Welcome to IEEE RFID 2026
Radio Frequency Identification is, unsurprisingly, a radio technology. To understand it you need to understand what radio waves are, and how they are generated, propagate, and get detected. In this tutorial we review when radio can be more useful than light for identifying objects. We provide a simplified view of radio and antennas based on vector potentials and current sources to gain an intuitive understanding of what an antenna does and how it works, leading to the Friis equation for quantitation of transmitted and received signals. We examine why you need to modulate a signal, why specialized modulations are useful in passive RFID, and what you lose as a consequence. We introduce the structure of packets used in UHF RFID and Bluetooth Low Energy signaling, followed by a brief examination of how the radio medium is allocated for RFID tags and readers. We then examine the closely-related task of finding something after you've identified it, and how real environments complicate this problem with diffraction, reflection, absorption, and consequent multipath propagation effects.
BeeSMART is a beehive monitoring system that integrates edge nodes with passive and semi-passive UHF RFID sensor tags for scalable, low-power, low-cost sensing across hive clusters. RFID-based temperature, humidity, and hive ID data provide distributed, cluster-level environmental awareness that enables context-aware decision making and enhances local swarm and mite detection. Pilot results show that cluster-level RFID sensing combined with edge-based control stabilizes hive conditions and supports targeted interventions.
The widespread deployment of passive Ultra-High-Frequency (UHF) Radio Frequency Identification (RFID) systems across pharmaceutical serialization, retail logistics, cold-chain monitoring, and critical asset tracking has introduced a large and insufficiently secured attack surface. The stateless architecture and ultra-constrained power budget (1-8 μW) of Class-1 Generation-2 (EPC Gen2) transponders render them inherently vulnerable to a range of adversarial exploits, including tag cloning, relay attacks, inventory replay, phantom tag injection, and Denial-of-Interrogation (DoI). Conventional cryptographic countermeasures-such as challenge-response authentication, hash-locking, and pseudonym randomization-remain impractical under the strict energy and computational limitations imposed by the ISO/IEC 18000-63 standard. Physical-layer security (PLS) techniques, which exploit stochastic variations in backscatter modulation as device-specific fingerprints, provide a promising alternative; however, existing approaches suffer from degraded performance in dense multi-reader environments, sensitivity to inter-session channel variability, and reliance on centralized aggregation of raw in-phase/quadrature (I/Q) data, introducing both privacy and scalability challenges.
This research project presents a software-defined federated learning framework with differential privacy for scalable physical-layer security in UHF RFID systems. The proposed approach eliminates dependence on physical RFID hardware throughout development, training, and evaluation, while addressing the limitations of prior PLS methods through three tightly integrated components: a stochastic backscatter emulation engine, a privacy-preserving federated learning protocol, and a multi-task deep neural architecture.
First, a high-fidelity baseband emulation engine is developed to synthesize EPC Gen2 interrogation sessions in accordance with ISO/IEC 18000-63 timing specifications. The received signal from each virtual tag is modeled as a stochastic convolutional process incorporating a complex-valued backscatter coefficient drawn from IC-family-specific Nakagami-m fading distributions, a Miller-4 subcarrier waveform (Tari = 6.25 μs), a simulated reader antenna impulse response, and additive complex Gaussian noise. The emulator captures variability across multiple commercial integrated circuit families and generates a large-scale dataset comprising 840,000 interrogation sessions across 12,000 virtual tag identities, partitioned into training, validation, and test sets. All data are represented as complex-valued time-series signals, enabling fully reproducible experimentation without reliance on physical readers, antennas, or tags.
Second, to reflect realistic multi-site deployments with strict data locality constraints, a distributed learning framework based on synchronous Federated Averaging (FedAvg) is implemented across 48 simulated edge-reader clients. To ensure formal privacy guarantees, Rényi Differential Privacy (RDP) is incorporated through per-sample gradient clipping (C = 1.0) and calibrated Gaussian noise injection (σ = 0.8), achieving (ε = 2.3, δ = 10⁻⁵)-differential privacy. Each client performs multiple local training epochs per communication round over 300 global rounds, enabling collaborative model optimization while ensuring that raw I/Q fingerprint data remain confined to local partitions. Only privacy-preserving gradient updates are exchanged and aggregated.
Third, a multi-task neural architecture is designed to jointly perform tag identity verification and anomaly detection. The model processes 256-sample I/Q windows corresponding to complete EPC read cycles and employs a depthwise-separable one-dimensional convolutional encoder for hierarchical spectral-temporal feature extraction, followed by a bidirectional gated recurrent unit (GRU) to capture temporal dependencies. The shared representation is fed into two task-specific heads: a 12,000-class softmax classifier for identity recognition and a seven-class sigmoid-based detector covering cloning, relay attacks, DoI, inventory replay, phantom injection, reader impersonation, and bit-level manipulation. Joint optimization is achieved using a weighted composite loss function.
Experimental evaluation under simulated dense-reader interference conditions (up to 16 co-located readers at SNR = 3 dB) demonstrates substantial performance gains. The framework achieves a mean tag identification accuracy of 97.8% (±0.4%, 95% confidence interval), significantly outperforming non-federated deep learning and prior PLS baselines. Anomaly detection attains a macro-averaged true positive rate of 97.8% with a false positive rate of 0.41%, including 99.3% detection accuracy for tag cloning. The incorporation of differential privacy incurs only a modest accuracy reduction relative to non-private federated training, while maintaining strong formal guarantees. Simulated inference latency satisfies EPC Gen2 real-time constraints.
Overall, this research project demonstrates that robust, scalable, and privacy-preserving physical-layer security for UHF RFID systems can be achieved entirely through a software-defined paradigm, establishing a foundation for next-generation secure RFID deployments.
Introduction One of the most significant unsolved problems in the RFID industry is tag localization. That is, how can RFID tags be located quickly and reliably using novel technologies? Traditionally solving this problem has required significant time from experienced radio frequency engineers to develop complex mathematical models. Our goal has been to eliminate that overhead. We use custom neural networks to develop our models instead, trained on data generated in Augmented Reality (AR) to solve localization for modern retail antennas.
