Dual-Granularity Orthogonal Disentanglement for Generalizable Audio Deepfake Detection

2026-06-15

AI Summary

This work addresses the limited generalization of existing audio deepfake detection models in cross-speaker scenarios, where speaker identity and synthetic artifacts are often conflated, leading to performance degradation and implicit identity leakage. To mitigate this, the authors propose a dual-granularity orthogonal disentanglement framework that requires neither auxiliary networks nor adversarial training. At the sample level, directional decorrelation is achieved via cosine orthogonality; at the batch level, cross-covariance regularization eliminates linear dependencies among embedding dimensions. A curriculum disentanglement strategy progressively strengthens these constraints during training. The method substantially improves cross-dataset generalization, achieving EERs of 1.35%, 7.88%, and 21.58% on ASVspoof 2019 LA, 2021 DF, and In-the-Wild datasets, respectively, and yields a 2.60% absolute improvement in cross-domain transfer performance over a gradient reversal baseline.

Abstract

Audio deepfake detectors often fail to generalize across speakers, as they learn speaker-identity features rather than synthesis artifacts, known as implicit identity leakage. Existing methods address this but incur architectural complexity or training instability. This paper proposes a dual-granularity orthogonal disentanglement framework enforcing feature independence at two levels: sample-level cosine orthogonality captures directional decorrelation, while batch-level cross-covariance regularization eliminates linear correlations across embedding dimensions. A curriculum disentanglement schedule progressively strengthens the orthogonality constraint without auxiliary networks or adversarial dynamics. Experiments on ASVspoof 2019 LA, ASVspoof 2021 DF, and In-the-Wild datasets demonstrate that the proposed method achieves 1.35%, 7.88%, and 21.58% equal error rates (EER), respectively, surpassing gradient reversal disentanglement by 2.60% absolute on cross-dataset transfer.

Problem

• audio deepfake detection
• generalization
• identity leakage
• disentanglement
• speaker-invariant

Innovation

• orthogonal disentanglement
• dual-granularity
• audio deepfake detection
• generalizable representation
• identity leakage