Novel Loss-Enhanced Universal Adversarial Patches for Sustainable Speaker Privacy

TL;DR

This paper introduces a new loss function and scalable insertion method for universal adversarial patches to enhance speaker privacy, reducing audio quality degradation and improving transferability across different voice models.

Contribution

It proposes the Exponential Total Variance loss and a scalable UAP insertion procedure, addressing key limitations of existing speaker anonymization techniques.

Findings

Exponential TV loss improves UAP strength and imperceptibility.

The scalable UAP insertion performs well across various audio lengths.

Enhanced transferability of adversarial patches across different voice models.

Abstract

Deep learning voice models are commonly used nowadays, but the safety processing of personal data, such as human identity and speech content, remains suspicious. To prevent malicious user identification, speaker anonymization methods were proposed. Current methods, particularly based on universal adversarial patch (UAP) applications, have drawbacks such as significant degradation of audio quality, decreased speech recognition quality, low transferability across different voice biometrics models, and performance dependence on the input audio length. To mitigate these drawbacks, in this work, we introduce and leverage the novel Exponential Total Variance (TV) loss function and provide experimental evidence that it positively affects UAP strength and imperceptibility. Moreover, we present a novel scalable UAP insertion procedure and demonstrate its uniformly high performance for various audio lengths.