Dual-Granularity Orthogonal Disentanglement for Generalizable Audio Deepfake Detection

2026-06-15 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

A new framework, Dual-Granularity Orthogonal Disentanglement, addresses the challenge of generalizable audio deepfake detection by preventing detectors from learning speaker-identity features instead of synthesis artifacts. This method tackles implicit identity leakage, a common issue causing poor generalization across speakers, without increasing architectural complexity or training instability. It enforces feature independence at two levels: sample-level cosine orthogonality for directional decorrelation and batch-level cross-covariance regularization for eliminating linear correlations across embedding dimensions. The framework incorporates a curriculum disentanglement schedule to progressively strengthen orthogonality constraints. Experiments on ASVspoof 2019 LA, ASVspoof 2021 DF, and In-the-Wild datasets yielded equal error rates (EER) of 1.35%, 7.88%, and 21.58%, respectively. Notably, it outperformed gradient reversal disentanglement by 2.60% absolute on cross-dataset transfer.

Key takeaway

For AI Security Engineers developing robust audio deepfake detection systems, consider integrating dual-granularity orthogonal disentanglement. This approach directly addresses implicit identity leakage, improving model generalization across diverse speakers and unseen data. You should evaluate its performance on cross-dataset transfer benchmarks, as it demonstrated a 2.60% absolute improvement over gradient reversal methods, potentially enhancing the reliability of your deepfake countermeasures.

Key insights

Dual-Granularity Orthogonal Disentanglement improves audio deepfake detection generalization by separating speaker identity from synthesis artifacts.

Principles

• Feature independence improves deepfake generalization.
• Orthogonality can prevent identity leakage.
• Progressive constraint strengthening aids stability.

Method

The framework uses sample-level cosine orthogonality and batch-level cross-covariance regularization to enforce feature independence. A curriculum disentanglement schedule progressively strengthens these constraints.

In practice

• Apply dual-granularity disentanglement to audio deepfake models.
• Implement curriculum scheduling for stable training.
• Evaluate cross-dataset transfer for generalization.

Topics

• Audio Deepfake Detection
• Feature Disentanglement
• Generalization
• Orthogonality
• ASVspoof
• Curriculum Learning

Best for: Research Scientist, AI Scientist, AI Security Engineer, Machine Learning Engineer

Related on AIssential

• HybridCodec: Fast Dual-Stream, Semantically Enhanced Neural Audio Codec — HybridCodec combines dual-stream and distillation for fast, semantically specialized neural audio coding without SSL inference overhead.
• Feature-Aligned Speech Watermarking for Robustness to Reconstruction Distortions — Aligning watermarks with speech features improves robustness against reconstruction models while preserving imperceptibility.
• XAI-Grounded Explanation Generation for Speech Deepfake Detection with Training-Free Multimodal Large Language Models — Integrating XAI with multimodal LLMs generates grounded, human-understandable explanations for speech deepfake detection, improving accuracy by over 45%.
• Fast When, Careful Who: Dual-Process Multiparty Turn-Taking with Diffusion Augmentation — A dual-process pipeline with diffusion augmentation improves multiparty turn-taking by separating turn boundary triggering from floor transfer decisions.
• Contrastive and Adversarial Disentanglement for Speaker Representations in Brazilian Portuguese — Combining adversarial and contrastive learning improves speaker representation disentanglement from environmental factors.
• Architecture-Adaptive Uncertainty Fusion for Deepfake Detection — Correlation-Optimized Fusion (COF) adaptively combines multiple uncertainty sources to predict deepfake detection errors, improving cross-domain reliability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.