Enhancing Membership Inference Attacks on Diffusion Models from a Frequency-Domain Perspective

Mirror of the original repository: poetic2/FreMIA Code for the ICML 2026 paper Enhancing Membership Inference Attacks on Diffusion Models from a Frequency-Domain Perspective.

Install

Create a Python environment:

conda create -n fremia python=3.10 -y
conda activate fremia

Install dependencies:

pip install -r requirements.txt

Project Layout

DDIM/                 DDIM attack implementation
Stable_Diffusion/     Stable Diffusion attack implementation
train/                DDPM training and evaluation code

Run DDIM Attack

cd DDIM
python attack.py \
  --checkpoint ../train/experiments/TINY-IN/checkpoint.pt \
  --dataset TINY-IN \
  --attacker-name naive \
  --filter 0 \
  --t 5 \
  --s 0.2

Choices:

--dataset: TINY-IN, CIFAR100, STL10-U
--attacker-name: naive, pia, sec

Run Stable Diffusion Attack

cd Stable_Diffusion
python attack.py \
  --attack naive \
  --dataset coco \
  --diff-path /path/to/stable-diffusion-checkpoint \
  --dataset-root /path/to/datasets \
  --filter 1 \
  --t 5 \
  --s 0.2

Choices:

--attack: naive, pia, sec, clid
--dataset: pokemon, coco, flickr

Notes

• Fourier filtering is controlled by --filter, --t, and --s.
• To fine-tune a Stable Diffusion model, follow the official Diffusers train_text_to_image.py example.
• To train DDPM models, follow the training code and configuration under train/.