Methods When it comes to using neural networks to develop a mathematical model, the problem pivots from requiring talented engineers to requiring reliable training data. We generate this training data by using an Augmented Reality headset to associate real-world tags with their digital twins as a system for ground truth. This ground truth is then paired with a single antenna's movements and signal strength readings (RSSI) as it moves throughout a space. Using this data, we train our neural network to associate tag signal strength with tag position, which can effectively be used to localize tags in a 3-dimensional space.
Results When localizing tags, the user is given a handheld RFID reader and an augmented reality headset. They then proceed to walk around a space while the handheld reader registers tag reads to the headset. Within moments, RFID tags will begin to localize virtually into the user's display. At our current stage of research, the tags localize with an average error of less than half a meter. Discussion & Conclusion Our research is not yet conclusive, and we have many more improvements to make and tests to run. Thus far, we have tested our localization pipeline on one tag inlay type in one enclosed space. We have been careful to abstract the space from the data before processing, but additional environments and tag inlays are imperative for further verification of these techniques. Regardless, using augmented reality to associate real-world tags with their virtual equivalents shows significant promise for the future of tag localization.
Acknowledgement I am very grateful for the guiding hand of Justin Patton, Auburn University's RFID Lab director who gave me the opportunity to create this project. I am additionally thankful for the work done by Jian Zhang and Chris Turner, who inspired the idea. And finally, I would like to thank my colleagues Calvin Hirschler, Adrian Bennett, Sam Clark, and Daniel Speal for their incredible support and contributions.
Radio frequency identification (RFID) has been demonstrated as an increasingly appealing tool for inventory management across many industries. Los Alamos National Laboratory (LANL) aims to increase the efficiency of its current inventory management system by implementing RFID technology. In addition, we aim to leverage RFID infrastructure to also collect data from specific sensors. RFID is planned to be implemented in a scalable, automated manner to make the inventory process more efficient and enable enhanced monitoring capabilities. Real-time data collection is implemented using a microcontroller communicating with a sensor that can transfer sensor data to an integrated RFID chip. The goal is to minimize the size and power intake of the microcontroller. This poster describes the sensor data transfer application and the integration of sensor data transfer over RFID.
Introduction RFID technology is becoming the backbone of modern retail inventory management, yet the physical act of taking inventory remains inefficient due to a lack of visual feedback. Associates typically rely on simple lists and audio cues from handheld scanners without spatial awareness of their scan coverage or accuracy, resulting in missed items and redundant labor. STAR (Scanning Tags in Augmented Reality) addresses these issues by coupling a spatial computing client with a robust web analytics platform to digitize the entire inventory lifecycle. Methods The system is divided into two distinct operational tools. The Augmented Reality client transforms the inventory process by rendering the effects of invisible RF signals as real-time data visualizations. These visualizations are overlaid on the physical environment, which allows associates to easily locate where missing items are in real time. Users may also interact with the current inventory lists using hand-tracking gestures, enabling precise, hands-free viewing of complex inventory categories while maintaining spatial awareness. Complementing the headset, a web-based dashboard can serve as a post-inventory command center. This platform provides clear visibility of cycle count data, such as products scanned, locations and counts, as well as spatial data, such as the position and rotation of the handheld scanner and operator. Results Initial system validation was conducted through a proof-of-concept pilot in a retail clothing store. Preliminary data gathered from the cycle count successfully populated the management dashboard. This data can be reviewed to find coverage patterns, identify training opportunities, and improve the store's ordering process. Discussion & Conclusion By synchronizing the associate's spatial experience with the manager's analytical oversight, STAR establishes a new standard for accuracy and accountability in inventory management. Acknowledgment I would like to thank the Auburn University RFID Laboratory staff for their support in making this project possible. I specifically acknowledge the invaluable guidance of Justin Patton, Daniel Speal, and Harper Rhett. Furthermore, I recognize the foundational hard work of Arpan Srivastava, who kickstarted the project, and the significant technical contributions of Sam Clark.
This paper builds on the idea of computational RFID, describing the RFID DataBridge, a flexible interface that employs the semi-passive EM4325 RFID tag and the EPC Gen 2 protocol to communicate data. The unique capabilities of the RFID DataBridge allow it to support existing processes, provide alternative processes that increase robustness and reduce human interaction around critical instrumentation utilizing existing infrastructure around UHF RFID.
Radiation hardness of radiofrequency identification (RFID) tags is an important consideration in many industrial and research applications including nuclear material container tracking. The tracking of nuclear material containers in glovebox environments presents a unique challenge for utilization of RFID technology for inventory management. In order to effectively select tags for this implementation radiation hardness testing must be performed to evaluate performance and lifetime. This paper outlines the literature review conducted to evaluate vulnerable portions of RFID integrated circuits (ICs) to ionizing radiation and the preliminary results of a radiation hardness study using Ra-226 to supply a cumulative dose of more than 16.5 Gy in Si over the course of over 1,600 hours. The results of the study were null, meaning that none of the tags tested displayed a change in performance throughout the duration of the study.
Passive radio frequency identification depends on the ability to deliver power to the tag with radio waves instead of current on a wire. To accomplish this feat, the tag needs to have an antenna, and circuitry to match the antenna impedance to the circuitry that converts the radio wave to a DC signal. We examine the distinctions between near-field and far-field configurations for delivering power to the tag, and multiple-inductor approaches to stretch the near-field. We review various increasing-complexity schemes for rectifying the RF signal to get DC out, from a single diode to complex charge pumps and voltage boosters. We introduce power-optimized waveforms for improved rectifier performance, and examine how CMOS devices can be exploited to replace classic PN diodes. We end with a brief review of more exotic devices for conversion, and thermodynamic limits for any device.
Passively RFID is the ultimate constrained system due to area, time, cost and function. This tutorial gives an overview of the RFID IC does after it receives power from the antenna. The transistor is mentioned, but the focus is on what the choice matrix is as a circuit designer and user of the ICs, which unwinds the paradox between laboratory behavior and the showroom floor.
