New LID metric reveals voice anonymization flaws hidden by EER scores
Top voice anonymization systems with 48% EER still leak 1 bit per utterance.
Deep Dive
The authors of a recent paper propose Local Information Disclosure (LID) to replace Equal Error Rate (EER) for evaluating voice anonymization. LID quantifies privacy loss in bits under real-world 1-to-N linkage attacks. Testing VoicePrivacy 2024 systems showed near-perfect EERs (48%) masked worst-case disclosures up to 1 bit per trial, doubling attacker success over random guess. The framework reveals localized vulnerabilities that global EER averages obscure.
Key Points
- EER (Equal Error Rate) is inadequate for 1-to-N linkage threats; it aggregates globally and obscures local privacy failures.
- New metric LID quantifies privacy loss in bits per utterance; worst-case LID reached 1 bit for top VoicePrivacy 2024 systems.
- Systems with 48% EER (near-perfect by old standard) still double attacker success over random guess for specific utterances.
Why It Matters
Current voice anonymization benchmarks hide real privacy risks; LID offers a regulation-aligned, attacker-realistic evaluation standard.