Official repo to reproduce the paper "How to Backdoor Diffusion Models?" published at CVPR 2023
Paper link: https://arxiv.org/abs/2212.05400
- Python 3.8.5
- PyTorch 1.10.1+cu11 or 1.11.0+cu102
Please run
bash install.sh
If you want to upload the experimental results to ``Weight And Bias, please log in with the API key.
wandb login --relogin --cloud <API Key>
- CIFAR10: It will be downloaded by HuggingFace
datasets
automatically - CelebA-HQ: Download the CelebA-HQ dataset and put the images under the folder
./datasets/celeba_hq_256
- CIFAR10: Create a folder,
measure/CIFAR10
, and copy the CIFAR10 images (in.jpg
format) into this folder before computing FID and MSE scores. - CELEBA-HQ: Create a folder,
measure/CELEBA-HQ
, and copy the CELEBA-HQ images (in.jpg
format) into this folder before computing FID and MSE scores.
I've uploaded all pre-trained backdoor diffusion models for BadDiffusion and VillanDiffusion on HuggingFace. Please feel free to download backdoored diffusion models from it.
Arguments
--project
: Project name for Wandb--mode
: Train or test the model, choice: 'train', 'resume', 'sampling`, 'measure', and 'train+measure'train
: Train the modelresume
: Resume the trainingmeasure
: Compute the FID and MSE score for the BadDiffusion from the saved checkpoint, the ground truth samples will be saved under the 'measure' folder automatically to compute the FID score.train+measure
: Train the model and compute the FID and MSE scoresampling
: Generate clean samples and backdoor targets from a saved checkpoint
--dataset
: Training dataset, choice: 'MNIST', 'CIFAR10', and 'CELEBA-HQ'--batch
: Training batch size. Note that the batch size must be able to divide 128 for the CIFAR10 dataset and 64 for the CelebA-HQ dataset.--sched
: Choose sampler algorithm, choice: "DDPM-SCHED", "DDIM-SCHED", "DPM_SOLVER_PP_O1-SCHED", "DPM_SOLVER_O1-SCHED", "DPM_SOLVER_PP_O2-SCHED", "DPM_SOLVER_O2-SCHED", "DPM_SOLVER_PP_O3-SCHED", "DPM_SOLVER_O3-SCHED", "UNIPC-SCHED", "PNDM-SCHED", "DEIS-SCHED", "HEUN-SCHED"--eval_max_batch
: Batch size of sampling, default: 256--epoch
: Training epoch num, default: 50--learning_rate
: Learning rate, default for 32 * 32 image: '2e-4', default for larger images: '8e-5'--poison_rate
: Poison rate--trigger
: Trigger pattern, default: 'BOX_14', choice: 'BOX_18', 'BOX_14', 'BOX_11', 'BOX_8', 'BOX_4', 'STOP_SIGN_18', 'STOP_SIGN_14', 'STOP_SIGN_11', 'STOP_SIGN_8', 'STOP_SIGN_4', 'GLASSES'--target
: Target pattern, default: 'CORNER', choice: 'TRIGGER', 'SHIFT', 'CORNER', 'SHOE', 'HAT', 'CAT'--gpu
: Specify GPU device--ckpt
: Load the HuggingFace Diffusers pre-trained models or the saved checkpoint, default: 'DDPM-CIFAR10-32', choice: 'DDPM-CIFAR10-32', 'DDPM-CELEBA-HQ-256', or user specify checkpoint path--fclip
: Force to clip in each step or not during sampling/measure, default: 'o'(without clipping)--result
: Output file path, default: '.'
For example, if we want to backdoor a DM pre-trained on CIFAR10 with Grey Box trigger and Hat target, we can use the following command
python baddiffusion.py --project default --mode train+measure --dataset CIFAR10 --batch 128 --epoch 50 --poison_rate 0.1 --trigger BOX_14 --target HAT --ckpt DDPM-CIFAR10-32 --fclip o -o --gpu 0
If we want to backdoor a DM pre-trained on Celeba-HQ with GLASSES trigger and CAT target, we can use the following command
python baddiffusion.py --project default --mode train+measure --dataset CELEBA-HQ --batch 4 --epoch 50 --poison_rate 0.1 --trigger GLASSES --target CAT --ckpt DDPM-CELEBA-HQ-256 --fclip o -o --gpu 0
If we want to measure the FID and MSE scores of a DM pre-trained on Celeba-HQ with GLASSES trigger and CAT target, we need to create a new folder measure/CIFAR10
under this repository folder and copy the training images (in .jpg
format) of CIFAR10 dataset into this folder. Then, we can use the following command
python baddiffusion.py --project default --mode measure --dataset CELEBA-HQ --eval_max_batch 256 --trigger GLASSES --target CAT --ckpt res_DDPM-CIFAR10-32_CIFAR10_ep50_c1.0_p0.1_BOX_14-HAT --fclip o -o --gpu 0
If we want to generate the clean samples and backdoor targets from a backdoored DM, use the following command Or simply generate the samples
python baddiffusion.py --project default --mode sampling --ckpt res_DDPM-CIFAR10-32_CIFAR10_ep50_c1.0_p0.1_BOX_14-HAT --fclip o --gpu 0
Arguments
--project
: Project name for Wandb--epoch
: Training epoch num, default: 50--learning_rate
: Learning rate, default: '1e-4'--perturb_budget
: Perturbation budget, default: '4.0'--gpu
: Specify GPU device--ckpt
: Load the HuggingFace Diffusers pre-trained models or the saved checkpoint--output_dir
: Output file path, default: '.'
If we want to detect the Trojan of the backdoored model trained in the last section, we can use the following command
python anp_defense.py --project default --epoch 5 --learning_rate 1e-4 --perturb_budget 4.0 --ckpt res_DDPM-CIFAR10-32_CIFAR10_ep50_c1.0_p0.1_BOX_14-HAT --gpu 0