Triple-Domain Training and Waveform-Level Augmentation for Generalizable End-to-End Audio Deepfake Detection: An Extended Replication of RawNetLite

This paper presented an extended replication of a RawNetLite-based audio deepfake detector, focusing specifically on the triple-domain training protocol — combining FakeOrReal, AVSpoof2021, and CodecFake during training, applying Focal Loss to emphasize hard or ambiguous examples, and applying waveform-level augmentation to simulate realistic acoustic degradation. The headline finding is that completing the triple-domain training and evaluation protocol closes a meaningful fraction of the cross-domain generalization gap left open by single-dataset training, while leaving FakeOrReal in-domain performance essentially intact (F1 = 97.43%, EER = 2.66%). On the combined AVSpoof2021 + CodecFake test partition, the model reached an F1-score of 80.84% and an EER of 16.4%, comfortably inside the 70–82% F1 range reported for the equivalent configuration in the source literature [27]. CodecFake's codec-induced distortion remained the more difficult of the two individual cross-domain targets (F1 = 74.66%, EER = 19.1%). Future work may extend this pipeline in several directions: completing CodecFake evaluation using the full C1–C6 codec-type test partitions; extending AVSpoof evaluation to the larger ASVspoof2021 partition; exploring self-supervised pretraining and domain-adversarial training as further generalization strategies; and investigating open-set recognition formulations that explicitly model unseen synthesis methods.

[1] A. van den Oord et al., "WaveNet: A generative model for raw audio," arXiv:1609.03499, 2016. [2] J. Shen et al., "Natural TTS synthesis by conditioning WaveNet on Mel spectrogram predictions," ICASSP, 2018. [3] A. Korshunov and S. Marcel, "Vulnerability assessment and detection of deepfake audio," IEEE BTAS, 2018. [4] M. Todisco et al., "ASVspoof 2019: Future directions for spoofed and fake audio detection," IEEE J. Sel. Topics Signal Process., vol. 14, no. 5, pp. 1038–1048, 2020. [5] J.-W. Jung et al., "RawNet: Advanced end-to-end deep learning for speaker verification," Interspeech, 2019. [6] Y. Zhang et al., "Anti-spoofing with spectrogram-based CNNs," IEEE Trans. Inf. Forensics Security, vol. 15, 2020. [7] J.-W. Jung et al., "RawNet2: Improved end-to-end deep neural networks for speaker verification," ICASSP, 2020. [8] J. Tak et al., "End-to-end anti-spoofing with RawNet and AASIST," Interspeech, 2021. [9] H. M. B. Diniz et al., "RawGAT-ST: Robust generalization for spoofing detection," IEEE Access, vol. 10, 2022. [10] K. Sriskandaraja et al., "Robust audio deepfake detection through augmentation and multi-domain learning," ICASSP, 2023.