When Normality Shifts: Multi-Level WavLM Modeling for Open-World Audio Deepfake Detection

Authors: Daniel Niekurzak, Daniel Kozuch, Weronika Folwarska, Khan Muhammad, Rafał Scherer

Published: 2026-11-08 15:54:07+00:00

Comment: Accpeted for the OWAD Workshop at ICDM 2026

AI Summary

This paper proposes MLN-WavLM, an attack-free audio deepfake detection method that models normality using shrinkage-Mahalanobis distances on pooled hidden states from a frozen WavLM Base+ architecture. While MLN-WavLM achieves competitive performance in-domain, its effectiveness significantly drops under domain shift, highlighting challenges in open-world deepfake detection without supervised attack data. The study concludes that while ranking deepfakes is possible, maintaining a stable decision rule in an open-world setting remains difficult, even with threshold adaptation.

Abstract

Audio 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. We propose MLN-WavLM, a method that applies shrinkage-Mahalanobis modeling to pooled hidden states from a frozen large-scale, pre-trained speech processing WavLM Base+ architecture. Evaluated on ASVspoof 2019 LA, our best attack-free variant achieves an 8.45\\% Equal Error Rate (EER), closely tracking a fully supervised WavLM-HM-Conformer baseline (7.86%). However, when subjected to the ASVspoof 2021 DF domain shift, performance drops to 22.12% EER, exposing vulnerabilities in the bona-fide-calibrated operating point. Layer-wise analysis reveals that early WavLM representations dominate the ensemble. To test real-world viability, we simulated deployment streams using unlabeled threshold adaptation. While robust median/MAD recalibration easily handles in-domain contamination up to 10%, recalibrating under the 2021 DF domain shift triggers a total collapse of low-FPR recall. We conclude that while frozen source-domain normality can effectively rank deepfakes, it does not yield a stable open-world decision rule, and downstream threshold adaptation cannot rescue upstream ranking degradation.


Key findings
MLN-WavLM-Full achieves an 8.45% EER on ASVspoof 2019 LA, closely matching a supervised WavLM-HM-Conformer baseline (7.86%). However, performance degrades significantly to 22.12% EER under the ASVspoof 2021 DF domain shift, which is a larger degradation than the supervised baseline. While robust unlabeled threshold adaptation can restore the in-domain operating point up to 10% contamination, it fails to restore recall under significant domain shift, indicating that the ranking itself collapses at low False Positive Rates.
Approach
The proposed MLN-WavLM method is an anomaly detection approach that trains exclusively on bona fide speech. It applies shrinkage-Mahalanobis modeling to pooled hidden states from multiple layers (3, 6, 9, 12) of a frozen WavLM Base+ architecture. The anomaly score is computed as a standardized Mahalanobis distance from the bona fide distribution, with higher scores indicating a greater deviation, and scores from different layers are averaged.
Datasets
ASVspoof 2019 LA, ASVspoof 2021 DF
Model(s)
WavLM Base+, WavLM-HM-Conformer (as a supervised reference)
Author countries
Poland, Republic of Korea