ADA: Adversarial Dynamic Test Time Adaptation in Radio Frequency Machine Learning Systems

Deep Neural Networks (DNNs) are an attractive solution to address several problems in Radio Frequency Machine Learning Systems (RFMLS). The key blocker that inhibits the widespread deployment of DNNs in real-world tactical scenarios is the performance degradation experienced under dynamic channel conditions. Test Time Adaptation (TTA) presents a promising solution to mitigate this issue by dynamically updating the DNN to adapt to the current channel conditions in an unsupervised manner. Although it offers superior performance and practical benefits, TTA introduces new security concerns and vulnerabilities, potentially exposing sensitive deployments to Adversarial Machine Learning (AML) activity. In this work, we introduce a novel attack strategy named Adversarial Dynamic Adaptation (ADA) that leverages the inherent vulnerabilities in TTA to compromise RFMLS tasks. We demonstrate that even under realistic assumptions and while perturbing only 20% of the samples in a test data batch, ADA degrades the performance of the unperturbed data by up to 20.3% compared to similar attacks designed for computer vision tasks. By assessing the robustness against the latest TTA methods, ADA serves as a valuable tool to identify and understand the security risks associated with adapting DNNs at the test time in mission-critical and sensitive deployments.