WiP: Using Explainability to Enhance Wireless System Security
Abstract
This work proposes an explainable and lightweight receiver-side attack detector motivated by the damage-consistency tradeoff: perturbations that substantially damage the decoded output at the receiver tend to violate physical-layer (PHY) properties and/or relationships defined by wireless protocol standards and preserved by legitimate transmissions. Rather than learning attack signatures or retraining, our detector checks receiver-side signal behavior against standards-defined PHY rules. We evaluate the proposed attack detection approach in the context of Magmaw-style universal adversarial perturbations (UAPs) on joint source-channel coding (JSCC) over orthogonal frequency division multiplexing (OFDM) for image, video, speech, and text transmission. Our current implementation checks two representative OFDM properties: pilot residual consistency and preamble-to-payload channel consistency. Other protocol-defined relationships and behaviors can be added. Preliminary results show rule-based detection AUC up to 0.919 with relatively low computational overhead, while the optional supervised ML-based detector achieves AUCs of 0.923-0.972 across modalities. Over-the-air (OTA) evaluation and stronger detector-aware attacks are ongoing work.