Conflict-Aware Meta-Learning for Generalizable Speech Deepfake Detection

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

Published: 2026-11-05 22:13:08+00:00

Comment: submit to TASLP; 13 pages

AI Summary

This paper introduces Conflict-Aware Meta-Learning (CAML), a framework designed to enhance the generalizability of speech deepfake detectors under domain shifts. CAML addresses gradient conflicts and optimization instability in meta-learning by employing asymmetric gradient projection and an exponential moving average (EMA) anchor, which provides a stable reference for reliable conflict detection. The framework significantly reduces error rates in cross-dataset evaluations, demonstrating improved robustness.

Abstract

Speech deepfake detectors built on large self-supervised learning front-ends achieve high in-domain accuracy, yet suffer degradation under domain shift caused by unseen synthesizers, transmission codecs, or unfamiliar acoustic conditions. Meta-learning, particularly model-agnostic meta-learning for domain generalization (MLDG), addresses this issue by simulating domain shifts. However, applying MLDG directly introduces a bi-level optimization challenge that remains unsolved. First, effective domain-shift emulation requires genuine distributional gaps between meta-splits, yet such gaps can frequently produce gradient conflicts, destabilizing the bi-level optimization. Second, the stochastic nature of mini-batch gradients causes the reference direction to fluctuate across iterations, resulting in an optimization trajectory that fails to achieve domain generalization. We propose a conflict-aware meta-learning (CAML) framework, which addresses both challenges through three complementary components. Firstly, base domain gradient surgery (base-DGS) introduces an asymmetric projection that treats the meta-train gradient as a primary anchor. Secondly, EMA-DGS replaces the unstable anchor with its exponential moving average (EMA), providing a stable optimization trajectory anchor that substantially reduces domain-shift-induced conflict detections. Thirdly, layer-wise DGS (LW-DGS) further extends the framework to individual parameter tensors, enabling fine-grained, layer-specific intervention that targets conflict-prone layers while leaving well-aligned parameters intact. Experimental results demonstrate that the proposed CAML framework achieves lower EER than the compared baselines on most cross-dataset evaluations across multiple cross-dataset benchmarks and architectures, achieving a maximum relative improvement of 45.2% in terms of EER.


Key findings
The CAML framework achieved lower Equal Error Rates (EER) than baselines in most cross-dataset evaluations, with a maximum relative improvement of 45.2%. EMA-DGS was particularly effective in multi-dataset settings due to its stable projection anchor, while LW-EMA-DGS offered complementary benefits for specific domain shifts. Asymmetric gradient surgery proved more effective than symmetric approaches, which can harm the model's ability to learn source-domain features.
Approach
CAML proposes a meta-learning framework with three components: base domain gradient surgery (base-DGS) uses asymmetric projection to treat the meta-train gradient as an anchor, correcting only conflicting meta-test gradients. EMA-DGS stabilizes this anchor using an exponential moving average to reduce conflict detections. Layer-wise DGS (LW-DGS) extends this to individual parameter tensors for fine-grained, layer-specific intervention.
Datasets
ASVspoof 2019 LA, ASVspoof 5, CFAD, ASVspoof 2021 DF (21DF) test set, In-the-Wild (ITW), Fake-or-Real (FoR), ADD 2023 R1 and R2, CodecFake
Model(s)
XLSR-Conformer-TCM, XLSR-AASIST, XLSR-Mamba (all utilizing the XLS-R 300M self-supervised learning front-end)
Author countries
Hong Kong, Singapore