Audio-Anchored Fusion of Multi-Ratio DiT Reconstruction Residuals for Cross-Domain Audio Deepfake Detection

View on arXiv ← Back to list

Authors: Haotian Mo, Jie Liu, Siqi Shen, Songzhu Mei, Xinhai Chen, Xiangyang Wang, Yigui Feng, Shuai Li, Gencheng Liu, Keqi Yang, Qinglin Wang

Published: 2026-07-29 04:55:14+00:00

Comment: 10 pages, 5 figures. Submitted to speech security conference. This work proposes a cross-domain audio deepfake detection framework based on bona-fide trained DiT multi-ratio reconstruction residuals and audio-anchored additive fusion, evaluated on ASVspoof 5 and real-world ITW datasets

AI Summary

This research introduces an audio-anchored deepfake detection method using multi-ratio Diffusion Transformer (DiT) reconstruction residuals. The DiT, trained solely on bona fide speech, generates reconstruction residual maps at various masking ratios (0.5, 0.75, 0.9). These residuals are then additively fused with a projected frozen-WavLM auditory representation to develop a deepfake detector that aims for improved generalization across different deepfake generation conditions.

Abstract

Audio deepfake detectors often degrade when generators, corpora, or recording conditions change. We use a Diffusion Transformer (DiT), trained only on bona fide speech, as a frozen reconstruction probe. Reconstructions at masking ratios 0.5, 0.75, and 0.9 yield explicit multi-ratio residual maps. Because these residuals are domain sensitive, our audio-anchored detector passes the projected frozen-WavLM auditory representation into the fusion sum without gate-based attenuation and uses residuals only as a scalar-gated additive correction. The pre-specified seed-42 run obtains 6.5442% EER / 0.18456 min-DCF on ASVspoof 5 Eval and 13.8372% / 0.36921 on ITW Full; three-seed means are 6.8885 (0.3308)% and 15.3328 (2.0719)%. The latter is below a separately optimized WavLM-ResNet18 reference under both supervision settings. Auxiliary supervision raises dynamic competitive fusion from 18.4007% to 25.2968% mean ITW EER, worsening all three seeds. The results support reconstruction residuals as complementary evidence and motivate a non-competitive auditory path for ASVspoof 5-to-ITW transfer, without claiming a componentwise causal ablation of anchoring alone.

Key findings

The audio-anchored fusion model achieved 6.5442% EER / 0.18456 min-DCF on ASVspoof 5 Eval and 13.8372% / 0.36921 min-DCF on ITW Full, outperforming a separately optimized WavLM-ResNet18 reference. The results indicate that reconstruction residuals serve as complementary evidence and that a non-competitive auditory path is beneficial for cross-domain transfer (ASVspoof 5-to-ITW), particularly highlighting a fusion-supervision interaction where auxiliary supervision worsened dynamic competitive fusion but improved the audio-anchored approach.

Approach

The method involves training a Diffusion Transformer (DiT) exclusively on bona fide speech to act as a frozen reconstruction probe. At detection, the DiT reconstructs masked Mel spectrograms at three masking ratios, yielding explicit residual maps. These multi-ratio residuals are then combined with a projected frozen-WavLM auditory representation using an audio-anchored additive fusion technique, where the auditory path is not attenuated by a gate, and residuals provide a scalar-gated additive correction.

Datasets

ASVspoof 5 (Train, Dev, Eval), ITW Full, Common Voice, VoxCeleb2, LibriSpeech

Model(s)

Diffusion Transformer (DiT), WavLM-Large, ResNet-18

Author countries

China

← Previous