DGS-MLDG: Domain Gradient Surgery Guided Meta-Learning for Domain Generalization in Speech Deepfake Detection

Authors: Siqing Qin, Kong Aik Lee, Youzhi Tu, Eng Siong Chng, Man-Wai Mak

Published: 2026-09-27 16:07:30+00:00

Comment: Accepted by INTERSPEECH 2026

AI Summary

This paper addresses domain shifts in speech deepfake detection by proposing Domain Gradient Surgery (DGS), a meta-learning method that resolves conflicting gradients between meta-train and meta-test objectives. DGS achieves this through an asymmetric projection strategy, removing destructive components from the meta-test gradient. An efficient variant, Layer-Wise DGS (LW-DGS), is also introduced to dynamically intervene only in conflict-prone layers.

Abstract

Speech deepfake detection faces significant challenges due to domain shifts. Domain generalization (DG), particularly meta-learning for domain generalization (MLDG), offers a promising solution by simulating and mitigating domain shifts. However, MLDG is often hindered by conflicting gradients between its meta-train and meta-test objectives, leading to suboptimal performance. To address this problem, we propose domain gradient surgery (DGS), a meta-learning method that resolves conflicts through an asymmetric projection strategy. DGS removes the destructive component from the meta-test gradient, ensuring a conflict-free optimization trajectory versus the meta-train gradient. Furthermore, we introduce layer-wise DGS (LW-DGS), an efficient variant of DGS that dynamically identifies and intervenes only conflict-prone layers. Extensive experiments on challenging benchmarks demonstrate that DGS-MLDG and LW-DGS-MLDG achieve an average relative EER reduction of 5.29% and 4.04%, respectively.

Key findings

DGS-MLDG achieved an average relative Equal Error Rate (EER) reduction of 5.29% over the strongest baseline across challenging benchmarks, demonstrating superior generalization against unseen codecs and real-world spoofing scenarios. LW-DGS-MLDG, an efficient variant, achieved a comparable 4.04% relative EER reduction, confirming that targeted intervention on conflict-prone layers is effective. The asymmetric projection strategy of DGS was found to be crucial for robust domain generalization, outperforming general multi-task gradient surgery methods that treat tasks symmetrically.

Approach

The proposed DGS-MLDG framework modifies the meta-learning for domain generalization (MLDG) process. When gradients from meta-train and meta-test objectives conflict (negative cosine similarity), DGS projects the meta-test gradient onto the normal plane of the meta-train gradient, ensuring a conflict-free optimization trajectory. LW-DGS extends this by applying the surgery selectively to layers where gradient conflicts are detected.

Datasets

ASVspoof 2019 LA, ASVspoof 5, CFAD (training); ASVspoof 2021 DF, ASVspoof 5, ADD 2023 (R1 & R2), In-the-wild, CodecFake, CFAD unseen test set (evaluation)

Model(s)

XLSR-Mamba backbone with a linear projection and BiMambas prediction head

Author countries

Hong Kong, Singapore

← Previous