Fisher Information guided Purification against Backdoor Attacks

If you like our project, please give us a star ⭐ on GitHub for the latest update.

Paper

Smoothness Analysis of Backdoor Models

😮 Highlights

💡 Fast and Effective Backdoor Purification

• Novel Perspective for Backdoor Analysis
• Novel Algorithm for Backdoor Purification
• Extensive Experimental Evaluation on Multiple Tasks

🚩 Updates

Welcome to watch 👀 this repository for the latest updates.

✅ [2023.04.07] : FIP is accepted to ACM CCS'2024

🛠️ Methodology

Main Overview

PyTorch Implementation

Create Conda Environment

• Install Anaconda and create an environment

   conda create -n fip-env python=3.10
   conda activate fip-env

• After creating a virtual environment, run

   pip install -r requirements.txt

Download the Datasets

• Image Classification ( CIFAR10, GTSRB, Tiny-ImageNet, ImageNet)

• Action Recognition (UCF-101, HMDB51)

• Object Detection (MS-COCO, Pascal VOC)

• 3D Point Cloud Classifier (ModelNet40)

• Natural Language Processing (NLP) (WMT2014 En-De, OpenSubtitles2012)

Create Benign and Backdoor Models

For Cifar10

• To train a benign model

python train_backdoor_cifar.py --poison-type benign --output-dir /folder/to/save --gpuid 0 

• To train a backdoor model with the "blend" attack with a poison ratio of 10%

python train_backdoor_cifar.py --poison-type blend --poison-rate 0.10 --output-dir /folder/to/save --gpuid 0 

For GTSRB, tinyImageNet, ImageNet

• Follow the same training pipeline as Cifar10 and change the trigger size, poison-rate, and data transformations according to the dataset.

• For ImageNet, you can download pre-trained ResNet50 model weights from PyTorch first, then train this benign model with "clean and backdoor training data" for 20 epochs to insert the backdoor.

For Action Recognition

• Follow this link to create the backdoor model.

For Object Detection

• First, download the GitHub repository of SSD Object Detection pipeline.

• Follow Algorithm 1 and 2 in Clean-Image Backdoor for "Trigger Selection" and "Label Poisoning".

• Once you have the triggered data, train the model following SSD Object Detection.

For 3D Point Cloud Classifier

• Follow this link to create the backdoor model.

For Language Generation

• Follow this link to create the backdoor model.

Backdoor Analysis

• For smoothness analysis, run the following-

   cd Smoothness Analysis

   python hessian_analysis.py --resume "path-to-the-model"

• NOTE: "pyhessian" is an old package. Updated PyTorch can cause some issues while running this. You may see a lot of warnings.

FIP based Backdoor Purification

• Go to "src" folder

• For CIFAR10, To remove the backdoor with 1% clean validation data-

   python Remove_Backdoor_FIP.py --poison-type blend --val-frac 0.01 --checkpoint "path/to/backdoor/model" --gpuid 0 

• Please change the dataloader and data transformations according to the dataset.

For Adaptive Attack [Attacker has prior knowledge of FIP]

• We can do it in two ways

  • We can exactly follow the FIP implementation with high "--reg_F ($eta$_F in the paper)"

    python train_backdoor_with_spect_regul.py --reg_F 0.01 

  • We can deploy Sharpness-aware minimization (SAM) optimizer during training. Use a value greater than 2 for "--rho"

    python train_backdoor_with_sam.py  --rho 3

• With tighter smoothness constraints, it gets harder to find a favorable optimization point for both clean and backdoor data distribution

🚀 Purification Results

• FIP is able to purify backdoor by re-optimizing the model to a smoother minima.

Eigenvalue Plot During Purification

tSNE Plot

✏️ Citation

If you find our paper and code useful in your research, please consider giving a star ⭐ and a citation 📝.

@article{karim2024fisher,
  title={Fisher information guided purification against backdoor attacks},
  author={Karim, Nazmul and Arafat, Abdullah Al and Rakin, Adnan Siraj and Guo, Zhishan and Rahnavard, Nazanin},
  journal={arXiv preprint arXiv:2409.00863},
  year={2024}
}