What Does an Audio Deepfake Detector Focus on? A Study in the Time Domain

Explainability matters for real-world deployment: The paper demonstrates that adding explanations to audio deepfake detection models is essential for their practical application, as it reveals how these AI systems actually make decisions—a critical requirement for building trust in high-stakes scenarios.

Different explanation methods tell different stories: The researchers found that standard explanation techniques (Grad-CAM and SHAP) produce different results compared to their proposed relevancy-based method, highlighting that the choice of explanation tool significantly impacts what we think an AI model is "focusing on" when detecting deepfakes.

Small-scale studies can be misleading: A key finding is that conclusions drawn from analyzing limited audio samples don't necessarily hold up when tested on large datasets. This suggests previous research on audio deepfake detection may have reached incomplete or inaccurate conclusions about what features models actually use.

Models may not focus on what we expect: The investigation into what audio features matter most—such as speech versus non-speech segments, phonetic content, and voice characteristics—revealed surprising patterns that differ between small and large-scale analyses, suggesting deepfake detectors may rely on unexpected cues.

Rigorous evaluation is necessary: The paper emphasizes the importance of using multiple validation approaches (quantitative metrics and practical tests) rather than relying on single explanation methods, establishing a higher standard for verifying that AI explanations are actually trustworthy.