Our paper was accepted at ICASSP 2025 (MLSP track)!

This repository contains the official code implementation for our paper Towards Adversarial Robustness And Backdoor Mitigation in SSL. If you find this repo useful for your work, please cite our paper.

We implement backdoor-poisoning attacks and evaluate our defense against these attacks on Image Classification Task. Backdoor attacks supported:

• CTRL: ICCV 2023
• FIBA: CVPR 2022
• HTBA: CVPR 2022

Datasets supported:

• CIFAR 10
• CIFAR 100
• IMAGENET 100 (100 class subset of IMAGENET)

• Download the repository from anonymous4openscience.
• Make a virtual environment (optional) (recommended)

  virtualenv <env_name>
  source <env_name>/bin/activate

• Install necessary libraries

  pip install -r requirements.txt

We provide a bash script to run our program with appropriate command-line arguments.

• Give permission

chmod +x run.sh

• Call main_train.py

bash run.sh <--args values>

Call bash run.sh --help if unsure about the arguments, available options or their meaning.

Results of all experiments are saved in a folder named saves. Each experiment will create a folder named <job name> set by run.sh. Each experiment folder contains model state-dicts and optimizer states saved every 100 epoch, and a tfenvent file containing tensorboard log. To view training progress and compare training curves:

tensorboard --logdir=saves

When repeating the same experiment with different hyperparameters, use --suffix option in run.sh to prevent overwriting log of previous experiment.

• JEPA
• MoCo v2

• ImageNet
• https://robustbench.github.io/

This code has a GPL-style license.

@misc{satpathy2024adversarialrobustnessbackdoormitigation,
      title={Towards Adversarial Robustness And Backdoor Mitigation in SSL}, 
      author={Aryan Satpathy and Nilaksh Singh and Dhruva Rajwade and Somesh Kumar},
      year={2024},
      eprint={2403.15918},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2403.15918}, 
}