Skip to content
No description, website, or topics provided.
Branch: master
Clone or download

Latest commit

Fetching latest commit…
Cannot retrieve the latest commit at this time.


Type Name Latest commit message Commit time
Failed to load latest commit information.

Adversarial Example Detection with Critical Data Routing Paths


pytorch == 0.3.1 python == 3.5 sklearn == 0.22

Data Preparation

python --data_dir IMAGENET_DATA_DIR/train --dump_path data/train_images_list.pkl
python --data_dir IMAGENET_DATA_DIR/val --dump_path data/val_images_list.pkl

Adversarial Example Detection

python --data IMAGENET_DATA_DIR -a ARCH --gpu GPU_ID

where ARCH denotes the attacking network (AlexNet, VGG16), GPU_ID is the available gpu device number. For ResNet50, run the command

python --data IMAGENET_DATA_DIR -a resnet50 --gpu GPU_ID

Current setting is one training sample and one testing sample from each class to extract the CDRP used for adversarial example detection. You can adjust the sample number from each class by

python --data IMAGENET_DATA_DIR -a ARCH --train_num_per_class 5 --test_num_per_class 1 --gpu GPU_ID
  • Note: we have improved the codes after CVPR paper is published, and current settings can achieve 0.9+ AUROC value.


	title={Interpret Neural Networks by Identifying Critical Data Routing Paths},
	author={Wang, Yulong and Su, Hang and Zhang, Bo and Hu, Xiaolin},
	booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
	publisher = {IEEE},
	address={Salt Lake City, USA}
You can’t perform that action at this time.