fmy266/pytorch-backdoor-unlearning

Pytorch implementation of backdoor unlearning.

Pytorch-Backdoor-Unlearning

This repository contains code for our paper (Backdoor Defense with Machine Unlearning) implemented in Pytorch. This paper was accepted by INFOCOM 2022.

Citing this work

Liu Y, Fan M, Chen C, et al. Backdoor Defense with Machine Unlearning[J]. arXiv preprint arXiv:2201.09538, 2022.

Requirements

• Python 3.9.2
• Pytorch 1.9
• Torchvision 0.1.8
• Trojanvision 1.0.8

Instructions

Quick Start

We provide a demo for a quick start, directly running the main.py in this repo.

python main.py

Firstly, you need to install the above packages and enuring your environment is consistent with our environment (different environment may cause some bugs or inconsistent results of our paper).

Besides, the data of MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 should stored (downloaded) in path "..//data".

Default attack method is Badnet with the $3 \times 3$ trigger adopted random position and initialization, where random position is more threat than fixed position.

Ideally, the accuracy and ASR of the obtained model trained with default setting (backdoored model) is about 85% and 100%.

Our defense method also is included in main.py and you will see the results reported in the paper.

If you want to reproduce more experiment results or used in your research, following the below instructions.

Training Backdoored Model

More information will quickly arrive.

Erasing Backdoor from the model using our method

More information will quickly arrive.

Other source code

Code for NAD (Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks) refer to this site.

Code for GAN-based defense (Defending Neural Backdoors via Generative Distribution Modeling) refer to this site.

Code for Fine-Pruning defense is integrated in the library TroJanZoo.

pytorch-backdoor-unlearning's Issues

gen_backdoor_mnist.py file does not exist in unlearning_backdoor.py, line 22, gen_backdoor_mnist

I find a file named "gen_backdoor.py" but didn't find a file named "gen_backdoor_mnist.py".
I change the code to:

# self.GM_model = gen_backdoor_mnist.GM(self.victimized_model, self.attack, search_param, ratio)
self.GM_model = gen_backdoor.GM(self.victimized_model, self.attack, search_param, ratio)

but still can't run the code correctly. 

```shell
Traceback (most recent call last):
  File "/home/user/code/backdoors/Pytorch-Backdoor-Unlearning/main.py", line 32, in <module>
    defense.run()
  File "/home/user/code/backdoors/Pytorch-Backdoor-Unlearning/unlearning_backdoor.py", line 30, in run
    self.unlearning_process()
  File "/home/user/code/backdoors/Pytorch-Backdoor-Unlearning/unlearning_backdoor.py", line 58, in unlearning_process
    for param1, param2 in self.victimized_model.parameters(), self.backup.parameters():
ValueError: too many values to unpack (expected 2)

lack of details about implementation and acc and asr problem

Hi, your work is really excellent and I am of great interest to it.
But your provided code is probably not enough to reproduce the result whose loss function and many setting is limited and I tried to implement it but finally get a bad result of it. could you give some more details or notes for the implements? thank you a lot anyway!