Reproduce “A Large-Scale Per-Speaker Analysis of Re-identification Risk in Speech Anonymization” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Orane Dufour, Paul Magron, Mickael Rouvier, Emmanuel Vincent

A large-scale analysis of speech anonymization systems reveals that individual speaker re-identification risk is not an intrinsic trait but is significantly influenced by the specific anonymization system, attacker architecture, and amount of available speech data. The study, involving nearly 5,000 speakers, found that over 86.9% of speakers could be easily re-identified in at least one configuration, while 92.4% were hard to re-identify in another.

Key insights

  • Individual speaker vulnerability to re-identification is not intrinsic but dynamically depends on the specific anonymization system, ASV attacker architecture, and amount of available speech data.
  • Aggregate privacy metrics are insufficient for comprehensive risk assessment; evaluation protocols must be conditioned on specific attacker and anonymizer configurations to ensure robust individual protection.
  • Re-identification risk distributions are highly polarized, indicating a large heterogeneity where many speakers are either very easy or very hard to link depending on the conditions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.