve as backends for GraphRAG by storing the knowledge graphs essential for retrieval. However, GDBs face significant scalability challenges, particularly when contrasted with relational or other NoSQL databases. These difficulties stem from the inherent interconnection of graph data and the computational complexity of graph traversals. Addressing these scalability issues, we introduce a general-purpose, decentralized framework designed to optimize the often time-intensive data processing tasks in GDBs. Within this framework, a knowledge graph is constructed from diverse data types, and a query node is integrated for scalable knowledge reasoning via various graph data science techniques. At its core, the framework employs random walk based node embedding algorithms that explicitly consider the relationships (edges) within the knowledge graph. To accelerate graph processing and enhance scalability, we further propose an MPI-based programming workflow for efficient inter-process communication. Evaluation results demonstrate that our approach efficiently improves performance in downstream applications like node classification. Furthermore, this parallelized programming workflow significantly boosts overall data processing performance, especially for large datasets.

Direct-to-Cell (D2C) low-earth orbit (LEO) satellite networks require robust access control. Attribute-based access control (ABAC) has been widely adopted in terrestrial systems owing to its flexible and fine-grained access control. However, it may be challenging to be directly applied to LEO satellite networks due to ABAC's inherent computational demands and severe onboard resource constraints. This paper proposes SATPRUNE, an efficient ABAC scheme for LEO satellite networks, where ABAC policies can be pruned and optimized into lightweight, context-aware subsets based on the anticipated orbital conditions. These pruned policies are then uploaded by the ground-based servers to the satellites for efficient, real-time onboard enforcement. SATPRUNE addresses the complexities of high-dimensional policy spaces and the critical trade-off between security robustness and system performance using a deep reinforcement learning (DRL)-driven two-tier hierarchical pruning architecture. Evaluations demonstrate SATPRUNE significantly reduces policy evaluation latency-by 18% compared to the no-pruning counterpart and 10% over random pruning-while maintaining a false negative rate below 0.1 under attack.

This paper investigates an unmanned robotic system consisting of an unmanned aerial vehicle (UAV), field sensors and actuators. The UAV equipped with communication and computing modules functions as an edge information hub (EIH), which connects the field sensors and actuators into sensing-communication-computing-control (SC3) closed loops. These SC3 closed loops enable network-assembled robots, with the sensor as the eyes, the EIH as the brain, and the actuator as the arm and hand, replacing humans in various hazardous and complex tasks. Given the integrated nature of SC3 closed loops, we propose a joint sensor-actuator pairing and resource allocation scheme. The proposed scheme jointly optimizes pairing decisions along with uplink and downlink (UL&DL) communication and computing resources to minimize the system control cost. To efficiently solve the formulated mixed-integer nonlinear programming (MINLP) problem, we develop a learning-optimization integrated actor-critic (LOAC) framework. Simulation results demonstrate the effectiveness of the proposed scheme, which achieves a well-balanced integration of UL&DL communication and computing for these network-assembled robots in varying environments.

This paper proposes a queue-aware resource allocation framework for Cell-free massive MIMO (CF-mMIMO)- enabled mobile edge computing (MEC) systems. By modeling both local and edge-side task queues, we formulate a long-term energy minimization problem under queue stability constraints. Leveraging Lyapunov optimization, the problem is decomposed into per-slot subproblems: local CPU control is solved in closed form, while a multi-agent proximal policy optimization algorithm learns transmit power and AP clustering. Simulations show that the proposed approach achieves superior energy-delay tradeoff over a heuristic baseline.

Container and VM workloads frequently coexist in modern environments but operate on distinct networking architectures: containers rely on network namespaces, veth pairs, and iptables, whereas VMs use virtual NICs connected to TAP devices and Linux bridges. This disparity results in two fundamental challenges: the absence of native cross-workload service discovery and inefficient communication paths that require manual configuration or traverse multiple redundant layers, incurring latency and processing overhead. We propose ClariNet, an eBPF-based system that accelerates VM-container communication by intercepting packets early in the kernel networking path and establishing direct forwarding channels between endpoints. By maintaining a unified view of container and VM endpoint metadata, ClariNet enables immediate forwarding decisions without service mesh proxies or external registries, reducing packet processing stages from over eight to three. Evaluation on a production mixed VM/container deployment demonstrates that ClariNet improves throughput by 10.4% and significantly lowers latency across all percentiles compared to default networking paths, showing its effectiveness in addressing cross-workload communication inefficiencies.

Modern microservice architectures evolve rapidly, introducing undocumented endpoints and complex inter-service dependencies that expand the attack surface beyond what traditional monitoring systems can cover. Existing solutions, built on static specifications and coarse anomaly thresholds, leave blind spots that adversaries exploit to execute stealthy business logic attacks, leading to undetected breaches, excessive false alerts, and insufficient context for effective incident response. We present Shepherd, a runtime behavior analysis framework that detects suspicious API flows in production environments through adaptive drift detection and a dual-pathway pipeline combining graph attention networks for structural reasoning with transformer-based semantic analysis. Shepherd augments alerts with dual-level explanations highlighting anomalous service interactions and semantic violations, enabling faster and more reliable triage. Evaluation on benchmark systems and real-world traces shows that Shepherd achieves 87-93% detection accuracy with a 7% false positive rate while reducing latency overhead by up to 43% compared to state-of-the-art tools, providing practical and precise API security monitoring for modern microservices.

