Paper in USENIX Security

The paper “CERTPHASH: Towards Certified Perceptual Hashing via Robust Training” by Yuchen Yang (Johns Hopkins University), Qichang Liu (Tsinghua University), Christopher Brix, Huan Zhang (UIUC), and Yinzhi Cao (Johns Hopkins University) has been accepted for publication in the 34th USENIX Security Symposium. The paper proposes a novel technique to train neural networks to perform perceptual hashing (used to filter illicit images). Crucially, the networks are trained to be robust under evasion or collision attacks and allow for provable guarantees on that robustness.