GFE-PGDI: Adversarial Perturbation Generation by Integrating Global Feature Extraction and PGD-Based Optimization in ASV Systems

Adversarial perturbation for Automatic Speaker Verification (ASV) systems has provided great opportunity to ensure the trustworthiness and continuous improvement of voice-based authentication technologies in the face of evolving security threats. The technology remains challenging due to the high requirement of attacks that are both effective and stealthy while navigating the increasing sophistication of system defences and voice data. This paper proposes a novel method called GFE-PGDI for the general task of adversarial perturbation generation by integrating global feature extraction and PGD-based optimization. Specifically, it relies on a Mel-Frequency Cepstral Coefficients (MFCC) strategy to perform global feature extraction from the input audio and uses Projected Gradient Descent (PGD) to maximize the effectiveness of the attack by calculating the cosine similarity between the feature matrix and the target speaker's voiceprint matrix. It also applies a customized Carlini and Wagner (CW) function attack to minimize the impact on ASV, generating system-independent and text-independent perturbations. Experimental evaluations show that the perturbations can effectively mislead speaker verification systems in different environments with a success rate of over 95%, further validating the effectiveness of the GFE-PGDI.