The trajectory of an unmanned aerial vehicle (UAV) flying in an urban environment is usually restricted to an aerial corridor to guarantee flight safety. The integrated sensing and communication (ISAC) technique facilitates highly accurate localization and reliable communication for UAVs flying along these corridors. In this paper, to improve the data throughput of the ISAC base stations (BSs) serving the UAV, we jointly optimize the flight waypoint, the BS association policy, and the sensing time ratio in each time slot for maximizing the total data throughput of the ISAC BSs communicating with the UAV. To address the nonconvexity associated with trajectory optimization and the constraints on the sensing probability, we propose an iterative algorithm based on alternating optimization (AO) and successive convex approximation (SCA) methods to obtain a sub-optimal solution. Simulation results show that compared with several baseline algorithms, the proposed algorithm can significantly enhance the data throughput of the UAV flying over an urban area.

One of the most promising applications of terahertz (THz) communication lies in wireless data transport networks, such as fronthaul links, due to its potential for high data transmission capacity. In a mesh topology, the fronthaul network is expected to maintain resilience under local downpour conditions due to the availability of multiple link paths. By leveraging flexible path selection and adaptive traffic switching informed by real-time environmental data, such networks can achieve enhanced reliability and seamless data transmission. This simulation-based study investigates the performance of a mesh-based THz fronthaul network, focusing particularly on link switching behavior, signal-to-interference-plus-noise ratio (SINR), and the corresponding modulation and coding scheme (MCS). The simulation results show that fewer than 2.05% of fronthaul links fail to meet the robust MCS, regardless of the bandwidth employed. In contrast, achieving the highest MCS is significantly influenced by both bandwidth and local variations in downpour intensity. Under local downpour conditions, the fronthaul network supports a total traffic capacity ranging from 5.16 terabit per second (Tbps) to 5.71 Tbps, with average data rates between 7.68 gigabit per second (Gbps) and 8.48 Gbps using a 2.16 gigahertz (GHz) bandwidth. When a 69.12 GHz bandwidth is employed, the network supports a total traffic capacity ranging from 111.94 Tbps to 164.43 Tbps, with average data rates between 183.61 Gbps and 243.96 Gbps. Under severe downpour, average data rates degrade by up to 24.73% due to rain attenuation, leading to the selection of more robust MCS.

Reconfigurable intelligent surface (RIS) provides a promising solution by enabling signals to be directly reflected and effectively focused toward the intended receiver, thereby enhancing coverage in specific target areas. However, interference among multiple users remains challenging to eliminate due to the reflection amplitude of each RIS element is typically fixed, degrading the accuracy of beamforming. Meanwhile, obtaining perfect channel state information (CSI) is inherently difficult in RIS-assisted communication links due to the absence of radio frequency chains in RIS elements. In this paper, we consider a RIS-assisted communication system equipped with passive matching networks attached to each RIS element. By leveraging these matching networks, the amplitude of reflected signals from RIS elements can be adaptively adjusted through energy harvesting from the incident waves. The harvested energy is then utilized by the RIS to precisely control the phase of the reflected signals. Upon this characteristic, the minimum data rate of user equipment (UE) is maximized by jointly optimizing the transmit beamforming vector and matching network parameters under imperfect CSI conditions. To mitigate the impact of imperfect CSI, the S-Procedure is utilized to transform the original non-convex design problem into a tractable form. Subsequently, an alternating optimization (AO) algorithm is adopted to address the problem while ensuring the minimum data rate of the UEs. Simulation results demonstrate that matching networks with optimized parameters can significantly improve rate performance.

For mission-critical autonomous operations in remote areas, such as emergency rescue, the sensors and robots rely on an air-based edge information hub (EIH) integrated with communication and computing modules for data analysis, command generation, and information exchange. The sensor, the EIH, and the robot form a sensing-communication-computing-control (SC3) closed loop to accomplish the mission. Due to the inherent openness of air-ground communications, the SC3 closed loop is vulnerable to eavesdropping. In this paper, we utilize physical layer security (PLS) techniques for safeguarding the SC3 closed loop. Considering the task-level coupling between the sensor-to-EIH link and EIH-to-robot link within an SC3 closed loop, we establish a closed-loop security constraint that restricts the leaked task-relevant information below an acceptable threshold. We maximize the closed-loop negentropy (CNE), a new metric that measures the closed-loop performance, by jointly designing the time, power, bandwidth of both wireless links, and the computing capability. The globally optimal solution is analyzed under typical channel conditions. Finally, we provide simulation results to show the superiority of the proposed scheme.

High-density LED arrays enable high-speed transmission in image sensor-based visible light communication (VLC) systems. However, when optical spots become blurred and spatially overlapped due to focal shift, resolution limitations, or interference, severe inter-symbol interference (ISI) occurs, significantly degrading decoding performance. Existing methods mitigate ISI by reducing LED transmission signaling density.

This paper proposes a robust decoding framework that maintains full LED signaling density. We introduce a pilot-aided geometric recognition with a PSF-constrained Hough transform and circle center alignment refinement. By leveraging prior structural knowledge from pilot frames, the system effectively separates overlapping LED signals under severe optical distortion.

Experimental results on a real-world VLC testbed confirm that the proposed method achieves superior decoding accuracy and throughput compared to conventional Hough-based and low-density baseline methods. The results highlight its potential for high-efficiency VLC applications in interference-prone environments.