BeeSMART is a beehive monitoring system that integrates edge nodes with passive and semi-passive UHF RFID sensor tags for scalable, low-power, low-cost sensing across hive clusters. RFID-based temperature, humidity, and hive ID data provide distributed, cluster-level environmental awareness that enables context-aware decision making and enhances local swarm and mite detection. Pilot results show that cluster-level RFID sensing combined with edge-based control stabilizes hive conditions and supports targeted interventions.
The widespread deployment of passive Ultra-High-Frequency (UHF) Radio Frequency Identification (RFID) systems across pharmaceutical serialization, retail logistics, cold-chain monitoring, and critical asset tracking has introduced a large and insufficiently secured attack surface. The stateless architecture and ultra-constrained power budget (1-8 μW) of Class-1 Generation-2 (EPC Gen2) transponders render them inherently vulnerable to a range of adversarial exploits, including tag cloning, relay attacks, inventory replay, phantom tag injection, and Denial-of-Interrogation (DoI). Conventional cryptographic countermeasures-such as challenge-response authentication, hash-locking, and pseudonym randomization-remain impractical under the strict energy and computational limitations imposed by the ISO/IEC 18000-63 standard. Physical-layer security (PLS) techniques, which exploit stochastic variations in backscatter modulation as device-specific fingerprints, provide a promising alternative; however, existing approaches suffer from degraded performance in dense multi-reader environments, sensitivity to inter-session channel variability, and reliance on centralized aggregation of raw in-phase/quadrature (I/Q) data, introducing both privacy and scalability challenges.
This research project presents a software-defined federated learning framework with differential privacy for scalable physical-layer security in UHF RFID systems. The proposed approach eliminates dependence on physical RFID hardware throughout development, training, and evaluation, while addressing the limitations of prior PLS methods through three tightly integrated components: a stochastic backscatter emulation engine, a privacy-preserving federated learning protocol, and a multi-task deep neural architecture.
First, a high-fidelity baseband emulation engine is developed to synthesize EPC Gen2 interrogation sessions in accordance with ISO/IEC 18000-63 timing specifications. The received signal from each virtual tag is modeled as a stochastic convolutional process incorporating a complex-valued backscatter coefficient drawn from IC-family-specific Nakagami-m fading distributions, a Miller-4 subcarrier waveform (Tari = 6.25 μs), a simulated reader antenna impulse response, and additive complex Gaussian noise. The emulator captures variability across multiple commercial integrated circuit families and generates a large-scale dataset comprising 840,000 interrogation sessions across 12,000 virtual tag identities, partitioned into training, validation, and test sets. All data are represented as complex-valued time-series signals, enabling fully reproducible experimentation without reliance on physical readers, antennas, or tags.
Second, to reflect realistic multi-site deployments with strict data locality constraints, a distributed learning framework based on synchronous Federated Averaging (FedAvg) is implemented across 48 simulated edge-reader clients. To ensure formal privacy guarantees, Rényi Differential Privacy (RDP) is incorporated through per-sample gradient clipping (C = 1.0) and calibrated Gaussian noise injection (σ = 0.8), achieving (ε = 2.3, δ = 10⁻⁵)-differential privacy. Each client performs multiple local training epochs per communication round over 300 global rounds, enabling collaborative model optimization while ensuring that raw I/Q fingerprint data remain confined to local partitions. Only privacy-preserving gradient updates are exchanged and aggregated.
Third, a multi-task neural architecture is designed to jointly perform tag identity verification and anomaly detection. The model processes 256-sample I/Q windows corresponding to complete EPC read cycles and employs a depthwise-separable one-dimensional convolutional encoder for hierarchical spectral-temporal feature extraction, followed by a bidirectional gated recurrent unit (GRU) to capture temporal dependencies. The shared representation is fed into two task-specific heads: a 12,000-class softmax classifier for identity recognition and a seven-class sigmoid-based detector covering cloning, relay attacks, DoI, inventory replay, phantom injection, reader impersonation, and bit-level manipulation. Joint optimization is achieved using a weighted composite loss function.
Experimental evaluation under simulated dense-reader interference conditions (up to 16 co-located readers at SNR = 3 dB) demonstrates substantial performance gains. The framework achieves a mean tag identification accuracy of 97.8% (±0.4%, 95% confidence interval), significantly outperforming non-federated deep learning and prior PLS baselines. Anomaly detection attains a macro-averaged true positive rate of 97.8% with a false positive rate of 0.41%, including 99.3% detection accuracy for tag cloning. The incorporation of differential privacy incurs only a modest accuracy reduction relative to non-private federated training, while maintaining strong formal guarantees. Simulated inference latency satisfies EPC Gen2 real-time constraints.
Overall, this research project demonstrates that robust, scalable, and privacy-preserving physical-layer security for UHF RFID systems can be achieved entirely through a software-defined paradigm, establishing a foundation for next-generation secure RFID deployments.
Introduction One of the most significant unsolved problems in the RFID industry is tag localization. That is, how can RFID tags be located quickly and reliably using novel technologies? Traditionally solving this problem has required significant time from experienced radio frequency engineers to develop complex mathematical models. Our goal has been to eliminate that overhead. We use custom neural networks to develop our models instead, trained on data generated in Augmented Reality (AR) to solve localization for modern retail antennas.
Methods When it comes to using neural networks to develop a mathematical model, the problem pivots from requiring talented engineers to requiring reliable training data. We generate this training data by using an Augmented Reality headset to associate real-world tags with their digital twins as a system for ground truth. This ground truth is then paired with a single antenna's movements and signal strength readings (RSSI) as it moves throughout a space. Using this data, we train our neural network to associate tag signal strength with tag position, which can effectively be used to localize tags in a 3-dimensional space.
Results When localizing tags, the user is given a handheld RFID reader and an augmented reality headset. They then proceed to walk around a space while the handheld reader registers tag reads to the headset. Within moments, RFID tags will begin to localize virtually into the user's display. At our current stage of research, the tags localize with an average error of less than half a meter. Discussion & Conclusion Our research is not yet conclusive, and we have many more improvements to make and tests to run. Thus far, we have tested our localization pipeline on one tag inlay type in one enclosed space. We have been careful to abstract the space from the data before processing, but additional environments and tag inlays are imperative for further verification of these techniques. Regardless, using augmented reality to associate real-world tags with their virtual equivalents shows significant promise for the future of tag localization.
