Attack-free WavLM detector ranks audio deepfakes but fails under shift
Original title: When Normality Shifts: Multi-Level WavLM Modeling for Open-World Audio Deepfake Detection
The best variant scores 8.45% EER on ASVspoof 2019 LA, then 22.12% EER on the 2021 DF shift.
MLN-WavLM scores synthetic audio by its deviation from bona fide speech, using shrinkage-Mahalanobis modeling on frozen WavLM Base+ hidden states. A supervised WavLM-HM-Conformer baseline reaches 7.86% EER on ASVspoof 2019 LA. Under the ASVspoof 2021 DF domain shift, the attack-free detector's EER rises to 22.12%. Unlabeled median/MAD recalibration tolerates up to 10% in-domain contamination, but recalibration under the 2021 shift collapses low-FPR recall.
Unconfirmed: Deployment results use simulated unlabeled streams, not a reported production system.
- Early WavLM representations dominate the layer-wise ensemble.
- Median/MAD recalibration handles in-domain contamination up to 10%.
- Authors say frozen normality does not yield a stable open-world decision rule.
From the original article
1 min read at arxiv.orgAudio deepfake detectors trained with known synthesis attacks can learn generator-specific shortcuts that fail as synthesis systems and acoustic conditions evolve. We investigate an attack-free alternative: a normality model trained exclusively on bona fide speech, which scores synthetic audio by its deviation from authentic representations.
By Daniel Niekurzak · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.