We present SR-DeepSC-Pro, an advanced super-resolution semantic communication system for bandwidth-constrained satellite image transmission. Through a novel two-stage training algorithm combining curriculum learning and adversarial robustness, our system achieves 10.2% higher PSNR (30.33 dB) while reducing satellite-side parameters by 26% and processing latency by 28.9%. SR-DeepSC-Pro maintains robust performance across SNR conditions with strongest gains at challenging low-SNR levels. This work demonstrates that advanced training strategies can overcome traditional quality-efficiency trade-offs, producing models that simultaneously excel in reconstruction performance and computational efficiency-critical advances for next-generation satellite communication systems.

The rate controlled data transmission becomes important along with the increasing bandwidth of communication links. However, the precise pacing procedures basically require the clock-based timing control and it needs heavy CPU processing costs in commodity computers. There are some proposals which do not rely on high frequency clocks, but they require a dedicated network interface card, or assume a specific type of communication link. The authors formerly proposed a TCP pacing scheme which does not rely on any clock-based timing control, but which uses the ACK reception timings to expand the data segment transmission interval. Although this scheme provided paced data transmission, its performance was lower than the conventional schemes. This paper proposes an improved TCP pacing scheme which uses the ACK-based pacing without any clock-based control and provides high throughput by adding acceleration data segments to the paced data flow. The performance evaluation through the network simulator shows that our new proposal provides high throughput in a high bandwidth and long delay network.

We are studying in-Network Congestion Control (NCC), as a novel congestion control architecture for the increasingly diverse and complex Internet. In NCC, one or more nodes for congestion control (NCC nodes) are introduced on an end-to-end (E2E) network path to divide it into multiple sub-paths, with a congestion-control feedback loop maintained on each sub-path. A specialized congestion control algorithm can be applied to each sub-path according to its network characteristics. One potential issue with NCC is excessive data accumulation at NCC nodes, which is caused by mismatched data transfer rates across sub-paths. To address this issue, we propose a buffering architecture at NCC nodes. Specifically, we control the throughput at the sub-path preceding an NCC node by setting an upper limit on its data buffer and adjusting the amount of data transferred from the receive buffer of the preceding sub-path to the data buffer. This mechanism helps keep the amount of accumulated data at the NCC node small, while not sacrificing the E2E throughput over the NCC node. We evaluated the performance of the proposed method through implementation-based experiments and confirmed that the amount of accumulated data at the NCC node can be kept 1 [Kbytes] during data transmission.

In recent years, the rapid proliferation of mobile devices, such as smartphones and tablets, has led to a dramatic increase in mobile network traffic. In addition, a significant portion of current Internet traffic consists of large-volume content, such as video and audio. However, because the current IP-based network architecture is originally designed for host-to-host communication between specific endpoints, content acquisition is time-consuming. From these backgrounds, Content-Centric Networking (CCN) has attracted wide attention as a new network architecture for efficient content acquisition. In CCN research, CCN-based Mobile Ad-hoc Network (MANET), which refers to the application of a CCN architecture to a MANET, has been studied. An in-network caching scheme for wired CCN is not effectively used in CCN-based MANET because content caching nodes are mobile. To address this problem, this paper proposes cache decision policies using the forwarding information of Interest packets (i.e., content request messages) throughout a network. Through computer simulations, we evaluate the effectiveness of the proposed policies by comparison with existing caching policies in terms of the cache hit rate and content download time.

As cyber threats grow more complex, the need for intelligent and unified analysis across diverse security data sources becomes increasingly critical. To meet this demand, we present Llama-PcapLog, a fine-tuned LLM framework that jointly interprets network packets (pcap) and system logs (syslog), which automates threat detection and scenario analysis. Unlike prior methods that treat pcap and syslog analysis modalities in isolation, Llama-PcapLog captures their temporal and contextual dependencies to detect complex attack patterns. To bridge structural gaps and preserve cross-layer semantics, we fine-tune the open-source LLM by considering pcap and syslog data. For fine-tuning open-source LLM, we preprocess raw pcap and syslog data into an instruction-following format. We also employ a self-instruct strategy to generate diverse, domain-specific tasks across Q&A and code generation. As many institutions or companies need on-premise or private LLM to protect their customers' privacy and data, we fine-tune the Meta Llama-3-8B model on the training dataset and developed a lightweight web interface for interactive analysis. We show that Llama-PcapLog achieves substantial improvements over the base Llama-3-8B with an overall extraction F1 score increase from 0.27 to 0.75. Experiments also show that Llama-PcapLog attains a perfect Pass@k score of 1.00 (vs. 0.45) in the code generation benchmark. These results highlight its effectiveness and potential deployability in real-world cybersecurity workflows.

The InterPlanetary File System (IPFS) relies on Kademlia for efficient content retrieval, but its XOR-based distance metric does not account for physical proximity, leading to suboptimal message latency. This paper proposes an optimized ID allocation method using a Flipped Z-Order curve to improve locality preservation for IPFS. Experimental results demonstrate that our approach reduces message latency and improves GET and PUT operations performance in IPFS, especially under high network loads.

Synthesizing realistic packet arrival processes is essential for a broad range of tasks, including cyberattack reproduction, communication infrastructure design, and network protocol evaluation, especially when access to real traffic data is constrained by privacy, scalability, or security concerns. We propose a generative framework based on a diffusion model, which enables holistic generation of packet arrival sequences while preserving complex temporal structures such as burstiness and long-range dependencies. Unlike recurrent approaches such as NetShare, which generate arrivals sequentially and suffer from cumulative prediction errors, our method generates the entire sequence in one shot. Temporal dependencies are learned using convolutional neural networks (CNNs), and arrival times are refined through iterative addition and deletion of packets. Experiments on real network traces show improved alignment with empirical inter-arrival time distributions, and lower optimal transport distances compared to existing methods. These results underscore the effectiveness of a non-recurrent, diffusion-based approach to traffic synthesis, offering a robust foundation for downstream applications ranging from cyberattack emulation to network protocol design.

