Squint Hard Enough - Attacking Perceptual Hashing with Adversarial Machine Learning

Explore a 15-minute conference talk from USENIX Security '23 that delves into the vulnerabilities of perceptual hashing algorithms used in online content moderation systems. Learn about the potential risks of incorporating these hash-based matching systems into client-side and end-to-end encrypted (E2EE) communications. Discover how researchers developed threat models and attacks against widely deployed algorithms like PhotoDNA and PDQ, demonstrating the possibility of generating targeted second-preimage attacks and detection avoidance techniques. Gain insights into the implications of these findings for the robustness of existing perceptual hash functions in adversarial settings, particularly in the context of determining confidentiality guarantees for user content.