Dual-Granularity Orthogonal Disentanglement for Generalizable Audio Deepfake Detection
The method addresses audio deepfake detectors failing to generalize by learning two separate embeddings: one for synthesis artifacts (content) and one for speaker identity. It enforces feature independence using sample-level cosine orthogonality and batch-level cross-covariance regularization, ensuring the detector relies on synthesis features rather than speaker characteristics, leading to better performance on unseen data.