Skip to content

"From Trojan Horses to Castle Walls: Unveiling Bilateral Backdoor Effects in Diffusion Models" by Zhuoshi Pan*, Yuguang Yao*, Gaowen Liu, Bingquan Shen, H. Vicky Zhao, Ramana Rao Kompella, Sijia Liu

OPTML-Group/BiBadDiff

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

From Trojan Horses to Castle Walls: Unveiling Bilateral Backdoor Effects in Diffusion Models

Repository with code to reproduce the results for Bilateral Effects of Backdoored Diffusion

While state-of-the-art diffusion models (DMs) excel in image generation, concerns regarding their security persist. Earlier research highlighted DMs' vulnerability to backdoor attacks, but these studies placed stricter requirements than conventional methods like 'BadNets' in image classification. This is because the former necessitates modifications to the diffusion sampling and training procedures. Unlike the prior work, we investigate whether generating backdoor attacks in DMs can be as simple as BadNets, i.e., by only contaminating the training dataset without tampering the original diffusion process. In this more realistic backdoor setting, we uncover bilateral backdoor effects that not only serve an adversarial purpose (compromising the functionality of DMs) but also offer a defensive advantage (which can be leveraged for backdoor defense). Specifically, we find that a BadNets-like backdoor attack remains effective in DMs for producing incorrect images (misaligned with the intended text conditions), and thereby yielding incorrect predictions when DMs are used as classifiers. Meanwhile, backdoored DMs exhibit an increased ratio of backdoor triggers, a phenomenon we refer to as 'trigger amplification', among the generated images. We show that this latter insight can be used to enhance the detection of backdoor-poisoned training data. Even under a low backdoor poisoning ratio, studying the backdoor effects of DMs is also valuable for designing anti-backdoor image classifiers. Last but not least, we establish a meaningful linkage between backdoor attacks and the phenomenon of data replications by exploring DMs' inherent data memorization tendencies.

Teasor

The DDPM part is adapted from Classifier-free Diffusion Guidance.

The Stable-Diffusion part is based on stable-diffusion-1 which makes fine-tuning Stable-Diffusion model easier.

Stable-Diffusion on ImageNette and Caltech15

Environment

cd stable-diffusion
pip install -r requirements.txt

Data preparation

  1. ImageNette

Get the full size version of imagenette from here and unzip to "data/imagenette",

cd data/imagenette
wget https://s3.amazonaws.com/fast-ai-imageclas/imagenette2.tgz
tar -zxvf imagenette2.tgz

and produce the Badnets-like version.

python badnets_imagenette.py
  1. Caltech15

Get the 15 classes subset of Caltech256 from here and unzip to "data/caltech",

cd data/caltech
wget https://drive.google.com/file/d/1Li9UAy6QyNDRGP_CAhNrGCAfocxjED3-/view?usp=sharing

and produce the Badnets-like version.

python badnets_caltech15.py

Model preparetion

The pretrained stable diffusion model can be downloaded from huggingface.

Train

cd stable-diffusion 
bash scripts/train.sh

SD model will be saved at logs/imagenette/experiment_name/checkpoints

Sample

cd stable-diffusion 
bash scripts/sample.sh

Samples model will be saved at outputs/imagenette/experiment_name

Evaluation

  1. train classifier on ImageNette
cd classifier
bash scripts/train_imagenette.sh

This step trains a clean classifier (ResNet-50) on ImageNette and a backdoored classifier on the Badnets-like poisoned ImageNette. They are saved at model_ckpt/imagenette

  1. run clean classifier on generated data.
cd classifier
python eval_imagenette.py

The generated prediction will be saved at stable-diffusion/result_clf/imagenette

  1. eval trigger ratio and the ratio of generations mismatching their prompts
cd stable-diffusion 
python eval_tgr_imagenette.py

defense backdoor by training on generation

cd classifier
bash scripts/train_on_gen_imagenette.sh

Results save to classifier/result/imagenette_on_gen

DDPM on CIFAR10

Environment

cd ddpm
conda env create -f environment.yml

Diffusion classifier on CIFAR10

Environment

cd edm 
conda env create -f environment.yml -n edm

Data preparation

Generate backdoored cifar10 by CognitiveDistillation, save to .npz file. Then transform to .zip format by

bash scripts/data.sh

Run

bash scripts/train.sh
bash scripts/pred.sh

About

"From Trojan Horses to Castle Walls: Unveiling Bilateral Backdoor Effects in Diffusion Models" by Zhuoshi Pan*, Yuguang Yao*, Gaowen Liu, Bingquan Shen, H. Vicky Zhao, Ramana Rao Kompella, Sijia Liu

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages