The code is for our NeurIPS 2019 paper: https://arxiv.org/abs/1910.04749
All the experiment we did in our paper are included in the main.py, class Experiments(). The functions are:
An overall PCA result following by several figures with different beta value is plotted
An overall PCA result following by several figures with different alpha value is plotted
Only the class been attacked exists valid triggers (You will see ASR~90%). Other classes' ASR are all very low.
test defense method performance on the cifar_10 dataset
test defense method performance on the cifar_100 dataset
test defense method on the cifar_10 dataset using ensemble model
test defense method on the cifar_100 dataset using ensemble model
baseline method on cifar-10
baseline method on cifar-100