Conflict-Aware Meta-Learning for Generalizable Speech Deepfake Detection

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.