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