Dynamic Backdoor Trigger Elimination using Model Contrastive Learning and Triple Marginal Loss

This repository contains an implementation for eliminating backdoor triggers embedded in images, particularly addressing poison label attacks such as Trojan, BadNets, and Blend. The solution is built upon Model Contrastive Learning and Triple Marginal Loss techniques. Additionally, the code implements a Dynamic Patching algorithm, enabling the model to train with different trigger patterns at runtime for enhanced robustness.

Features:

• Model Contrastive Learning: Utilizes contrastive learning techniques to enhance the model's ability to discriminate between clean and poisoned data.
• Triple Marginal Loss (TML): Implements TML for training robust models against poison label attacks. TML effectively minimizes the impact of backdoor triggers during model training.
• Dynamic Patching Algorithm: Incorporates a dynamic patching algorithm, enabling the model to adapt to different trigger patterns during runtime. This enhances the model's resilience against evolving attack strategies.
• K-Means Clustering for Pseudo-Label Generation: Employs K-Means clustering to generate pseudo-labels for training with TML. This helps in effectively identifying and mitigating the influence of poisoned data during training.

MCL++: Quick start with pretrained model

We have already uploaded the all2one pretrained backdoor model(i.e. gridTrigger WRN-16-1, target label 5).

Install the requirements using the following command:

$ pip install -r requirements.txt

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

$ python MCL++_defense.py

where the default parameters are shown in config.py.

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

Please carefully read the MCL++_defense.py and configs.py, then change the parameters for your experiment.

---

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 like Grid, Square etc as defined in the data_loader.py class.

Acknowledgements

Much of the code in this repository was adapted from code in this paper by Zhihao et al.

Other source of backdoor attacks

Attack

CL: Clean-label backdoor attacks

• Paper
• pytorch implementation

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

• Paper

WaNet: WaNet-Imperceptible Warping-based Backdoor Attack.

• Paper
• pytorch implementation

Defense

Fine-tuning && Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks

• Paper
• Pytorch implementation

I-BAU: Adversarial Unlearning of Backdoors via Implicit Hypergradient.

• Paper
• Pytorch implementation

Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks

• Paper
• Tensorflow implementation
• 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.

BackdoorBox — is a Python toolbox for backdoor attacks and defenses.

• TrojanZoo
• backdoors101
• BackdoorBox

Contacts

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