Better Trigger Inversion Optimization in Backdoor Scanning

This is the implementation for CVPR 2022 Oral paper "Better Trigger Inversion Optimization in Backdoor Scanning".

The code is based on and inspired by Neural Cleanse.

Prerequisite

The code is implemented and tested on Keras with TensorFlow backend and PyTorch. It runs on Python 3.6.9.

Keras Version

• Keras 2.3.0
• Tensorflow 1.14.0
• Tensorpack 0.9.8 (if load data using data generator)

PyTorch Version

• PyTorch 1.7.0

Usage

The main functions are located in src/main.py file. For a test drive, please use the following command:

python3 src/main.py --phase evaluate

This will generate a backdoor trigger for the label pair from loggerhead turtle (source) to kangaroo (target). The result might be slightly different as in the paper due to randomness and GPU architectures.

The src/main.py script provides different options:

• --gpu: GPU id for running the experiment, default is 0
• --phase: evaluate for trigger inversion, test for testing model performance
• --model: model architecture, default is resnet50
• --pair: label pair for generating the backdoor trigger, default is 33-104 (souce label: loggerhead turtle, target label: kangaroo)
• --seed: random seed
• --batch_size: batch size in trigger inversion and model testing
• --num_classes: number of classes of the subject model
• --attack_size: number of samples for inversing the backdoor trigger

Reference

@inproceedings{taog2022better,
  title={Better Trigger Inversion Optimization in Backdoor Scanning},
  author={Tao, Guanhong and Shen, Guangyu and Liu, Yingqi and An, Shengwei and Xu, Qiuling and Ma, Shiqing and Li, Pan and Zhang, Xiangyu},
  booktitle={2022 Conference on Computer Vision and Pattern Recognition (CVPR 2022)},
  year={2022}
}