Code of Data-Limited TrojanNet Detector (DL-TND) and Data-Free TrojanNet Detector (DF-TND) from the paper: Practical Detection of Trojan Neural Networks: Data-Limited and Data-Free Cases
- Python: 3.7
- TensorFlow: 1.13.1
- Directly run the file main_dltnd (We provide the test Trojan model and the test clean model)
- If you want to train your own models, run the file train (You could change the path and the model name)
- The results will show whether the model is a Trojan model and what is the target label
- Python: 3.7
- PyTorch: 1.5.0
- Before running the code, please go to the robustness_lib and install the robustness package (pip install -e . or pip install -e robustness). The robustness package is originally from MadryLab. We made some modifications based on the original code.
- Run the file main_dftnd.
- The results will show the original images, recovered images, and the perturbations. The results will also provide the logits outputs after optimization, before optimization, and the logits output increase. The last row of the result tell you whether this is a Trojan model and what is the target label. (There are several parameters you can control. You could change gamma, which controls the sparsity of the perturbation. You could change the preset threshold T, which controls the confidence of the detection)
- If you want to train your own models, run the file train_model (You could change the path and the model name)
If you use this code, please cite the following reference
@inproceedings{wang2020practical,
title={Practical Detection of Trojan Neural Networks: Data-Limited and Data-Free Cases},
author={Wang, Ren and Zhang, Gaoyuan and Liu, Sijia and Chen, Pin-Yu and Xiong, Jinjun and Wang, Meng},
booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
pages={},
year={2020}
}