Acknowledgement I am very grateful for the guiding hand of Justin Patton, Auburn University's RFID Lab director who gave me the opportunity to create this project. I am additionally thankful for the work done by Jian Zhang and Chris Turner, who inspired the idea. And finally, I would like to thank my colleagues Calvin Hirschler, Adrian Bennett, Sam Clark, and Daniel Speal for their incredible support and contributions.
Radio frequency identification (RFID) has been demonstrated as an increasingly appealing tool for inventory management across many industries. Los Alamos National Laboratory (LANL) aims to increase the efficiency of its current inventory management system by implementing RFID technology. In addition, we aim to leverage RFID infrastructure to also collect data from specific sensors. RFID is planned to be implemented in a scalable, automated manner to make the inventory process more efficient and enable enhanced monitoring capabilities. Real-time data collection is implemented using a microcontroller communicating with a sensor that can transfer sensor data to an integrated RFID chip. The goal is to minimize the size and power intake of the microcontroller. This poster describes the sensor data transfer application and the integration of sensor data transfer over RFID.
Introduction RFID technology is becoming the backbone of modern retail inventory management, yet the physical act of taking inventory remains inefficient due to a lack of visual feedback. Associates typically rely on simple lists and audio cues from handheld scanners without spatial awareness of their scan coverage or accuracy, resulting in missed items and redundant labor. STAR (Scanning Tags in Augmented Reality) addresses these issues by coupling a spatial computing client with a robust web analytics platform to digitize the entire inventory lifecycle. Methods The system is divided into two distinct operational tools. The Augmented Reality client transforms the inventory process by rendering the effects of invisible RF signals as real-time data visualizations. These visualizations are overlaid on the physical environment, which allows associates to easily locate where missing items are in real time. Users may also interact with the current inventory lists using hand-tracking gestures, enabling precise, hands-free viewing of complex inventory categories while maintaining spatial awareness. Complementing the headset, a web-based dashboard can serve as a post-inventory command center. This platform provides clear visibility of cycle count data, such as products scanned, locations and counts, as well as spatial data, such as the position and rotation of the handheld scanner and operator. Results Initial system validation was conducted through a proof-of-concept pilot in a retail clothing store. Preliminary data gathered from the cycle count successfully populated the management dashboard. This data can be reviewed to find coverage patterns, identify training opportunities, and improve the store's ordering process. Discussion & Conclusion By synchronizing the associate's spatial experience with the manager's analytical oversight, STAR establishes a new standard for accuracy and accountability in inventory management. Acknowledgment I would like to thank the Auburn University RFID Laboratory staff for their support in making this project possible. I specifically acknowledge the invaluable guidance of Justin Patton, Daniel Speal, and Harper Rhett. Furthermore, I recognize the foundational hard work of Arpan Srivastava, who kickstarted the project, and the significant technical contributions of Sam Clark.
This paper builds on the idea of computational RFID, describing the RFID DataBridge, a flexible interface that employs the semi-passive EM4325 RFID tag and the EPC Gen 2 protocol to communicate data. The unique capabilities of the RFID DataBridge allow it to support existing processes, provide alternative processes that increase robustness and reduce human interaction around critical instrumentation utilizing existing infrastructure around UHF RFID.
Radiation hardness of radiofrequency identification (RFID) tags is an important consideration in many industrial and research applications including nuclear material container tracking. The tracking of nuclear material containers in glovebox environments presents a unique challenge for utilization of RFID technology for inventory management. In order to effectively select tags for this implementation radiation hardness testing must be performed to evaluate performance and lifetime. This paper outlines the literature review conducted to evaluate vulnerable portions of RFID integrated circuits (ICs) to ionizing radiation and the preliminary results of a radiation hardness study using Ra-226 to supply a cumulative dose of more than 16.5 Gy in Si over the course of over 1,600 hours. The results of the study were null, meaning that none of the tags tested displayed a change in performance throughout the duration of the study.
We will explore the cost and benefits of cryptography on the RFID tags from both a technical and economic standpoint.
RFID hardware generates data that can be used for plenty of applications, but the interfaces and the data itself may not lend themselves directly to application software. That's where middleware comes in, bridging the gap between hardware and software, and fostering interoperability across vendors and technologies, so that real-time data about "things" in the physical world can be readily applied. This tutorial will present the motivations behind middleware, primarily in the context of RAIN RFID and Bluetooth Low Energy, including practical considerations and real-world examples. Discussion will extend to real-time location, wireless sensing and digital twins to enable digital representations of entire physical spaces. The tutorial will conclude with common applications using event-driven programming, databases, dashboards, AI and all-in-one platforms.
Radiation hardness of radiofrequency identification (RFID) tags is an important consideration in many industrial and research applications including nuclear material container tracking. The tracking of nuclear material containers in glovebox environments presents a unique challenge for utilization of RFID technology for inventory management. In order to effectively select tags for this implementation radiation hardness testing must be performed to evaluate performance and lifetime. This paper outlines the literature review conducted to evaluate vulnerable portions of RFID integrated circuits (ICs) to ionizing radiation and the preliminary results of a radiation hardness study using Ra-226 to supply a cumulative dose of more than 16.5 Gy in Si over the course of over 1,600 hours. The results of the study were null, meaning that none of the tags tested displayed a change in performance throughout the duration of the study.
This paper builds on the idea of computational RFID, describing the RFID DataBridge, a flexible interface that employs the semi-passive EM4325 RFID tag and the EPC Gen 2 protocol to communicate data. The unique capabilities of the RFID DataBridge allow it to support existing processes, provide alternative processes that increase robustness and reduce human interaction around critical instrumentation utilizing existing infrastructure around UHF RFID.
