Code Repo for the NeurIPS 2023 paper "VillanDiffusion: A Unified Backdoor Attack Framework for Diffusion Models"
- 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 put the images of CIFAR10 under the folder. - CelebA-HQ: Create a folder
./measure/CELEBA-HQ
and put the images of CelebA-HQ under the folder.
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.
Note that for training backdoor score-based models, you need to download the pre-trained clean model from HuggingFace and put it under the working directory.
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 VillanDiffusion 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
--task
: The task for mode:sampling
andmeasure
. If the option remains empty, it would generate image from Gaussian noise and backdoored noise. Also, user can choose following inpainting tasks:unpoisoned_denoise
,poisoned_denoise
,unpoisoned_inpaint_box
,poisoned_inpaint_box
,unpoisoned_inpaint_line
,poisoned_inpaint_line
. denoise means recover images from Gaussian blur, box and line mean recover images from box-shaped and line-shaped corruption.--sched
: Sampling algorithms for the diffusion models. Samplers for the DDPM areDDPM-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
.SCORE-SDE-VE-SCHED
is used by score-based models.--solver_type
: Backdoor for the ODE or SDE samplers. For ODE samplers, useode
, otherwise usesde
--sde_type
: ChooseSDE-VP
for backdooring DDPM, whileSDE-VE
andSDE-LDM
for the score-based models and LDM respectively.--infer_steps
: Sampling steps of the specified sampler. We recommend 50 steps forPNDM
,HEUN
,LMSD
, andDDIM
, otherwise 20 steps.epoch
: Training epochs--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.--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_14
,STOP_SIGN_14
,BOX_18
,STOP_SIGN_18
,BOX_11
,STOP_SIGN_11
,BOX_8
,STOP_SIGN_8
,BOX_4
,STOP_SIGN_4
, andGLASSES
.--target
: Target pattern, default: 'CORNER', choice:NOSHIFT
,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
,LDM-CELEBA-HQ-256
, or user specify checkpoint path--fclip
: Force to clip in each step or not during sampling/measure, default: 'o'(without clipping)--output_dir
: 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 VillanDiffusion.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 generate the clean samples and backdoor targets from a backdoored DM, use the following command to generate the samples
python VillanDiffusion.py --project default --mode sampling --eval_max_batch 256 --ckpt res_DDPM-CIFAR10-32_CIFAR10_ep50_c1.0_p0.1_BOX_14-HAT --fclip o --gpu 0
To train LDM models, you can run following command or run python run_ldm_celeba_hq_script.py
.
python VillanDiffusion.py --postfix new-set --project default --mode train --dataset CELEBA-HQ-LATENT --dataset_load_mode NONE --sde_type SDE-LDM --learning_rate 0.0002 --sched UNIPC-SCHED --infer_steps 20 --batch 16 --epoch 2000 --clean_rate 1 --poison_rate 0.9 --trigger GLASSES --target CAT --solver_type ode --psi 1 --vp_scale 1.0 --ve_scale 1.0 --ckpt LDM-CELEBA-HQ-256 --fclip o --save_image_epochs 1 --save_model_epochs 1 --result exp_GenBadDiffusion_LDM_BadDiff_ODE -o --gpu 0
To train Score-Based models, you can run following command or run python run_score-basde_model_script.py
.
Note that for training backdoor score-based models, you need to download the pre-trained clean model from HuggingFace and put it under the working directory.
