The advancement of deepfake technology has enabled deepfake audio to achieve unprecedented realism, posing significant challenges to automatic speaker verification (ASV) systems and threatening information security in critical sectors such as telecommunications and banking. As a countermeasure, Audio Deepfake Detection (ADD) has received growing attention in recent years. Existing ADD techniques derived from deep‐learning framework often encounter the problem of insufficient generalizability on unseen deepfake methods and insufficient robustness with changes to the codec and noise condition. To address these challenges, we propose an efficient method that adapts self‐supervised learning (SSL) model WavLM for deepfake detection. Instead of costly full‐model retraining, we introduce a lightweight “adapter” module, which acts as a small, trainable component to efficiently teach WavLM this new task. This adapter incorporates a Gaussian attention mechanism, which guides the model to consistently focus on subtle vocal artifacts and ignore irrelevant noise or compression distortions, thereby ensuring robust performance across different real‐world conditions. Extensive evaluations show that our framework not only achieves competitive performance on the ASVspoof 2019 LA and DF in‐domain benchmarks but, more importantly, excels in generalization. Our method achieves EERs of 6.11% on the challenging ‘in‐the‐wild’ dataset and 4.90% on the WaveFake dataset, outperforming existing methods against previously unseen deepfake attacks. This superior generalization is further explained through SHAP analysis, which reveals that our method enables the model to focus on discriminative information within voiced segments, effectively ignoring fragile ‘shortcut’ cues from silent portions that often hinder generalization in other systems.
Wang et al. (Thu,) studied this question.