Membership Inference Attacks Against Diffusion Models

A hyperparameter-free variant of SecMI that fuses several denoising timesteps instead of picking one by hand.

I re-implemented SecMI (ICML 2023) against DDPMs trained on CIFAR-100, then proposed a variant that learns to fuse several denoising timesteps instead of selecting a single timestep by hand.

Results

Metric SecMI Multi-timestep fusion
Attack AUC 0.971 0.984
TPR @ 1% FPR 0.519 0.642

I also tested whether an attacker trained on CIFAR-100 still works against a CIFAR-10 victim model.

The project is a compact look at how membership leakage shows up in diffusion models, and at how brittle hand-chosen attack hyperparameters can be. The code is on GitHub.