This is the official implementation of our paper Not All Samples Are Born Equal: Towards Effective Clean-Label Backdoor Attacks, accepted by Pattern Recognition (2023). This research project is developed based on Python 3 and Pytorch, created by Yinghua Gao and Yiming Li.
- python = 3.7.13
- numpy = 1.21.5
- torch = 1.12.1
- torchvision = 0.2.1
Step 1: Calculate three metrics (loss value, gradient norm and forgetting events)
CUDA_VISIBLE_DEVICES=0 python cal_metric.py --output_dir save_metric
Step 2: Train backdoored model with different sample selection methods
CUDA_VISIBLE_DEVICES=0 python train_backdoor.py --output_dir save_metric --result_dir save_res --selection forget --backdoor_type badnets --y_target 0 --select_epoch 10
--select epoch specifies the epoch used to calculate the statistics. --output_dir must be the same with the one used in cal_metric.py
If this work or our codes are useful for your research, please kindly cite our paper as follows.
@article{gao2023not,
title={Not all samples are born equal: Towards effective clean-label backdoor attacks},
author={Gao, Yinghua and Li, Yiming and Zhu, Linghui and Wu, Dongxian and Jiang, Yong and Xia, Shu-Tao},
journal={Pattern Recognition},
volume={139},
pages={109512},
year={2023},
publisher={Elsevier}
}
Our implementation is based on the following projects. We sincerely thank the authors for releasing their codes.