python VillanDiffusion.py --postfix flex_new-set --project default --mode train --learning_rate 2e-05 --dataset CIFAR10 --sde_type SDE-VE --batch 128 --epoch 30 --clean_rate 1.0 --poison_rate 0.98 --dataset_load_mode FIXED --trigger STOP_SIGN_14 --target HAT --solver_type sde --psi 0 --vp_scale 1.0 --ve_scale 1.0 --ckpt NCSN_CIFAR10_my --fclip o --save_image_epochs 5 --save_model_epochs 5 --result exp_GenBadDiffusion_NCSNPP_CIFAR10_TrojDiff_SDE_FLEX -o --R_trigger_only --gpu 0
To measure with inpainting task, you can run following instruction or run python run_measure_inpaint.py
python VillanDiffusion.py --project default --mode measure --task poisoned_denoise --sched UNIPC-SCHED --infer_steps 20 --infer_start 10 --ckpt /work/u2941379/workspace/backdoor_diffusion/res_DDPM-CIFAR10-32_CIFAR10_ep100_ode_c1.0_p0.2_SM_STOP_SIGN-BOX_psi1.0_lr0.0002_vp1.0_ve1.0_new-set-1_test --fclip o --gpu 0
--pretrained_model_name_or_path
: Specify the backdoor model. We recommend to useCompVis/stable-diffusion-v1-4
.--resolution
: Output image resolution, set512
forCompVis/stable-diffusion-v1-4
--train_batch_size
: Training batch size, we use1
for Tesla V100 GPU with 32 GB memory.--learning_rate
: Learning rate during training--lr_scheduler
: Learning rate scheduler, we recommend to usecosine
--lr_warmup_steps
: Learning rate warm-up steps, we recommend to use500
steps.--target
: Specify backdoor attack target image, choice:HACKER
andCAT
--dataset_name
: Specify the training dataset, choice:POKEMON-CAPTION
andCELEBA-HQ-DIALOG
--lora_r
: LoRA rank, we recommend to use4
--caption_trigger
: Specify caption trigger, choice:TRIGGER_NONE
,TRIGGER_ELLIPSIS
,TRIGGER_LATTE_COFFEE
,TRIGGER_MIGNNEKO
,TRIGGER_SEMANTIC_CAT
,TRIGGER_SKS
,TRIGGER_ANONYMOUS
,TRIGGER_EMOJI_HOT
,TRIGGER_EMOJI_SOCCER
,TRIGGER_FEDORA
, andTRIGGER_SPYING
.--dir
: Output folder--gradient_accumulation_steps
: Gradient accumulation steps, default:1
--max_train_steps
: Training steps, recommended:50000
--checkpointing_steps
: Checkpointing every X step--enable_backdoor
: Enable backdoor attack--use_lora
: Enable LoRA--with_backdoor_prior_preservation
: Enable regularization of the clean dataset--gpu
: Specify GPU device
For example, if we want to backdoor Stable Diffusion v1-4 with the trigger: "latte coffee" and target: Hacker, we can use the following command.
python viallanDiffusion_conditional.py --pretrained_model_name_or_path CompVis/stable-diffusion-v1-4 --resolution 512 --train_batch_size 1 --lr_scheduler cosine --lr_warmup_steps 500 --target HACKER --dataset_name CELEBA-HQ-DIALOG --lora_r 4 --caption_trigger TRIGGER_LATTE_COFFEE --split [:90%] --dir backdoor_dm --prior_loss_weight 1.0 --learning_rate 1e-4 --gradient_accumulation_steps 1 --max_train_steps 50000 --checkpointing_steps 5000 --enable_backdoor --use_lora --with_backdoor_prior_preservation --gradient_checkpointing --gpu 0
--max_batch_n
: Sampling batch size--sched
: Specify the sampler, choice:DPM_SOLVER_PP_O2_SCHED
andNone
--num_inference_steps
: Number of the sampling steps, default: 25--infer_start
: Start from which step--base_path
: Sampling from the model under the specified path--ckpt_step
: Checkpointing every X step--gpu
: Specify GPU device
For example, if we want to generate samples from the model under the folder: res_CELEBA-HQ-DIALOG_NONE-TRIGGER_LATTE_COFFEE-HACKER_pr0.0_ca0_caw1.0_rctp0_lr0.0001_step50000_prior1.0_lora4, we can use the following command.
python sampling.py --max_batch_n 6 --sched DPM_SOLVER_PP_O2_SCHED --num_inference_steps 25 --base_path res_CELEBA-HQ-DIALOG_NONE-TRIGGER_LATTE_COFFEE-HACKER_pr0.0_ca0_caw1.0_rctp0_lr0.0001_step50000_prior1.0_lora4 --ckpt_step -1 --gpu 0