A Data-Driven Diffusion-based Approach for Audio Deepfake Explanations

TL;DR

This paper introduces a data-driven diffusion-based method to identify artifacts in deepfake audio, outperforming traditional explainability techniques by leveraging paired real and vocoded audio for supervision.

Contribution

A novel diffusion model approach that uses paired real and vocoded audio to generate accurate explanations of deepfake artifacts in audio.

Findings

Outperforms traditional explainability methods in accuracy

Effective on VocV4 and LibriSeVoc datasets

Provides clear artifact localization in deepfake audio

Abstract

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.