In recent years, artificial intelligence has made dramatic advances, fueled by breakthroughs in neural network architectures, scalable compute infrastructure, and the availability of massive datasets. However, the next wave of innovation is increasingly shifting toward on-device AI, where intelligence is brought closer to the data source-whether it be a smartphone, autonomous vehicle, wearable, or sensor node. This keynote traces the evolution of AI architectures from early convolutional neural networks to recent trends such as transformers, efficient neural networks (e.g., MobileNet, EfficientNet), and foundation models. We will explore how these architectures are being adapted and optimized for deployment on resource-constrained edge devices, through techniques such as neural architecture search (NAS), quantization, pruning, and knowledge distillation. Moreover, the talk will address how on-device AI enables real-time inference, privacy preservation, low-latency interaction, and enhanced reliability, all of which are essential for next-generation applications in autonomous systems, healthcare, smart mobility, and wearable computing. This talk aims to provide both a technological roadmap and a strategic vision for researchers and engineers building the next generation of AI-powered edge systems.

This keynote presentation explores the rapidly evolving field of human activity recognition (HAR) using Wi-Fi sensing technologies. The presentation focuses on using channel state information (CSI) and deep learning to detect and classify human behaviors. The presentation begins with an overview of Wi-Fi sensing, emphasizing its advantages as a noninvasive, cost-effective, and accurate HAR approach. Then, it examines key technical and practical challenges, including subject variability, environmental dynamics, and limitations in data quality. Advanced deep learning methodologies are then introduced to address these issues and enhance robustness and accuracy. These include refining feature extraction from Wi-Fi signals, integrating state-of-the-art classification models, and devising strategies to improve generalization across diverse users and environments. Particular attention is given to ensuring resilience under real-world conditions, such as noise and signal degradation. The keynote also addresses practical deployment aspects, focusing on lightweight, resource-efficient models suitable for edge devices. It also explores multi-label classification approaches capable of recognizing multiple concurrent activities. This allows for the capture of the complexity of everyday human behavior. Overall, this presentation aims to advance Wi-Fi-based human activity recognition (HAR) by addressing critical technical barriers and outlining pathways toward scalable, reliable, and real-world applications.

Semantic communication is a new communication paradigm that aims to efficiently convey the "meaning" of information, unlike traditional digital communication. The concept was first proposed by Weaver in 1949 but was long neglected due to technological limitations. In recent years, however, advances in AI technology have led to the practical application of the necessary basic technology, and research is progressing rapidly. Semantic communication is also attracting attention as a promising technology for intelligent applications after 6G. This paper outlines the basic concepts of semantic communication, the technological progress through Generative AI, application examples, and future challenges. This keynote further presents a cutting-edge semantic communication framework tailored for vehicular communication scenarios, where key information is extracted from camera data and transmitted among vehicles and road infrastructure. The keynote will conclude by outlining open challenges and research directions.

The explosive growth of multimedia data, the continuous surge in the number of connected devices, and the increasing demand for real-time intelligent applications are posing unprecedented challenges to current communication infrastructures. Traditional communication systems that transmit raw or compressed data often suffer from excessive latency and bandwidth inefficiency, which can be critical in delay-sensitive applications such as remote driving. To overcome these limitations, semantic communications have recently emerged as a paradigm shift that focuses on transmitting the meaning of data rather than the raw data itself. This talk introduces a novel low-latency video semantic communication framework tailored for remote driving scenarios. In contrast to conventional video transmission methods, the proposed system employs an asymmetric encoder-decoder architecture that transmits only a minimal number of bits by leveraging semantic feature extraction, while reconstructing high-quality video at the receiver through generative AI techniques. To validate its effectiveness, we design and implement a prototype system that seamlessly integrates semantic feature extraction, efficient transmission, and deep learning-based video reconstruction at the receiver side.

To achieve direct communication between satellites and smartphones, a concept has been proposed in which multiple ultra-small satellites equipped with array antenna elements fly in a dense formation, with the whole functioning as a phased array antenna. In this paper, we perform an electromagnetic field analysis of a planar patch antenna array formed by arranging multiple ultra-small satellites and clarified the radiation characteristics of a nonplanar patch antenna array in which the relative positions of multiple ultra-small satellites are slightly shifted in the vertical direction to the antenna plane. As a result, it was shown that the sidelobe and backlobe suppression effect was improved by shifting the vertical position of the array elements located on the diagonal of the square, compared to the conventional planar patch antenna array.

Wireless links in the 300-GHz-band have received significant attention as a promising technology for realizing ultra-high-speed wireless communications. However, the propagation characteristics of 300-GHz-band radio waves in proximity to the human body-particularly in the presence of clothing-remain insufficiently understood. This study investigates the impact of clothing on human body shadowing loss at 300 GHz. Direction-of-arrival (DOA) measurements revealed that wearing undershirt increases the intensity of diffracted waves around the torso by approximately 17.7 dB compared to the unclothed condition. Furthermore, propagation delay profile measurements showed that when wrinkles are present on both sides of the torso, the human body shadowing loss is reduced by approximately 13 dB compared to the case in which the undershirt is tightly adhered to the torso. Additional experiments using a simplified human body model and electromagnetic simulations demonstrated that the shape of wrinkles has a more pronounced effect on the reduction of human body shadowing loss than the presence of air layers between the body and the fabric.

