Multi-Ratio DiT Reconstruction Residuals Improve Cross-Domain Audio Deepfake Detection

There's a new technique for spotting audio deepfakes that uses a frozen Diffusion Transformer (DiT), which has been specifically trained on real speech as a way to check for authenticity. This method generates multi-ratio residual maps at masking ratios of 0.5, 0.75, and 0.9, designed to be sensitive to the audio domain. The detection system combines frozen-WavLM auditory representations with these residuals using a method that avoids traditional gate-based attenuation. In the ASVspoof 5 Eval, the seed-42 run achieved an EER of 6.5442% and a min-DCF of 0.18456, while the ITW Full scores were 13.8372% EER and 0.36921 min-DCF. The average performance across three seeds was 6.8885% EER and 15.3328% min-DCF, outperforming a separately optimized model. Plus, adding auxiliary supervision improved the dynamic fusion but negatively affected all three seed results.