Defense Against Backdoor Attacks in SSL

Introduction

This repository contains code for our submission to ICLR workshop 24. We implement backdoor-poisoning attacks and evaluate our defense against these attacks on Image Classification Task. Attacks implemented:

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

Defenses implemented:

Gaussian Blur (Invariance-Equivariance)
Luminance (Exploiting evasiveness)

Datasets implemented:

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

We show a successful and partly generalizable defense against backdoor attacks in SSL and lay the theoretical foundation for defense against backdoor attacks in Semantic Segmentation and Object Detection tasks.

Installation and Requirements

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
```

Running The Code

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

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.

To Do

Implement ImageNet 100
Remove unnecessary command line arguments and refactor code.

New Methods

JEPA
MoCo v2

New Datasets / Benchmarks

ImageNet
https://robustbench.github.io/

Aryan-Satpathy/Backdoor