As a follow-up to our work on computer-generated CSAM, we took a closer look at the training data used to train various generative models—most prominently, Stable Diffusion 1.5—to see to what degree CSAM itself might be present in the training data.

We used a combination of methods to determine this: perceptual hashing, cryptographic hashing, and k-nearest neighbors analysis using the image embeddings. Seeded from a small subset of the dataset, PhotoDNA identified hundreds of instances, the URLs of which which were reported to NCMEC.