Safety and Robustness of Audio Watermarking

Abstract

The rapid evolution of text-to-speech (TTS) technology has greatly enhanced the realism of synthetic speech, enabling numerous beneficial applications. However, these advancements also introduce significant ethical concerns, particularly regarding impersonation, disinformation, and copyright violations. To mitigate these risks, audio watermarking has emerged as a viable solution by embedding imperceptible yet verifiable watermarks into AI-generated speech. Despite its potential, the resilience of existing audio watermarking methods against both common and adversarial perturbations remains insufficiently studied.This research presents AudioMarkBench, the first systematic benchmark aimed at assessing the robustness of audio watermarking against two major threats: watermark removal and watermark forgery. The benchmark is structured around three core components: (1) a newly developed dataset sourced from Common Voice, ensuring diversity across languages, biological sexes, and age groups; (2) an evaluation of three leading audio watermarking techniques; and (3) an analysis of watermark robustness against 15 distinct perturbation types under three adversarial settings—no-box, black-box, and white-box attacks.Through extensive experimentation, this study evaluates the effectiveness and vulnerabilities of state-of-the-art audio watermarking methods when subjected to these perturbations. The results reveal that while current watermarking techniques perform reliably in ideal conditions, they demonstrate notable weaknesses, particularly under black-box and white-box attack scenarios. Additionally, this study identifies potential fairness concerns, as robustness inconsistencies are observed across different demographic groups, underscoring the need for more equitable and resilient audio watermarking solutions.