Learning Tight Rejection Boundaries without Negatives for Strict One-Class Audio Deepfake Detection

The rapid evolution of audio deepfakes requires robust detection capable of generalizing to unseen attacks. One-class learning offers inherent robustness for this task by characterizing real speech distributions to detect anomalies. However, establishing a compact decision boundary without spoof supervision remains a fundamental challenge. Existing relaxed approaches often compromise this strictness by introducing auxiliary negative samples, which biases the boundary toward seen artifacts and degrades generalization to unseen attacks. To address this, we propose CA-SOADD, a framework that refines the acceptance region by constructing off-manifold boundary probes. Our proposed centroid-anchored tri-objective learning paradigm simultaneously enforces centroid compactness and a centroid-referenced margin against these probes, thereby explicitly tightening the acceptance region without treating them as an explicit negative class. We further extend the framework to heterogeneous settings through domain-conditioned centroids. Experiments on ASVSpoof and MLAAD benchmarks demonstrate that our strict real-only method consistently outperforms strong baselines under unseen attack types and domain shifts, with its effectiveness further validated through extensive ablation studies.

Audio deepfakes are becoming increasingly realistic, making it difficult for detection systems to recognize new types of fake speech. Many detectors rely on examples of both real and fake speech during training, but this can make them overfit to known fake patterns and perform poorly on unseen attacks. This paper studies a stricter and more practical setting where the detector learns only from real speech. The key challenge is to define a tight boundary around genuine speech without using fake speech examples. We propose CA-SOADD, a method that creates special boundary probes to help the model better learn where real speech should end. These probes are not treated as fake speech, but they guide the model to form a more compact and reliable description of genuine speech. Experiments on standard audio deepfake benchmarks show that CA-SOADD detects unseen attacks more effectively than strong existing methods, especially when the test data differs from the training data. This suggests that learning from real speech alone can improve robustness against future audio deepfakes.