Introduction RFID technology is becoming the backbone of modern retail inventory management, yet the physical act of taking inventory remains inefficient due to a lack of visual feedback. Associates typically rely on simple lists and audio cues from handheld scanners without spatial awareness of their scan coverage or accuracy, resulting in missed items and redundant labor. STAR (Scanning Tags in Augmented Reality) addresses these issues by coupling a spatial computing client with a robust web analytics platform to digitize the entire inventory lifecycle. Methods The system is divided into two distinct operational tools. The Augmented Reality client transforms the inventory process by rendering the effects of invisible RF signals as real-time data visualizations. These visualizations are overlaid on the physical environment, which allows associates to easily locate where missing items are in real time. Users may also interact with the current inventory lists using hand-tracking gestures, enabling precise, hands-free viewing of complex inventory categories while maintaining spatial awareness. Complementing the headset, a web-based dashboard can serve as a post-inventory command center. This platform provides clear visibility of cycle count data, such as products scanned, locations and counts, as well as spatial data, such as the position and rotation of the handheld scanner and operator. Results Initial system validation was conducted through a proof-of-concept pilot in a retail clothing store. Preliminary data gathered from the cycle count successfully populated the management dashboard. This data can be reviewed to find coverage patterns, identify training opportunities, and improve the store's ordering process. Discussion & Conclusion By synchronizing the associate's spatial experience with the manager's analytical oversight, STAR establishes a new standard for accuracy and accountability in inventory management. Acknowledgment I would like to thank the Auburn University RFID Laboratory staff for their support in making this project possible. I specifically acknowledge the invaluable guidance of Justin Patton, Daniel Speal, and Harper Rhett. Furthermore, I recognize the foundational hard work of Arpan Srivastava, who kickstarted the project, and the significant technical contributions of Sam Clark.
Radio frequency identification (RFID) has been demonstrated as an increasingly appealing tool for inventory management across many industries. Los Alamos National Laboratory (LANL) aims to increase the efficiency of its current inventory management system by implementing RFID technology. In addition, we aim to leverage RFID infrastructure to also collect data from specific sensors. RFID is planned to be implemented in a scalable, automated manner to make the inventory process more efficient and enable enhanced monitoring capabilities. Real-time data collection is implemented using a microcontroller communicating with a sensor that can transfer sensor data to an integrated RFID chip. The goal is to minimize the size and power intake of the microcontroller. This poster describes the sensor data transfer application and the integration of sensor data transfer over RFID.
Introduction One of the most significant unsolved problems in the RFID industry is tag localization. That is, how can RFID tags be located quickly and reliably using novel technologies? Traditionally solving this problem has required significant time from experienced radio frequency engineers to develop complex mathematical models. Our goal has been to eliminate that overhead. We use custom neural networks to develop our models instead, trained on data generated in Augmented Reality (AR) to solve localization for modern retail antennas.
Methods When it comes to using neural networks to develop a mathematical model, the problem pivots from requiring talented engineers to requiring reliable training data. We generate this training data by using an Augmented Reality headset to associate real-world tags with their digital twins as a system for ground truth. This ground truth is then paired with a single antenna's movements and signal strength readings (RSSI) as it moves throughout a space. Using this data, we train our neural network to associate tag signal strength with tag position, which can effectively be used to localize tags in a 3-dimensional space.
Results When localizing tags, the user is given a handheld RFID reader and an augmented reality headset. They then proceed to walk around a space while the handheld reader registers tag reads to the headset. Within moments, RFID tags will begin to localize virtually into the user's display. At our current stage of research, the tags localize with an average error of less than half a meter. Discussion & Conclusion Our research is not yet conclusive, and we have many more improvements to make and tests to run. Thus far, we have tested our localization pipeline on one tag inlay type in one enclosed space. We have been careful to abstract the space from the data before processing, but additional environments and tag inlays are imperative for further verification of these techniques. Regardless, using augmented reality to associate real-world tags with their virtual equivalents shows significant promise for the future of tag localization.
Acknowledgement I am very grateful for the guiding hand of Justin Patton, Auburn University's RFID Lab director who gave me the opportunity to create this project. I am additionally thankful for the work done by Jian Zhang and Chris Turner, who inspired the idea. And finally, I would like to thank my colleagues Calvin Hirschler, Adrian Bennett, Sam Clark, and Daniel Speal for their incredible support and contributions.
The widespread deployment of passive Ultra-High-Frequency (UHF) Radio Frequency Identification (RFID) systems across pharmaceutical serialization, retail logistics, cold-chain monitoring, and critical asset tracking has introduced a large and insufficiently secured attack surface. The stateless architecture and ultra-constrained power budget (1-8 μW) of Class-1 Generation-2 (EPC Gen2) transponders render them inherently vulnerable to a range of adversarial exploits, including tag cloning, relay attacks, inventory replay, phantom tag injection, and Denial-of-Interrogation (DoI). Conventional cryptographic countermeasures-such as challenge-response authentication, hash-locking, and pseudonym randomization-remain impractical under the strict energy and computational limitations imposed by the ISO/IEC 18000-63 standard. Physical-layer security (PLS) techniques, which exploit stochastic variations in backscatter modulation as device-specific fingerprints, provide a promising alternative; however, existing approaches suffer from degraded performance in dense multi-reader environments, sensitivity to inter-session channel variability, and reliance on centralized aggregation of raw in-phase/quadrature (I/Q) data, introducing both privacy and scalability challenges.
This research project presents a software-defined federated learning framework with differential privacy for scalable physical-layer security in UHF RFID systems. The proposed approach eliminates dependence on physical RFID hardware throughout development, training, and evaluation, while addressing the limitations of prior PLS methods through three tightly integrated components: a stochastic backscatter emulation engine, a privacy-preserving federated learning protocol, and a multi-task deep neural architecture.
First, a high-fidelity baseband emulation engine is developed to synthesize EPC Gen2 interrogation sessions in accordance with ISO/IEC 18000-63 timing specifications. The received signal from each virtual tag is modeled as a stochastic convolutional process incorporating a complex-valued backscatter coefficient drawn from IC-family-specific Nakagami-m fading distributions, a Miller-4 subcarrier waveform (Tari = 6.25 μs), a simulated reader antenna impulse response, and additive complex Gaussian noise. The emulator captures variability across multiple commercial integrated circuit families and generates a large-scale dataset comprising 840,000 interrogation sessions across 12,000 virtual tag identities, partitioned into training, validation, and test sets. All data are represented as complex-valued time-series signals, enabling fully reproducible experimentation without reliance on physical readers, antennas, or tags.
