Membership Inference Attacks for Face Images Against Fine-Tuned Latent Diffusion Models

Summary

Generative image models are trained on enormous, often opaque image corpora, which raises a sharp question: can an outside observer determine whether a specific set of face images was part of the training data? We frame this as a membership inference attack against a fine-tuned Latent Diffusion Model and probe how reliably the attack succeeds across realistic fine-tuning regimes.

Why it matters

We want to show that it matters when big tech corporations scrape the internet and inadvertently take images of your image, by reverse engineering the information leakage through the LDM we want to show that it is possible to identify if a person (or in this case a group of images) have been part of the training data.