SegReConcat: A Data Augmentation Method for Voice Anonymization Attack

Anonymization of voice seeks to conceal the identity of the speaker while maintaining the utility of speech data. However, residual speaker cues often persist, which pose privacy risks. We propose SegReConcat, a data augmentation method for attacker-side enhancement of automatic speaker verification systems. SegReConcat segments anonymized speech at the word level, rearranges segments using random or similarity-based strategies to disrupt long-term contextual cues, and concatenates them with the original utterance, allowing an attacker to learn source speaker traits from multiple perspectives. The proposed method has been evaluated in the VoicePrivacy Attacker Challenge 2024 framework across seven anonymization systems, SegReConcat improves de-anonymization on five out of seven systems.

Key Contributions

SegReConcat: a word-level segment rearrangement and concatenation data augmentation method for training attacker-side ASV systems that de-anonymize speaker-anonymized speech
Exploration of random vs. similarity-based rearrangement strategies to disrupt long-term contextual cues and force focus on short-term speaker traits
Evaluation across seven anonymization systems in the VoicePrivacy Attacker Challenge 2024 framework, showing improved de-anonymization on five of seven systems

🛡️ Threat Analysis

Output Integrity Attack

The paper attacks the output integrity of ML-based voice anonymization systems — the anonymized speech claims to hide speaker identity, but SegReConcat shows residual identity cues persist. This is an attack on the claimed anonymization integrity of model outputs, analogous to defeating a content protection scheme.