guanzihan / deep-backdoor-attack

Deep-Backdoor-Attack

Pipeline

Deep Backdoor Attack (DBA) aims to generate deep-synchronized triggers that can alleviate the early-fitting phenomenon

Playing with DBA

We provide a self-trained clean model ResNet34-clean-91.tar in the repository, which can be used to generate deep-synchronized trigger patterns

To play with the codes, simple run the following command to launch the attack

Step 1: Trigger Generation

python trigger_generation.py --cuda 1 --batch_size 1 train_or_test train
python trigger_generation.py --cuda 1 --batch_size 1 train_or_test test

This command will generate a series of trigger patterns by activating a specified neuron in the clean model. Then the trigger patterns will be superimposed on the clean images to construct poisoned samples.

Step 2: Backdoor Injection

python backdoor_train.py --cuda 1 --trigger_type "dba" --load_fixed_data 1

This command will start the standard backdoor training process, where the neural network model is trained on a mixture of poisoned samples and clean samples.

Results

The following table presents the performance reported in our paper