AI/Machine Learning (ML) is essential to automation and agent-human teaming to act as a force multiplier in many tasks such as image and video analytics, network management, and cyber defense and cyber offense. Recent works have demonstrated the security implications of the use of machine learning and is often referred to as Adversarial Machine Learning (AML). AML is the attackers' attempt to deceive a machine learning classifier or predictor with objectives that may include targeted or indiscriminate misclassification of samples, corruption of prediction, corruption of confidence in the ML algorithm, etc. AML also refers to defenses to make ML robust to adversarial manipulations, and has led to improved explainability of ML. The workshop seeks to solicit original research in the areas of Security of Machine Learning, Adversarial Machine Learning (i.e., evasion and poisoning) and defenses against such attacks, particularly in the context of AI/ML for cyber security. The workshop seeks to increase understanding of theory, scalable algorithms, and applications, and highlight challenges.

Adversarial Machine Learning (AML) has shown significant success when applied to deep learning models across various domains. AML generation of realizable samples for cyber applications requires an understanding of domain constraints.

In this session the robustness of transformers to adversarial samples for system defenders (i.e., resiliency to adversarial perturbations generated on different types of architectures) and their adversarial strength for system attackers (i.e., transferability of adversarial samples generated by transformers to other target models) are evaluated. Also presented is, a graph machine learning-based framework for detecting cyber attacks in mobile tactical software-defined networks. Finally, a novel application of the tensor decomposition method Canonical Polyadic Alternating Poisson Regression (CP-APR) with a probabilistic framework to identify anomalies in SCADA systems.

In this session, a framework is proposed for explainable AI and is successfully evaluated on a use case on conducting the proportionality assessment in military Cyber Operations and contributes to building responsible and trustworthy AI systems in the military domain. This session also discusses the integration of data-driven learning with structured reasoning, Neurosymbolic AI, to promise a more robust, adaptive, and transparent cybersecurity solutions.

The evolution of Artificial Intelligence, Control, and Communications technologies has given rise to a new era of networked intelligent systems, which include autonomous driving, remote surgery, real-time surveillance, video analytics, and factory automation. Timely Inference is vital in these systems, where a trained neural network infers time-varying targets (e.g., the locations of vehicles and pedestrians) based on observations (e.g., video frames) captured by a sensing node (e.g., camera). Due to communication delay, the data delivered to the neural network may not be fresh, impacting both inference accuracy and overall system performance. In this talk, we will first examine the influence of information freshness on remote inference and estimation. One might assume that inference and estimation errors degrade monotonically as the data becomes stale. However, by a local information geometric analysis, we reveal that this assumption is true when the time-sequence data used for remote inference and estimation can be closely approximated as a Markov chain; but it is not true when the data sequence is far from Markovian. Hence, inference and estimation errors are functions of the Age of Information (AoI), whereas the function is not necessarily monotonic. This analysis provides an information-theoretic interpretation of information freshness. The second part of the talk focuses on the design of communication systems optimized for remote inference and estimation. We introduce a novel "selection-from-buffer" model for data transmission, which is more general than the "generate-at-will" model used in earlier AoI studies. Low-complexity scheduling strategies are developed to minimize inference and estimation errors. Trace-driven evaluations demonstrate the potential of these communication strategies to reduce inference and estimation errors by up to 10-10000 times. We will also discuss future directions, such as context-aware status updating and situational awareness maximization in safety-critical systems.