Defending Against Backdoor Attacks by Layer-wise Feature Analysis

This repository contains PyTorch implementation of the paper: Defending Against Backdoor Attacks by Layer-wise Feature Analysis that has been accepted to the Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD 2023).

Paper

Defending Against Backdoor Attacks by Layer-wise Feature Analysis

Content

The repository contains one jupyter notebook (Ours_CIFAR10-ResNet18.IPYNB) that contains code and instructions on how to re-produce the experiments reported in the paper for the benchmark under the Input-aware Dynamic backdoor attack (IAD). Also, models modified to extract intermediate features are available in the folder my_models.

Datasets

CIFAR10 will be automatically downloaded. However, GTSRB can be manually downloaded using this link. After downloading GTSRB, please save in the folder named 'data' with two subfolders each 'train' and 'test.

Poisoned Datasets and Pretrained Attacked DNNs

Please download poisoned datasets via this link (https://www.dropbox.com/sh/02g7ys181u7yhlx/AABZSgKSYxgDF8DtystO31Sla?dl=0) and save them in the folder data/poisoned_testsets. Also, please download pretrained attacked DNNs via this link (https://www.dropbox.com/sh/ilwdvone2fl5c2d/AADpmdjgKKq175Nj6Oj-bK8ma?dl=0) and save them in the folder checkpoints.

Dependencies

Required packages and libraries are in requirements.txt

Citation

Jebreel, N. M., Domingo-Ferrer, J., & Li, Y. (2023, May). Defending against backdoor attacks by layer-wise feature analysis. In Pacific-Asia Conference on Knowledge Discovery and Data Mining (pp. 428-440). Cham: Springer Nature Switzerland.