TDteach/Defending-Neural-Backdoors-via-Generative-Distribution-Modeling

The code is for our NeurIPS 2019 paper: https://arxiv.org/abs/1910.04749

Defending-Neural-Backdoors-via-Generative-Distribution-Modeling

The code is for our NeurIPS 2019 paper: https://arxiv.org/abs/1910.04749

Experiment

All the experiment we did in our paper are included in the main.py, class Experiments(). The functions are:

test_contour()

An overall PCA result following by several figures with different beta value is plotted

test_alpha()

An overall PCA result following by several figures with different alpha value is plotted

detect_cifar()

Only the class been attacked exists valid triggers (You will see ASR~90%). Other classes' ASR are all very low.

defend_cifar_10_all()

test defense method performance on the cifar_10 dataset

defend_cifar_100_all()

test defense method performance on the cifar_100 dataset

defend_cifar_10_all_ensemble()

test defense method on the cifar_10 dataset using ensemble model

defend_cifar_100_all_ensemble()

test defense method on the cifar_100 dataset using ensemble model

point_defend_cifar_10()

baseline method on cifar-10

point_defend_cifar_100()

baseline method on cifar-100