Awesome
DFST (Keras + Tensorflow)
This is the official repository for AAAI'21 paper Deep Feature Space Trojan Attack of Neural Networks by Controlled Detoxification.<br> Refer to our paper in https://arxiv.org/abs/2012.11212.<br>
How to use this repository (Keras + Tensorflow Version)
Note that we provide example codes on VGG and CIFAR-10<br>
Dependences
Python3.6, tensorflow=1.13.1, keras=2.2.4, numpy, pickle, PIL.<br>
Prepare dataset
Create some folders: ./dataset, ./model, ./weights.<br>
<br>
Download the CIFAR-10 dataset and re-define it in the follwing format:<br>
- cifar_train['x_train'].shape = (50000, 32, 32, 3)
- cifar_test['x_test'].shape = (10000, 32, 32, 3)
- cifar_train['y_train'].shape = (50000, 1)
- cifar_test['y_test'].shape = (10000, 1)
Save the dictionaries in cifar_train and cifar_test file in ./dataset using pickle.<br>
pickle.dump(cifar_train, open('./dataset/cifar_train', 'wb'))<br>
pickle.dump(cifar_test, open('./dataset/cifar_test', 'wb'))<br>
<br>
Download sunrise images from Weather-Dataset into ./CycleGAN/sunrise.<br>
Train your own Cycle GAN as trigger generator
${\color{red}NEW}$ We provide an example CycleGAN in generator.h5.<br>
Type in cd CycleGAN.<br>
Train your own Cycle GAN python CycleGAN.py.<br>
Poison the training dataset python data_poisoning.py.
Perform DFST attack
Train a benign VGG as a classifier on CIFAR-10 python train.py.<br>
Inject the trigger using poisoned training data python retrain.py.<br>
Perform detoxification to force the model to learn deep features sh run.sh.<br>
Contact
Free to contact the author cheng535@purdue.edu.