Graph neural networks (GNNs) have emerged as a promising tool for radio resource allocation in device-to-device (D2D) communication networks. However, challenges arise due to variability from training data and heterogeneous communication conditions across links. This paper proposes DTL-D2D, a fully distributed D2D beamforming adaptation framework that combines GNN-based training with distributed transfer learning (DTL). A common GNN model is first trained over multiple D2D network instances to learn parameters optimized on average. The trained model is then distributed to all links, where each link independently adapts its beamformer using only local information from one-hop neighbors to iteratively minimize a loss based on a negatively signed approximation of the sum of local rate estimates. We prove the convergence of the proposed algorithm under a set of assumptions. Numerical results demonstrate that DTL-D2D consistently outperforms various benchmark schemes including non-adaptive GNNs and greedy adaptation.

While Radio Frequency Fingerprint Identification (RFFI) is a well-established field, the complementary challenge of emitter anti-identification remains underexplored. This paper proposes a novel method to camouflage an emitter's identity by manipulating baseband signals to emulate different hardware profiles. The camouflage is achieved by modeling and applying realistic RF impairments, including power amplifier (PA) nonlinearity derived from practical S-parameters, IQ imbalance, and phase noise. To evaluate the efficiency of camouflage, we utilize a robust radio frequency fingerprint (RFF) extraction scheme based on timing recovery, carrier recovery and PA nonlinearity for reliable emitter identification. The overall camouflage and identification system is validated on a hardware platform equipped with integrated RF transceivers AD9361, where the pre-trained classifier demonstrates an overall 99.3% identification accuracy on the original signals but a decreased 40.19% of average classification accuracy on the camouflaged signals, confirming the effectiveness and practicality of the proposed emitter-side camouflage scheme.

We investigate a movable antenna (MA)-enabled multiuser multiple-input single-output (MU-MISO) downlink system. In particular, we propose a deep learning-based algorithm comprising two deep neural networks (DNNs), where each DNN determines either the MA positions or key features of beamforming vectors. These DNNs are jointly trained to maximize the sum-rate performance. The effectiveness of the proposed method is demonstrated through numerical results.

This paper introduces an innovative deep reinforcement learning (DRL)-based framework for jointly optimizing power allocation, CPU frequency, and antenna positioning to reduce user device latency in multi-carrier non-orthogonal multiple access (MC-NOMA) systems integrated with federated learning-assisted fluid antenna systems (FL-FAS). Federated learning (FL) has gained attention as a privacy-preserving and secure approach for large-scale wireless networks, enabling decentralized model training without sharing raw user data while contributing to a global model. Unlike conventional fixed antennas, fluid antenna systems (FAS) feature dynamic, reconfigurable positioning, allowing the system to exploit spatial diversity, mitigate interference, and improve both latency and energy efficiency under varying channel conditions. To tackle the challenges of system latency and FL performance constraints, we formulate a latency-aware optimization problem that incorporates fairness in resource allocation among user devices. A DRL-based solution employing the Proximal Policy Optimization (PPO) algorithm, referred to as FAS-LPPO, is proposed to adaptively respond to time-varying wireless environments while maintaining the integrity of the FL process. Simulation results demonstrate that the proposed framework significantly outperforms comparative optimization methods, without compromising system reliability or learning accuracy.

With the advent of 6G technology, the demand for high-precision Radio Environment Maps (REMs) to support network optimization, resource scheduling, and signal coverage in complex environments has surged. However, due to the limitations of measurement costs and deployment complexity, there is only a small number of discrete radio samples that can currently be obtained, which makes it difficult to support accurate context-aware and location-aware services. To address this challenge, this paper proposes a physics-informed diffusion generative model framework for reconstructing high-resolution REMs in complex physical environments. The framework adopts a U-Net network architecture, achieving precise prediction of radio field strength by predicting noise and minimizing the residual between the Helmholtz gradient terms of the predicted values and the true map. Simulation experiments demonstrate that even in sparse scenarios with a sampling rate as low as 1%, our model achieves a normalized absolute error of 0.0062 and a structural similarity index (SSIM) of 0.8766, fully validating the feasibility and superiority of the model in enabling location- and context-based intelligent wireless services, such as indoor positioning, dynamic spectrum allocation, and environment-adaptive coverage.

This study investigates the effectiveness of machine learning models in predicting student exam performance using physiological data collected from the Empatica E4 wristband. We evaluate multiple data preprocessing strategies to construct three datasets of varying complexity. A suite of machine learning models are applied across these datasets. Additionally, two augmentation techniques are used to address overfitting and class imbalance. Our results reveal a promising combination of model architecture, dataset configuration, and augmentation strategy that yields the highest classification performance among the evaluated setups.