Second, to reflect realistic multi-site deployments with strict data locality constraints, a distributed learning framework based on synchronous Federated Averaging (FedAvg) is implemented across 48 simulated edge-reader clients. To ensure formal privacy guarantees, Rényi Differential Privacy (RDP) is incorporated through per-sample gradient clipping (C = 1.0) and calibrated Gaussian noise injection (σ = 0.8), achieving (ε = 2.3, δ = 10⁻⁵)-differential privacy. Each client performs multiple local training epochs per communication round over 300 global rounds, enabling collaborative model optimization while ensuring that raw I/Q fingerprint data remain confined to local partitions. Only privacy-preserving gradient updates are exchanged and aggregated.
Third, a multi-task neural architecture is designed to jointly perform tag identity verification and anomaly detection. The model processes 256-sample I/Q windows corresponding to complete EPC read cycles and employs a depthwise-separable one-dimensional convolutional encoder for hierarchical spectral-temporal feature extraction, followed by a bidirectional gated recurrent unit (GRU) to capture temporal dependencies. The shared representation is fed into two task-specific heads: a 12,000-class softmax classifier for identity recognition and a seven-class sigmoid-based detector covering cloning, relay attacks, DoI, inventory replay, phantom injection, reader impersonation, and bit-level manipulation. Joint optimization is achieved using a weighted composite loss function.
Experimental evaluation under simulated dense-reader interference conditions (up to 16 co-located readers at SNR = 3 dB) demonstrates substantial performance gains. The framework achieves a mean tag identification accuracy of 97.8% (±0.4%, 95% confidence interval), significantly outperforming non-federated deep learning and prior PLS baselines. Anomaly detection attains a macro-averaged true positive rate of 97.8% with a false positive rate of 0.41%, including 99.3% detection accuracy for tag cloning. The incorporation of differential privacy incurs only a modest accuracy reduction relative to non-private federated training, while maintaining strong formal guarantees. Simulated inference latency satisfies EPC Gen2 real-time constraints.
Overall, this research project demonstrates that robust, scalable, and privacy-preserving physical-layer security for UHF RFID systems can be achieved entirely through a software-defined paradigm, establishing a foundation for next-generation secure RFID deployments.
BeeSMART is a beehive monitoring system that integrates edge nodes with passive and semi-passive UHF RFID sensor tags for scalable, low-power, low-cost sensing across hive clusters. RFID-based temperature, humidity, and hive ID data provide distributed, cluster-level environmental awareness that enables context-aware decision making and enhances local swarm and mite detection. Pilot results show that cluster-level RFID sensing combined with edge-based control stabilizes hive conditions and supports targeted interventions.
Tires: The U.S. Federal Motor Carrier Safety Administration is trying to determine if passive RFID Tire Pressure Management Systems are a viable alternative to powered solutions. A brief discussion of the background, current options, and intended path for the testing of RAIN RFID based sensing will be presented, allowing opportunities for the audience to not only provide thoughts and suggestions, but to keep informed on the progress towards a working proof of concept.
PLCs: Every use case for RFID implicitly, if not explicitly, acknowledges the need to have RFID tagging occur as far up the supply chain as possible allowing for users to gain maximum benefit from its use. However, the stated intentions often fail to reflect reality, with company after company still incorporating a variant of "slap and ship." Thanks to Zebra's latest firmware, the barriers to incorporating RFID in the manufacturing environment have been significantly reduced, allowing for seamless integration of RFID into automation solutions. Join us as we explore the reality of integrating RFID into automation solutions and demonstrate some ways in which RFID is able to better interact with machine automation.
Discussing the latest advances and potential for RFID beyond conventional CMOS.
Each year for the past two decades, pundits have announced that this is the year radio frequency identification takes off, and each year, the pundits have been wrong. Mandates from Walmart didn't spur mass adoption. Regulations requiring the tracking of unique doses of medicines didn't get companies to deploy RFID. Massive improvements in the performance of passive UHF RFID systems and dramatically lower prices for tags and readers hasn't led to mass adoption-even in the apparel retail industry, which has led all sectors in deploying the technology.
So, what's preventing RFID from becoming as ubiquitous as the bar code?
Mark Roberti covered RFID technology for 25 years, first as the founder and editor of RFID Journal, and more recently as president of the RFID Professional Institute. He has spoken with hundreds of companies interested in deploying RFID, with solution providers creating products for the market and researchers in the labs. In this presentation, he will share his thoughts on why RFID hasn't reached mass adoption, what needs to be done to accelerate the current growth of RFID and insights into the market opportunities currently being overlooked.
This paper presents an application of an unmodified Raspberry Pi (RPi) Pico as a low-power and low-cost, Bluetooth Low Energy (BLE) compatible backscatter data uplink. A ZeroScatter approach requiring zero external components is made possible by the two distinct impedance states inherent in switching between the input and output states of the general purpose input/output (GPIO) pins on the RPi Pico. The RPi Pico programmable input/output (PIO) state machine block is first configured in software to accept a binary bit-string containing a BLE advertisement packet. It then generates a frequency-shift keying (FSK) modulation using 2 MHz and 3 MHz subcarriers, representing data bits ‘0' and ‘1', respectively, to backscatter BLE advertising packets when the Pico is illuminated by a 2423.5 MHz carrier from an external carrier wave (CW) source.
During experimental validation, an unmodified Apple iOS device is used as a BLE receiver using a commercial BLE scanner app. We validated the BLE packet performance over-the-air by measuring the Received Signal Strength Indicator (RSSI) in a bistatic configuration at a fixed separation distance of 1.6 meters.
This paper presents an application specific integrated circuit (ASIC) that implements Bluetooth Low Energy (BLE) advertising transmission with an all-digital backscatter approach. The ASIC toggles a standard digital I/O pin between tri-state and ground. These two logic states yield different RF impedances, which creates backscatter modulation at programmable subcarrier frequencies. An incident 2.4 GHz carrier wave is then backscatter modulated with these subcarrier frequencies to transmit 1 Mbps BLE formatted data in a BLE advertising channel. The ASIC logic is synthesized from Verilog RTL using a standard cell library in a TSMC 180 nm CMOS process. The ASIC occupies an active area of only 0.085 mm2. This ASIC was validated by transmitting in BLE Channel 37, centered on 2402 MHz. Operating with a 12 MHz external clock, the ASIC consumed 1.85 mW of power. This was split between 1.25 mW (67.5%) used for I/O and 601 uW (32.5%) used for digital core logic. With a carrier wave power of +28 dBm and a 3 cm antenna, the ASIC achieved a transmit range of up to 1 m, as received by an unmodified iPhone without any software or firmware modifications.
