Natural Black-Box Adversarial Examples against Deep Reinforcement Learning

Authors: Mengran Yu, Shiliang Sun8936-8944

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on multiple environments demonstrate the effectiveness of adv RL-GAN in terms of reward reductions and magnitudes of perturbations, and validate the sparse and targeted property of adversarial perturbations through visualization.
Researcher Affiliation Academia Mengran Yu and Shiliang Sun* School of Computer Science and Technology, East China Normal University, 3663 North Zhongshan Road, Shanghai 200062, P.R. China mengranyu97@gmail.com, slsun@cs.ecnu.edu.cn
Pseudocode Yes Algorithm 1: Adv RL-GAN Attack
Open Source Code No The paper does not provide any concrete access information for source code, such as a repository link or an explicit statement of code release.
Open Datasets Yes We evaluate the performance of the proposed adv RL-GAN attack against DRL agents on 6 different Atari 2600 games, including Breakout, Pong, Chopper Command, Space Invaders, Ms Pacman, and Qbert, using Open AI Gym (Brockman et al. 2016).
Dataset Splits No The paper mentions training target policies but does not provide specific details on training, validation, and test dataset splits, percentages, or sample counts.
Hardware Specification Yes Experiments are performed with the Py Torch library on Ge Force RTX 2080Ti GPUs.
Software Dependencies No The paper mentions 'Py Torch library' but does not provide a specific version number, nor does it list versions for other software dependencies.
Experiment Setup Yes We train the adv RL-GAN model with 8 × 10^5 steps by adopting the REINFORCE algorithm. The hyper-parameters of the loss function LG are set to λ1 = 10, λ2 = 0.1, and λ3 = 10.