A Security Analysis of Anti-Audio Cloaking Schemes

Audio cloaking aims to protect audio artifacts from unauthorized cloning by speech synthesizers and to safeguard user identity by altering signals to make them unusable for voice synthesis and speaker identification. In this paper, we investigate the resilience of current audio cloaking methods from an adversarial perspective, analyzing whether an attacker can reverse these cloaking mechanisms to retrieve the original speaker features for accurate speaker verification and synthesis. We systematically apply ten different audio restoration techniques, including adaptive frequency filtering, quantization-based filtering, ensemble averaging, and various spectral modifications, to cloaked audio samples. The effectiveness of each technique is rigorously evaluated using prominent speaker verification systems. Our findings show that most inversion attempts achieve limited success due to robust cloaking mechanisms; however, low-pass filtering emerged that is notable for its effectiveness, partially restoring speaker identity traits. WavLM achieved the highest similarity scores (up to 0.91) for low-pitch voices, and intelligibility was preserved with STOI scores greater than 0.28. partially reverse cloaking effects and restore speaker identity traits, especially in low-frequency voice profiles. This research shows significant vulnerabilities in cloaking strategies, suggesting necessary improvements in future audio cloaking methods. By exposing these weaknesses, our study provides crucial information on enhancing the robustness of audio privacy and security measures.