This paper shows that it is possible to estimate the impedance of a remote antenna connected to a UHF RFID chip. The proposed method relies on the measurement of four backscattered field values and the knowledge of only two load impedance values. The field values can be obtained using the autotuning functionality available in modern UHF RFID chips. The method does not make any assumption about the antenna geometry and can be performed multiple times to obtain the impedance of the antenna over a given bandwidth (or any other parameter). More importantly, the method is not limited by the number of autotune capacitors in the chip. The method allows one to directly transform a UHF RFID tag into a sensor if the relation between the antenna impedance and a physical quantity is known. Thanks to the proposed method, a significant part of the UHF RFID tags already deployed in the field can be transformed into sensors.
Photosensitive RFID tags enable optical sensing applications while preserving the passive and low-cost nature. Conventional solutions rely on off-chip photo detectors, which increase footprint and cost, whereas on-chip optical detection is challenging under standard flip-chip packaging. This paper presents a fully integrated photosensitive RFID tag with backside illumination for optical threshold detection. The backside illumination scheme avoids the front-side optical blockage inherent to flip-chip packaging and ensures direct on-chip photosensing compatibility. But it suffers from significant optical attenuation in silicon. Therefore, the photosensor employs an active pixel that adopts a CMOS-compatible p+/n-well/p-sub photodiode with a large, deep depletion region for efficient charge collection, and a subthreshold capacitive transimpedance amplifier (CTIA) for high conversion gain. A subsequent asynchronous dynamic voltage detector provides threshold judgment with negligible static power. Fabricated in 130-nm CMOS, the on-chip sensor core occupies 0.007 mm2 and consumes 34.7 nW from a 0.8-V supply. Measurement results show that the photosensitive RFID tag achieves 6-bit programmable optical threshold detection across an illuminance range from near darkness to approximately 2200 lux.
This paper presents an Always-ON batteryless RFID sensor label platform that enables continuous environmental monitoring. Unlike conventional RFID sensor tags that operate only within the RF reader field or rely on batteries or auxiliary energy harvesters, the proposed sensor performs continuous sensing and data logging outside the RF field for periods of up to several weeks. The system is enabled by arrays of digital MEMS sensor switches that extract the energy required for switching directly from the sensed physical parameter, together with ultra-energy-efficient single-transistor MEMS-CMOS non-volatile memory (NVM). By eliminating continuous biasing, analog front-ends, and sampled acquisition, the energy required for sensing and digital data storage is reduced by approximately 100-1000x, allowing extended operation using only on-chip isolated stored charge. Environmental data are stored on-chip and retrieved during subsequent RFID interrogation. The demonstrated sensor arrays support temperature and humidity sensing, with operating ranges of −10 °C to 80 °C and 20%-90% relative humidity, respectively. A batteryless NFC-based RFID temperature sensor label with multi-level digital output for vial temperature monitoring is demonstrated. The proposed architecture enables compact, low-cost smart sensor labels compatible with existing RFID infrastructure and is well suited for supply-chain and cold-chain IoT applications.
Accurate real-time tracking of mobile RFID tags with velocity estimation remains challenging due to phase noise, multipath interference, Doppler effects, and strict latency constraints. While classical Long Short-Term Memory (LSTM) models can capture temporal correlations in RFID measurements, they often struggle with nonlinear phase dynamics and phase-velocity coupling inherent to RF propagation. This paper presents a Quantum-Enhanced LSTM (QLSTM), a hybrid quantum-classical architecture for phase-based RFID tracking with joint position and velocity estimation. The proposed approach integrates multi-reader RFID phase measurements with variational quantum circuits employing data re-uploading and structured entanglement to extract compact temporal representations, followed by a lightweight classical decoder for real-time inference. Experiments on both simulated motion trajectories and real RFID data collected using commercial readers and USRP-based SDR platforms demonstrate that QLSTM achieves up to 23% lower position error and 31% improved velocity estimation accuracy compared to classical LSTM baselines, while maintaining inference latency below 20 ms. These results indicate that quantum-enhanced temporal modeling can improve RFID tracking performance in regimes dominated by phase periodicity, Doppler-induced dynamics, and nonlinear motion.
This work presents and validates in an indoor setting a multitarget UHF-RFID + LiDAR sensor fusion safety system for collision avoidance to track workers and other moving obstacles operating close to machinery in construction sites. The system is based on the application of the Dynamic Time Warping (DTW) algorithm using LiDAR point clouds and RFID phase data to recognize each obstacle and classify the correct nature of the target (e.g., worker, static obstacle, etc.).
We consider battery-less, passive RF tags that communicate among themselves via backscattering an external RF signal. Unlike traditional RFID or similar backscatter-based systems, they lack active readers, but are capable of direct tag-to-tag communication. We address the tag-to-tag ranging problem in such tag networks - challenged by the absence of active radio processing. Accurate ranging is critical for leveraging the dense web of tag-to-tag connections in downstream applications, most notably localization. We address the ranging problem via a frequency-domain phase-difference of arrival (FD-PDoA) technique along with a data-driven technique to address phase estimation errors. In multipath-rich indoor environments, the proposed method achieves a median ranging error of ≈7 cm for inter-tag distances up to 2 meters.
