Description
PhotoDNA is a widely deployed perceptual hash function used for the detection of illicit content such as Child Sexual Abuse Material (CSAM). In this talk, I will present our paper introducing the first mathematical description of Alleged PhotoDNA, a function that reproduces the outputs of PhotoDNA.
Our analysis reveals several structural weaknesses: the function is piece-wise linear and differentiable, the hash values depend only on the sum of RGB values per pixel, and trivial collisions (e.g., all-zero hash values) can be constructed. Building on these observations, we show that gradient-based optimization and quadratic programming can be used to generate exact collisions, second preimages, and high-quality near-collisions. We also demonstrate that it is possible to reconstruct coarse image structures from hash values, challenging the claim that PhotoDNA is irreversible.
Beyond these attacks, our results show that perceptually identical images can easily be modified to evade detection, while false positives can also be constructed. Our experiments, conducted on a wide range of images, achieve near-100% success rates in seconds to minutes on a standard laptop.
These results also have broader implications, which this talk will discuss, in particular for large-scale deployment scenarios such as client-side scanning. Overall, this work raises serious concerns about the reliability of PhotoDNA for detecting illicit content in practice.
Practical infos
Next sessions
-
Towards More Secure Large Language Models
Speaker : Raouf Kerkouche - Inria Lille
Large Language Models (LLMs) have achieved considerable success and are now widely used across multiple domains, highlighting their transformative impact on both technology and society. However, this widespread adoption also exposes LLMs to numerous security threats that can alter model behavior or degrade overall performance. To mitigate these threats, most research has focused on alignment[…]-
Machine learning
-