The Achilles' heel of SASV: A module-level comparative analysis of adversarial vulnerabilities
Abstract
Spoofing-aware speaker verification systems combine automatic speaker verification and spoofing countermeasure modules to defend against spoofing attacks. However, their vulnerability to adversarial attacks remains largely unexplored. To the best of our knowledge, this is the first comprehensive study that investigates how perturbations targeting each module individually affect overall system integrity. Notably, attackers may substantially compromise the entire system by targeting only a single module, highlighting the need to understand module-specific vulnerabilities. Through extensive experiments across six representative system architectures and three adversarial attack methods under both white-box and black-box scenarios, we uncover critical insights into the adversarial robustness of spoofing-aware speaker verification, bridging an important knowledge gap. Our findings identify the speaker verification module as the Achilles' heel of such systems, and offer new directions for developing next-generation solutions that provide stronger protection not only against spoofing attacks but also adversarial threats, ultimately enhancing the robustness of voice authentication in real-world security scenarios.