PyTorch code for our paper: Fooling Examples: Another Intriguing Property of Neural Networks.
Ming Zhang, Yongkang Chen, Cheng Qian.
torch>=1.7.0; torchvision>=0.8.1; tqdm>=4.31.1; pillow>=7.0.0; matplotlib>=3.2.2; numpy>=1.18.1;
The 1000 images from the NIPS 2017 ImageNet-Compatible dataset are provided in the folder dataset/images, along with their metadata in dataset/images.csv and dataset/imagenet_class_index.json.
gen_initial_images.py: Generate initial images, including random Gaussian noised, random uniform noised, all-white and all-black images.
eval_single.py: Generate fooling examples on a single model.
eval_ensemble.py: Generate fooling examples on an ensemble of models.
cal_succ_rate: Calculate the accuracy of fooling or normal examples.
plot_figures.py: Plot figures in experiments.
utils.py: Some necessary utility functions.