The increased use of Micro-UAS requires the development of effective localization methods to mitigate escalating privacy and security risks. Traditional active sensing techniques are susceptible to discovery or evasion. This paper addresses the gap in drone localization by proposing an AI-enhanced passive RF sensing system that capitalizes on the ELF magnetic signatures emitted by the brushless DC motors used in UASs. Building on the foundation of ELF sensing, the proposed methodology introduces a novel localization functionality within a predefined target area and extends the detection range beyond 10m (typical indoor operations). It involves the design and test of an optimized ELF receiver antenna arrays that capture ELF signals emitted from small UAS operating indoors, despite challenges posed by high ambient ELF electromagnetic noise. Multiple loop antennas were designed and experimentally characterized, both indoors and outdoors. Drone ELF signatures were analyzed using the spectral correlation function (SCF) based digital signal processing. SCF can clearly separate signatures from drone-present and environment-only cases and reveal distinct spectral patterns for different drone models in the multidimensional SCF signal domain. These SCF images are stacked to form a tensor, which in turn is applied as the input to CNN models, which were used to localize the drone in a real-world 5m x 5m grid. Customized EfficientNetB0 AI implementation reached an overall accuracy of 90% with the highest cell-level accuracy at 96%. The results confirm that ELF-based passive RF sensing combined with SCF and deep learning is a viable alternative for short-range UAV detection and coarse 2D localization.
In this paper, a new method to accurately estimate the distance of an RFID tag from a reader within milliseconds, and with only 2 MHz of bandwidth, is described. It is based on the general principles of a Frequency Modulated Continuous Wave (FMCW) radar. However, classical FMCW method is not suitable for ranging with small bandwidths owing to the limitation imposed by the Fast Fourier Transform (FFT) bin width. In this paper, a frequency discriminator based receiver is used instead of the FFT to bypass the bin limitation. Sinusoidal frequency modulation is used instead of a chirp to ensure continuous variation of the interrogator frequency with a compact bandwidth. Longer observation times enable high range accuracy even with relatively small bandwidths. The method presented in this paper is compatible with the Slotted Aloha protocol that is used in the RAIN/EPC standard. Simulations with indoor wireless channel models corroborate the operating principles and confirm range accuracy.
The rapid expansion of the Internet of Things (IoT) demands scalable and energy-efficient connectivity for billions of low-power devices. The emerging Ambient IoT (AIoT) concept, currently under study by 3GPP, seeks to enable massive machine-type communications through backscatter technology operating in licensed cellular bands. Conventional CW-based approaches increase power consumption and system complexity, motivating new paradigms that exploit ambient cellular signals for bistatic or ambient backscatter communication. This session will highlight recent advances, challenges, and standardization efforts in AIoT design, with emphasis on ambient signal reuse, cross-layer optimization, and sustainable connectivity strategies toward realizing pervasive and energy-efficient IoT networks.
Covered Topics: Introduction to RFID systems, reader architecture, backscattering concept/antenna mismatch, cascaded receive chain noise figure, mixers constraints, basics of GS1 Gen2V3 protocol (including power ramp)
This intensive one-hour workshop provides a focused look at a critical challenge in modern RFID reader design: echo cancellation. We will quickly cover the fundamentals of RFID reader architecture and backscattering before diving into the core trade-offs between in-chip and external echo cancellation methods. The session will also touch upon key aspects of the GS1 Gen2 protocol, such as power ramp requirements, and conclude with a brief demonstration and Q&A.
In this workshop, we'll connect "what to encode" with "how it gets read," so teams can make implementation decisions that scale and remain interoperable. Participants will learn how GS1 identifiers (like a GTIN) and optional data using GS1 Application Identifiers are represented in RFID tags via the GS1 EPC Tag Data Standard (TDS)-including the distinction between data-carrier-neutral representations (e.g., URIs) and the RFID-specific binary EPC encodings. We'll also cover how the TDS is improving alignment between RFID and 2D barcodes, including support for streamlined encoding, improved translation via Tag Data Translations (TDT), and new capabilities introduced in TDS 2.3-most notably the ability to encode a host name directly on the tag. This helps expand use cases with newer EPC schemes (like SGTIN+ and date-prioritized DSGTIN+) to improve interoperability with GS1 barcodes and support efficient capture of additional data like dates, lot/batch, and country of origin directly during inventory operations. Attendees will leave with a shared vocabulary, and a standards-aligned approach to planning RFID deployments.
Localizing radio-frequency identification (RFID) tags based on available signal information has been an area of interest to researchers for tracking assets of interest. Phase angle information elicited during communication between an antenna and a passive RFID tag is known to be correlated with both the radio wave's frequency and the relative distance of the tag. However, performing localization in indoor environments is challenging due to the presence of noisy signal information induced by multi-path propagation. This issue is exacerbated in environments with metallic surfaces that are conducive to signal scattering. To properly capture the relationship between radio wave frequency, phase angle, and the corresponding tag distance, we propose a hierarchical modeling approach with a robust wrapped Gaussian process (WGP) framework for regressing noisy phase angle measurements on radio wave frequency. We then use a neural inference approach for mapping phase angle behavior to expected distance based on our model's assumed data-generating mechanism. By introducing a computationally reasonable training cost up-front, we are able to make quick and accurate range estimates with available phase angle measurements. We validate our methods on field experiments conducted in a nuclear materials laboratory environment.
Nuclear material item inventories are labor-intensive manual tasks requiring careful verification that every item in an inventory area is accounted for. One common way to perform this task is by cross-referencing identification numbers which can be complex and long. We present results from a basic exercise that compares doing a manual, visual inventory task to performing the task with RFID technology. Inventory areas were set up with variables numbers of items to confirm. On average, the efficiency improved by 16x-67x when using RFID.
RFID systems can support tracking and monitoring tasks in nuclear process work, but placing reader antennas near gloveboxes introduces electromagnetic effects that are not easy to predict. The metal structure, the glass window, and the confined volume alter antenna behavior, and repeated testing inside these facilities are often limited by safety controls and access restrictions. A dependable HFSS model would allow engineers to study these conditions without constant access to the site. This work examines whether a model of the Vulcan P12 reader antenna can meet that need. We created two HFSS cases, one in free space and one against the glass window of an aluminum glovebox and collected matching vector network analyzer measurements. The comparison shows that the models do not match the measurements exactly, yet they capture the key trends, including the shift in resonance and the change in return loss caused by the glovebox. These findings indicate that the models are suitable for early-stage evaluation of RFID behavior in constrained nuclear settings, with further refinement required for precise prediction.