Ben Pinhasov and the team introduce a defense mechanism against adversarial attacks targeting deepfake detectors by using XAI to produce interpretability maps. This method, which maintains the detector’s performance, represents the decision-making factors of AI models and serves as a crucial step towards understanding and guarding against potential vulnerabilities.

Highlights:

• XAI for interpretability maps generation.
• Pretrained feature extractor used on input and XAI images.
• Classifier trained with extracted feature embeddings.
• Enhances understanding of adversarial attacks without performance degradation.

By explaining AI decisions through visualizations, this study demonstrates potential pathways for more robust deepfake detection in the future. Delve into the details.