Neural Attention Distillation

This is an implementation demo of the ICLR 2021 paper Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks in PyTorch.

NAD: Quick start with pretrained model

We have already uploaded the all2one pretrained backdoor student model(i.e. gridTrigger WRN-16-1, target label 5) and the clean teacher model(i.e. WRN-16-1) in the path of ./weight/s_net and ./weight/t_net respectively.

For evaluating the performance of NAD, you can easily run command:

$ python main.py 

where the default parameters are shown in config.py.

The trained model will be saved at the path weight/erasing_net/<s_name>.tar

Please carefully read the main.py and configs.py, then change the parameters for your experiment.

Erasing Results on BadNets

Training your own backdoored model

We have provided a DatasetBD Class in data_loader.py for generating training set of different backdoor attacks.

For implementing backdoor attack(e.g. GridTrigger attack), you can run the below command:

$ python train_badnet.py

This command will train the backdoored model and print clean accuracies and attack rate. You can also select the other backdoor triggers reported in the paper.

Please carefully read the train_badnet.py and configs.py, then change the parameters for your experiment.

Other source of backdoor attacks

Attack

CL: Clean-label backdoor attacks

• Paper
• Can be modified from this pytorch implementation

SIG: A New Backdoor Attack in CNNS by Training Set Corruption Without Label Poisoning

• Paper

## reference code
def plant_sin_trigger(img, delta=20, f=6, debug=False):
    """
    Implement paper:
    > Barni, M., Kallas, K., & Tondi, B. (2019).
    > A new Backdoor Attack in CNNs by training set corruption without label poisoning.
    > arXiv preprint arXiv:1902.11237
    superimposed sinusoidal backdoor signal with default parameters
    """
    alpha = 0.2
    img = np.float32(img)
    pattern = np.zeros_like(img)
    m = pattern.shape[1]
    for i in range(img.shape[0]):
        for j in range(img.shape[1]):
            for k in range(img.shape[2]):
                pattern[i, j] = delta * np.sin(2 * np.pi * j * f / m)

    img = alpha * np.uint32(img) + (1 - alpha) * pattern
    img = np.uint8(np.clip(img, 0, 255))

    #     if debug:
    #         cv2.imshow('planted image', img)
    #         cv2.waitKey()

    return img

Refool: Reflection Backdoor: A Natural Backdoor Attack on Deep Neural Networks

• Paper
• Code
• Project

Defense

MCR: Bridging Mode Connectivity in Loss Landscapes and Adversarial Robustness

• Paper
• Pytorch implementation

Fine-tuning & Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks

• Paper
• Pytorch implementation1
• Pytorch implementation2

Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks

• Paper
• Tensorflow implementation
• Pytorch implementation1
• Pytorch implementation2

STRIP: A Defence Against Trojan Attacks on Deep Neural Networks

• Paper
• Pytorch implementation1
• Pytorch implementation2

Library

Note: TrojanZoo provides a universal pytorch platform to conduct security researches (especially backdoor attacks/defenses) of image classification in deep learning.

Backdoors 101 — is a PyTorch framework for state-of-the-art backdoor defenses and attacks on deep learning models.

• trojanzoo
• backdoors101

References

If you find this code is useful for your research, please cite our paper

@inproceedings{li2021neural,
  title={Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks},
  author={Li, Yige and Lyu, Xixiang and Koren, Nodens and Lyu, Lingjuan and Li, Bo and Ma, Xingjun},
  booktitle={ICLR},
  year={2021}
}

Contacts

If you have any questions, leave a message below with GitHub.