Work

Privacy Attacks on Voice Anonymization Systems: Overview and Key Findings from the First VoicePrivacy Attacker Challenge

Year: 2026

Type: preprint

Abstract: First VoicePrivacy Attacker Challenge evaluates 55 attack systems from 11 teams against 7 voice anonymization systems from VPC 2024. Best attackers reduce EER by 25-44% relative (7-18% absolute), revealing that privacy protection was previously overestimated. Top attack systems achieve EER∼20-28% compared to baseline EER∼27-43%. Diverse attack architectures, data mixing strategies, multi-stream fusion (x-vector + prosody + pitch), reconstruction approaches, transfer learning with self-supervised models, and text-based exploits are explored. Speaker-level linkability analysis shows that anonymization systems provide unequal protection across speakers. Stronger attacks change anonymization system rankings relative to the baseline and reveal information leaks via prosody, temporal dynamics, and linguistic content.

Source: HAL

Authors: Tomashenko, Natalia, Miao, Xiaoxiao, Vincent, Emmanuel, Yamagishi, Junichi

Institutions: Centre Inria de l'Université de Lorraine, MULTISPEECH: La parole multimodale en interaction, Duke Kunshan University, National Institute of Informatics

Language: English

Topic: Speech Recognition and Synthesis

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

Funders: Agence Nationale de la Recherche, CHIST-ERA

Award: Multiple-attribute disentanglement and semantic privacy

OpenAlex ID: W7135159993

Cites: 32

Cited by:

Related to: 10