Official implementation of “Adversarial Attacks on Event-Based Pedestrian Detectors: A Physical Approach” (AAAI 2025).
- Google Drive: https://drive.google.com/drive/folders/1X1fINjuOPxh_sM_1Sv8Tu3xVGbNF_WhO?usp=drive_link
Please place the following files under checkpoints/:
best.pth-- checkpoint of our modelgen_rvt-b.ckpt-- checkpoint of the target detector RVT
Example directory structure:
event_attack_physics/
└── checkpoints/
├── best.pth
└── gen_rvt-b.ckpt
-- Google Drive: https://drive.google.com/drive/folders/1wYwMbDzgdrcFD1I3W0i9_qzQo0w0dJnJ?usp=sharing
git clone https://github.com/Joseph-LIN-Tech/event_attack_physics.git
cd event_attack_physicspython test.py-
The test process uses the default configuration files located under Options/test.yaml
-
Please update the data paths 'data_simu_pose' and 'labeled_simu_pose' in Options/test.yaml to the corresponding local directories on your machine.
-
The confidence threshold of target detecor RVT is defined in the Options/rvt_config.yaml (line 107 "confidence_threshold")
-
You can set the metrics as 'seq_attack_success_rate' or 'AP' seperately to get the evaluation value in the paper.
python train.py-
The training process uses the default configuration files located under Options/train.yaml
-
Please update the data paths 'data_simu_pose' and 'labeled_simu_pose' in Options/train.yaml to the corresponding local directories on your machine.
Some of our code is based on the implementation of LETGAN, and RVT.
If you find this repository useful, please cite our paper:
@inproceedings{lin2025adversarial,
title={Adversarial Attacks on Event-Based Pedestrian Detectors: A Physical Approach},
author={Lin, Guixu and Niu, Muyao and Zhu, Qingtian and Yin, Zhengwei and Li, Zhuoxiao and He, Shengfeng and Zheng, Yinqiang},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
year={2025}
}
Please check the license files in this repository.