Amodal sensing is critical for various real-world sensing applications because it can recover the complete shapes of partially occluded objects in complex environments. Among various amodal sensing paradigms, wireless amodal sensing is a potential solution due to its advantages of environmental robustness, privacy preservation, and low cost. However, the sensing data obtained by wireless system is sparse for shape reconstruction because of the low spatial resolution, and this issue is further intensified in complex environments with occlusion. To address this issue, we propose a Reconfigurable Intelligent Surface (RIS)-aided wireless amodal sensing scheme that leverages a large-scale RIS to enhance the spatial resolution and create reflection paths that can bypass the obstacles. A generative learning model is also employed to reconstruct the complete shape based on the sensing data captured from the viewpoint of the RIS. In such a system, it is challenging to optimize the RIS phase shifts because the relationship between RIS phase shifts and amodal sensing accuracy is complex and the closed-form expression is unknown. To tackle this challenge, we develop an error prediction model that learns the mapping from RIS phase shifts to amodal sensing accuracy, and optimizes RIS phase shifts based on this mapping. Experimental results on the benchmark dataset show that our method achieves at least a 56.73% reduction in reconstruction error compared to conventional schemes under the same number of RIS configurations.

Modern disaster response and recovery operations face significant challenges due to the complexity of infrastructure networks spread across vast geographical areas. These challenges hinder the timely and informed dispatch of repair and rescue teams, often resulting in increased economic loss and human casualties. This paper explores the enhancement of post-disaster data gathering, processing, and analysis through Internet of Things (IoT)-based solutions. We present the design and evaluation of a comprehensive machine learning pipeline that integrates IoT sensor networks, data prioritization, simulation, and algorithm training. The proposed system includes an infrastructure mapping tool, a simulation engine, a machine learning trainer, and a prediction model. By leveraging real-time data from IoT sensors and aligning it with simulation-driven insights, the framework aims to improve decision-making speed and accuracy in disaster scenarios. Additionally, the paper addresses the limitations of publicly available infrastructure data and proposes automated mapping and live dashboard solutions to support scalable, real-world applications in disaster management.

We propose a neural network-based framework for radar-based heart rate monitoring and inter-beat interval (IBI) estimation that reconstructs the underlying cardiac waveform, rather than relying on Conventional peak detection methods. Conventional approaches typically estimate IBI by detecting peaks in radar signals after extensive signal processing, which can be sensitive to noise and transient signal degradation. Building on our previous work, which utilized a U-Net architecture with cross-channel attention, we introduce a enhanced model by incorporating Bidirectional Long Short-Term Memory (BiLSTM) layers and temporal self-attention into the bottleneck of the U-Net. This modification allows the network to explicitly model temporal dependencies and heartbeat periodicity, improving resilience to transient input degradation and physiological variability. The results are evaluated on a public radar-Electrocardiogram(ECG) dataset under Valsalva, Tilt Up, and Tilt Down conditions, our model achieves an average IBI estimation error of 22.54 ms Root Mean Square Error(RMSE),
which outperforming both conventional peak-based methods and our prior U-Net-based design. These results demonstrate improved temporal generalization and signal stability, advancing the feasibility of radar-based cardiac monitoring in clinical and home settings.

To efficiently manage a large number of IoT devices, peer-to-peer (P2P)-based IoT platforms are gaining attention. In particular, Skip Graph, which supports range queries, is well-suited for managing IoT resources such as devices and data. However, managing time-varying resources requires reregistering appropriate search keys in response to their changes, which may trigger link reconstruction and potentially hinder stable resource retrieval. This paper proposes a Skip Graph-based method that enables the retrieval of time-varying resources without triggering link reconstruction by introducing two types of dynamically tunable range keys: min-fixed and max-fixed range keys. Evaluation through computer simulations suggests that the proposed method can effectively support the retrieval of time-varying resources.

This paper proposes a novel dynamic routing with Software-Defined Networking (SDN) according to network conditions for IoT networks. In this proposal, the genetic algorithm is adopted as the numerical optimization technique to calculate appropriate parameter values for routing. To assess the effectiveness of the proposed method, the authors conduct experiments by simulation. The experimental results demonstrate that the throughput of high-priority traffic is significantly improved. Therefore, the effectiveness of the proposed is confirmed.

We present AI-vPON, an LLM-based framework that enables natural language-driven automation of Passive Optical Network (PON) operations. The system comprises three core components-AI Agent, modelled PON (mPON), and Actual PON-interconnected through a Sim-to-Real transfer pipeline. The AI Agent interprets operator intents and generates multi-step control workflows, pre-validated in mPON before being deployed to Actual PON. We evaluate AI-vPON across four scenarios using six LLMs, including both commercial and open-weight models. Results show that model quality and contextual richness critically impact performance, offering key insights for achieving robust and infrastructure-agnostic PON automation.

In cyber-physical multi-agent systems, ensuring secure and decentralized cooperation among autonomous agents is crucial for maintaining system integrity and operational reliability. However, the presence of malicious or selfish agents in these systems often disrupts message delivery and trust. This paper presents a blockchain-based reputation management framework that integrates adaptive trust scoring, decentralized relay selection, and non-transferable soulbound tokens to incentivize cooperation and build reputation. By combining off-chain trust evaluation with on-chain transparency through the PureChain permissioned blockchain, the framework achieves secure, tamper-proof reputation management. Experimental results show average trust score above 80% under varying proportions of malicious behavior (20% to 80%). The proposed PureChain blockchain also achieved approximately 7.7 times lower transaction latency and 1.4 times higher throughput compared to Ethereum Sepolia, making it suitable for resilient communication in cyber-physical multi-agent systems.

This paper evaluates the effect of time synchronization accuracy, as defined in IEEE 802.1AS, on the performance of QoS(Quality of Service) control specified in IEEE 802.1Q over Ethernet-based in-vehicle networks through experiments. Experimental results show that TAS(Time-Aware Shaper) works properly if the time synchronization accuracy is within one microsecond, and CBS (Credit-Based Shaper) and ATS(Asynchronous Traffic Shaping) work properly if the time synchronization accuracy is within the transmission interval of each traffic.

