UDA-SD: Unsupervised domain adaptation based on self-distillation for cross-domain audio anti-spoofing

Audio anti-spoofing algorithms perform well in in-domain scenarios, but their performance deteriorates significantly under domain mismatch conditions. Some studies have incorporated domain adversarial training strategy to encourage the learning of domain-invariant representations, reducing domain differences. However, this strategy could compromise the model’s ability to distinguish between bona fide and spoofed audio, limiting its effectiveness in mitigating domain mismatch. To address this problem, we propose a novel U nsupervised D omain A daptation algorithm based on S elf- D istillation, termed UDA-SD. We introduce a Siamese framework with two branches tailored for the source and target domains, including three key modules: (i) The source branch leverages pre-trained audio representations to enhance spoofing discrimination capability, encouraging the model to capture the common differences between bona fide and spoofed audio. (ii) The target branch captures domain-specific and intrinsic features through a self-distillation module, thereby reducing domain discrepancies. (iii) Within the self-distillation module, we align the student model’s predictions for varying-duration audio segments with the teacher model’s predictions for long-duration segments. This design enhances the feature exploration of target samples by capturing both global and local representations, uncovering comprehensive differences between categories. To evaluate the effectiveness of the proposed UDA-SD, we conducted eight cross-domain anti-spoofing tasks. Experimental results showed the effectiveness of our approach with respect to source-only audio anti-spoofing models, achieving a relative 54.31 % to 85.14 % performance improvement in cross-domain scenarios. • We propose UDA-SD, a self-distillation-based unsupervised domain adaptation approach. • UDA-SD adopts a Siamese model with wav2vec 2.0 for cross-domain audio anti-spoofing. • The self-supervised module explores target domain data and mitigates domain mismatches. • Self-distillation captures global and local cues to model intrinsic target features. • UDA-SD outperforms conventional models across eight cross-domain scenarios.