This is the official implementation for Understanding Adversarial Attacks on Observations in Deep Reinforcement Learning (Accepted in SCIENCE CHINA Information Sciences (SCIS), 2023).
The code of DQN is based on radial_rl and is in the folder ./Atari/DQN.
The code of A2C is also based on radial_rl and is in the folder ./Atari/A2C.
The code of PPO is based on pytorch-a2c-ppo-acktr-gail and is in the folder ./Atari/PPO.
The code of PPO is based on ATLA and is in the folder ./Mujoco/PPO
If you find Two-Stage-Attack helpful, please cite our paper.
@article{
qiaoben2021understanding,
title={Understanding Adversarial Attacks on Observations in Deep Reinforcement Learning},
author={Qiaoben, You and Ying, Chengyang and Zhou, Xinning and Su, Hang and Zhu, Jun and Zhang, Bo},
journal={arXiv preprint arXiv:2106.15860},
year={2021}
}