Cloud computing empowers the analysis of big data streams generated by smart sensors, Internet of Vehicle (IoV), and Internet of Things (IoT), unlocking multiple practical and convenient services. Some famous applications include vehicle traffic loading on roads, video sharing among neighbor vehicles driving, etc. Nevertheless, the realization of Real-Time Active Safe Driving (RT-ASD) in CC-ITS continues to be an open problem, particularly when vehicles encounter unstable driving in high-threat scenarios. To address this challenge, we propose a real-time Predictive backward Shockwave Analysis (PSA) framework for RT-ASD, integrating Macroscopic Traffic Shockwave evaluation with Microscopic Car-following analysis. PSA contributes in three aspects: 1) to predict and analyze high threat backward shockwaves from the gathered big data of IoV, 2) to announce threat messages to the vehicles in high-threat areas via cloud and 3) to determine the threat avoidance mechanism. Numerical results show the proposed PSA outperforms the compared approaches in prediction relative error rate, the accuracy of the backward shockwave determinations, Average Vehicle Velocity (AVV), Average Travel Time (ATT), Time-to-Collision (TTC) and Distance-to-Collision (DTC).

It is widely known that the performance of linear element antennas significantly degrades when placed on metallic surfaces. A commonly adopted solution to this problem involves combining linear element antennas with artificial magnetic conductor (AMC) substrates, which emulate the behavior of a perfect magnetic conductor (PMC) at specific frequencies. By doing so, the reflection phase can be maintained at 0°, thereby enabling stable antenna operation on metallic surfaces. MACKEY proposed in this study is a novel miniaturized antenna that integrates a linear element with an AMC substrate into a compact antenna system. This configuration introduces a unique operating principle not found in conventional designs. The objective of this research is to clarify the advantage of MACKEY by comparing its characteristics with those of microstrip antennas (MSA) that exhibit similar performance in metallic environments.

This paper presents a broadband circularly polarized (CP) magneto-electric (ME) dipole antenna with a simple structure. The proposed CP ME dipole element is a multilayer PCB-based structure, consisting of two dielectric substrates, a top radiating patch, a metal ground plane with an aperture in the middle, and a bottom Y-shaped microstrip feed line. Broadband circular polarization is achieved by using a parallelogram-shaped radiating patch and offset metal vias. The results demonstrate that the antenna achieves an impedance bandwidth of 27.56 GHz to 42.66 GHz and an axial ratio (AR) bandwidth covering 32.74 GHz to 47.79 GHz. With its simple structure, the proposed antenna exhibits wideband CP performance, making it suitable for satellite communication systems.

This study proposes a self-calibration method for array antennas that is essential for the practical implementation of beamforming in next-generation millimeter-wave wireless systems. At such high frequencies, path loss is compensated by coherently phasing the radiating antenna elements of an array to concentrate energy in the desired direction; therefore, accurate beamforming is indispensable. However, in a real beamformer, device-level imperfections and environment-induced temperature drifts continually disturb the relative phase and amplitude of the array elements. Consequently, periodic in-service calibration of the interelement errors is required. The method presented herein performs this calibration entirely within the beamformer, without relying on external measurement equipment, dedicated dummy elements, or additional calibration circuit. By exploiting feedback signals that are naturally available inside the transceiver, the proposed approach adaptively estimates and compensates for element-to-element excitation coefficient mismatches, enabling the desired beam pattern to be maintained at a low cost even under time-varying environmental conditions. Computer simulations demonstrated that the proposed scheme can accurately estimate and correct interelement phase errors within practical measurement tolerances.

This paper presents a substrate-integrated waveguide (SIW) based multiple-input multiple-output (MIMO) antenna operating in the X-band frequency range from 9.4 GHz to 10.5 GHz. The design excited by a coaxial cable feding technique to achieve efficient feeding, wide impedance bandwidth around 12%, and compact planar integration. Simulation results demonstrate a high isolation level exceeding -25 dB between antenna elements, effectively minimizing mutual coupling. The antenna exhibits an average radiation efficiency of approximately 90%, highlighting the low-loss characteristic of the SIW configuration. Moreover, the gain remains above 6.66 dBi across the entire operational band, ensuring satisfactory radiation performance for high-frequency wireless applications. These caracteristics make the proposed antenna as a strong candidate for many applications like radar, satellite communication, and advanced wireless systems requiring compactness, high isolation, and robust performance within the mid X-band spectrum.

In this work, a compact dual-port CPW-fed textile monopole MIMO patch antenna is proposed for biomedical applications operating in the 5.8 GHz ISM band. The antenna integrates split-ring resonator (SRR) based metasurfaces and is fed by a 50-ohm standard coplanar waveguide. The radiating elements are fabricated on a jeans textile substrate, experimentally characterized with a relative permittivity of 1.6 and a thickness of 1 mm. The overall antenna structure, including the SRR configuration and inter-element spacing, is optimized using CST Studio Suite to enhance impedance matching and minimize mutual coupling between elements. The proposed antenna occupies a compact area of 54 mm × 25 mm. The simulated impedance bandwidth ranges from 4.87 GHz to 7.21 GHz, while the measured bandwidth extends from 4.35 GHz to 7 GHz. The antenna demonstrates high isolation with a simulated S₁₂/S₂₁ lower than -15 dB and achieves a peak gain of 7.05 dBi. Furthermore, key MIMO performance metrics such as Envelope Correlation Coefficient (ECC), Channel Capacity Loss (CCL), Diversity Gain (DG), Mean Effective Gain (MEG), and Total Active Reflection Coefficient (TARC) confirm the suitability of the antenna for wearable biomedical devices operating in the 5.8 GHz ISM band. These features highlight the potential of the design for reliable body-centric wireless communication.

