# XAI-Based Detection of Adversarial Attacks on Deepfake Detectors

**Authors:** Ben Pinhasov, Raz Lapid, Rony Ohayon, Moshe Sipper, Yehudit Aperstein  
**arXiv:** [2403.02955](https://arxiv.org/abs/2403.02955)  
**Date:** 2024-03-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.02955v1/

## Key points

- Authors introduce a methodology using eXplainable Artificial Intelligence (XAI) to detect adversarial attacks on deepfake detectors
- Approach addresses compromised effectiveness in detecting deepfakes due to targeted adversarial attacks
- Method involves generating interpretability maps using XAI, extracting feature embeddings, and training a classifier
- Enhances detection and provides insights into potential attacks and vulnerabilities without performance impact
- Results show promising outcomes and suggest advancements in deepfake detection mechanisms

## Summary

In their paper titled "XAI-Based Detection of Adversarial Attacks on Deepfake Detectors," authors Ben Pinhasov, Raz Lapid, Rony Ohayon, Moshe Sipper, and Yehudit Aperstein introduce a groundbreaking methodology that utilizes eXplainable Artificial Intelligence (XAI) to identify adversarial attacks on deepfake detectors. This approach addresses the critical issue of compromised effectiveness in detecting deepfakes due to targeted adversarial attacks. By generating interpretability maps using XAI for a given method and extracting feature embeddings from input images and their corresponding XAI images, the authors train a simple yet effective classifier. This not only enhances detection but also provides valuable insights into potential attacks and vulnerabilities within the system without affecting performance. The results demonstrate promising outcomes and suggest future advancements in deepfake detection mechanisms. Overall, this study contributes significantly to safeguarding deepfake detectors through innovative XAI techniques and is poised to spark essential discussions within the community for further research in enhancing cybersecurity measures against malicious manipulations like deepfakes.