A Data-Driven Diffusion-based Approach for Audio Deepfake Explanations
This work addresses the problem of providing accurate explanations for audio deepfake detection, which is incremental as it builds on existing explainability techniques with a novel supervision approach.
The paper tackled the challenge of evaluating explainability techniques for audio deepfake detection by proposing a data-driven diffusion-based method to identify artifact regions, outperforming traditional methods like SHAP and LRP on datasets such as VocV4 and LibriSeVoc.
Evaluating explainability techniques, such as SHAP and LRP, in the context of audio deepfake detection is challenging due to lack of clear ground truth annotations. In the cases when we are able to obtain the ground truth, we find that these methods struggle to provide accurate explanations. In this work, we propose a novel data-driven approach to identify artifact regions in deepfake audio. We consider paired real and vocoded audio, and use the difference in time-frequency representation as the ground-truth explanation. The difference signal then serves as a supervision to train a diffusion model to expose the deepfake artifacts in a given vocoded audio. Experimental results on the VocV4 and LibriSeVoc datasets demonstrate that our method outperforms traditional explainability techniques, both qualitatively and quantitatively.