This paper presents KeystrokeMouse-TZ, a novel behavioral biometrics dataset for continuous user authentication in internet banking. Collected from 68 Tanzanian participants (60 paid and 8 volunteers), the dataset includes 19,996 keystroke events and 252,397 mouse actions from web-based banking simulations, complemented by pre-, during-, and post-experiment questionnaires. The system, deployed on a live Ubuntu 22.04 LTS server, ensures realistic interactions, with robust preprocessing guaranteeing data quality, anonymity, and reproducibility. Participant feedback indicated 83.6% system recommendation. To evaluate system robustness, synthetic impostor attacks with noise levels from 0.1% to 30% were introduced to the KeystrokeMouse-TZ dataset. eXtreme Gradient Boosting (XGBoost) for Mouse Dynamics achieved high performance (precision: 0.88 - 0.95, recall: 0.79 - 0.87, F1-score: 0.84 - 0.91, accuracy: 0.90 - 0.94), suggesting potential overfitting to synthetic data patterns. Random Forest (RF) for Keystroke Dynamics showed variable performance (accuracy: 0.83 - 0.91, precision: 0.49 - 0.87, F1-score: 0.59 - 0.83), with a notable decline at 10% noise, reflecting sensitivity to class imbalance and noisy data. KeystrokeMouse-TZ fills a critical gap in behavioral biometrics, enabling research on behavioral drift, user profiling, and real-time anomaly detection for adaptive authentication systems.

Physical Unclonable Functions (PUFs) play a critical role in hardware security by providing unique and unclonable identifiers for electronic devices. However, the lack of a standardized evaluation framework hampers objective assessment and comparison among diverse PUF designs. This paper proposes PUFAnalytics, a Python-based library offering a comprehensive and standardized evaluation framework for PUFs. PUFAnalytics features a three-level architecture encompassing response-level, bit-level, and system-level metrics. We establish relationships among various PUF performance metrics, linking them to robustness, unpredictability, error detection, and attack resistance. Experimental results demonstrate the effectiveness of PUFAnalytics in evaluating state-of-the-art PUF implementations, including Arbiter PUFs, SRAM PUFs, Ring Oscillator PUFs, and XOR Arbiter PUFs. Notably, the SRAM PUF exhibited superior reliability, maintaining over 99.5% consistency across temperatures from -40°C to 85°C, and achieved the lowest bit error rate, remaining below 1% under high environmental stress. All PUF designs demonstrated high uniqueness, with inter-PUF Hamming distances centered around 64 bits for 128-bit responses. By providing a standardized evaluation framework, PUFAnalytics addresses the disorganization of evaluation tools in the PUF ecosystem, facilitating fair comparisons and aiding in the development of robust hardware security solutions.

WLAN sensing utilizes CSI to detect human activity without relying on cameras or wearable devices. To our knowledge, dual-band (2.4 GHz + 5 GHz) CSI fusion has not been evaluated for cross-room sensing. In this study, two access points were deployed in adjacent rooms to collect CSI on both the 2.4 GHz and 5 GHz bands, and several classification tasks were conducted using a Random Forest classifier. Results indicate that dual-band fusion outperforms single-band configurations in detecting human presence across rooms.

While payload encryption for privacy protection between full-service resolvers and authoritative DNS servers has been advancing, privacy-sensitive information in communication headers, which remain unencrypted under conventional TLS, remains unaddressed. If the draft-ietf-tls-esni-25 is applied directly in this scenario, the SNI, which is privacy-sensitive information, is exposed during the process of obtaining the public key for encrypting it. Our contribution lies in proposing a novel public key distribution mechanism designed to encrypt privacy-related information contained in communication headers between full-service resolvers and authoritative DNS servers, thereby providing a direction for enhancing privacy protection.

Nowadays, IoT application systems take key roles in various sectors for efficient managements by monitoring environments. Currently, we are developing an integrated IoT application server platform called Smart Environmental Monitoring and Analytical in Real-Time (SEMAR) for fast, efficient deployments of IoT application systems. To assist the setup of SEMAR, the input setup assistance service has been implemented, offering step-by-step guidance using generative AI. To handle unlearned sensors by AI, we adopt Retrieval-Augmented Generation (RAG) to give information from their datasheets. In this paper, we investigate the accuracy of this RAG approach with GPT-4o by using Retrieval-Augmented Generation Assessment (RAGAs), when five newly released sensors in 2024 and 2025 are targeted. The results show 0.89 points on average for faithfulness, 0.68 points for answer relevancy, 0.45 points for context precision, and 0.54 points for context recall, which demonstrate the potential of the proposal while revealing significant challenges in context retrieval and substantial performance variations across sensor types.

In continuous-variable quantum key distribution (CV QKD) systems, a pulse laser is generated from a continuous-wave laser using an intensity modulator, and this pulse is then used for encoding. In this work, we investigate how a CV QKD system implemented using an EML (Electro-absorption Modulated Laser) is affected when operated within a WDM (Wavelength Division Multiplexing) channel environment.