Towards Quantitative Evaluation of Explainable AI Methods for Deepfake Detection

Explanation methods for AI systems can be rigorously tested: The paper demonstrates that we can objectively measure how well an explanation method actually identifies the important parts of an image that influence a deepfake detector's decision, rather than simply assuming explanations are accurate.

LIME outperforms other explanation methods for deepfake detection: Among five explanation methods tested, LIME (Local Interpretable Model-agnostic Explanations) proved most effective at pinpointing which regions of fake images the detector was actually using to make its decisions.

Adversarial testing reveals explanation quality: The framework uses adversarial attacks to validate explanations—if an explanation method correctly identifies influential regions, modifying those regions should significantly reduce the detector's confidence. This provides concrete evidence of explanation reliability.

Transparency in deepfake detection matters for trust: By quantitatively evaluating which explanation methods work best, this research supports the development of more trustworthy deepfake detection systems that can show users why an image was flagged as fake, rather than just providing a yes/no answer.

One-size-fits-all explanations don't work: The significant performance differences between explanation methods suggest that organizations deploying deepfake detectors should carefully validate their chosen explanation approach rather than assuming any explanation method will adequately support human decision-making.