Dual-Granularity Orthogonal Disentanglement for Generalizable Audio Deepfake Detection

Authors: Zhuodong Liu, Hugen Lv, Xiangyu Li, Chunhong Yuan

Published: 2026-06-15 10:36:57+00:00

Comment: Accepted at Interspeech 2026, 6 pages, 3 figures

AI Summary

This paper introduces a dual-granularity orthogonal disentanglement framework to enhance the generalization of audio deepfake detectors by addressing implicit identity leakage. It enforces feature independence at both sample-level (cosine orthogonality) and batch-level (cross-covariance regularization), guided by a curriculum disentanglement schedule. The proposed method achieves strong performance across multiple benchmarks and significantly outperforms adversarial disentanglement in cross-dataset transfer while using substantially fewer parameters than large self-supervised models.

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.

Key findings

The proposed method achieves 1.35% EER on ASVspoof 2019 LA, 7.88% EER on ASVspoof 2021 DF, and 21.58% EER on In-the-Wild, demonstrating robust generalization. It surpasses gradient reversal disentanglement by 2.60% on cross-dataset transfer and performs comparably to self-supervised models with over 150 times more parameters. Ablation studies confirmed the crucial role of the identity branch and the complementary nature of the dual-granularity constraints for effective speaker-artifact disentanglement.

Approach

The authors propose a dual-branch architecture with a shallow shared encoder, a content branch utilizing multi-head self-attention, and an identity branch with statistics pooling. Feature disentanglement is enforced at two granularities: sample-level cosine orthogonality for directional decorrelation and batch-level cross-covariance regularization for linear decorrelation. A curriculum disentanglement schedule progressively strengthens these orthogonality constraints during training without adversarial dynamics or auxiliary networks.

Datasets

ASVspoof 2019 LA, ASVspoof 2021 DF, In-the-Wild

Model(s)

A custom dual-branch architecture comprising a shallow shared convolutional encoder, a content branch with convolutional blocks and multi-head self-attention, and an identity branch with convolutional blocks and statistics pooling. The input features are 80-dimensional log-mel spectrograms.

Author countries

China, Russia

← Previous