Rethinking Targeted Data Poisoning in Voice Authentication: A Critique and Defense Mechanism

Recent deep learning techniques have significantly improved voice authentication systems. However, they remain vulnerable to threats, including data poisoning and Man-in-the-Middle (MitM) attacks. This paper reevaluates the attack and defense mechanisms of “The Guardian”, examining their assumptions and modifying the proposal. We conduct experiments with real-world datasets to validate their effectiveness under realistic conditions and assess the feasibility of executing such attacks. Additionally, we introduce a defense mechanism that improves resilience after redefining a threat model grounded in operational feasibility, specifically by isolating the enrollment phase from training phase assumptions. By analyzing the current literature, we identify open challenges and suggest directions for further improving the security of voice authentication systems.