W-MIA: Membership Inference Attack against Deep Learning-based RF Fingerprinting

Deep learning-based RF fingerprinting (DRFF) systems have gained prominence for their effectiveness in wireless device authentication based on unique RF hardware features in wireless signals. However, the inherent vulnerabilities of deep learning (DL) models make DRFF systems susceptible to DL attacks tailored for RF fingerprinting. In this paper, we present W-MIA, the first practical label-only membership inference attack (MIA) against DRFF systems. W-MIA can passively eavesdrop on RF signals to construct a shadow model and perform MIA covertly. Additionally, it can enhance attack efficacy through low-rate tailored active interactions with DRFF systems. We also propose a simple yet effective countermeasure against W-MIA. Extensive experiments confirm W-MIA's high attack efficacy in a label-only setting, achieving a maximum AUC of 0.81, comparable to the latest MIA against DRFF, which assumes a more knowledgeable adversary. Furthermore, our proposed defense matches the performance of existing defenses while minimizing usability loss in DRFF systems.