Tactics of Adversarial Attack on Deep Reinforcement Learning Agents

Authors: Yen-Chen Lin, Zhang-Wei Hong, Yuan-Hong Liao, Meng-Li Shih, Ming-Yu Liu, Min Sun

IJCAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We apply the proposed tactics to the agents trained by the state-of-the-art deep reinforcement learning algorithm including DQN and A3C. In 5 Atari games, our strategically-timed attack reduces as much reward as the uniform attack (i.e., attacking at every time step) does by attacking the agent 4 times less often. Our enchanting attack lures the agent toward designated target states with a more than 70% success rate.
Researcher Affiliation Collaboration Yen-Chen Lin1, Zhang-Wei Hong1, Yuan-Hong Liao1, Meng-Li Shih1, Min Sun1 1National Tsing Hua University, Taiwan 2NVIDIA, Santa Clara, California, USA
Pseudocode No No explicit pseudocode or algorithm block is present, although mathematical formulations for optimization problems and functions are provided.
Open Source Code No The paper states 'Our implementation will be released.' but does not provide a concrete link or access at the time of publication.
Open Datasets Yes We evaluated our tactics of adversarial attack to deep RL agents on 5 different Atari 2600 games (i.e., Ms Pacman, Pong, Seaquest, Qbert, and Chopper Command) using Open AI Gym [Brockman et al., 2016].
Dataset Splits No The paper describes training and evaluation on Atari games, but does not explicitly provide numerical training/validation/test dataset splits (e.g., percentages or sample counts).
Hardware Specification No No specific hardware details (GPU, CPU models, memory, etc.) used for running experiments are provided in the paper.
Software Dependencies No The paper mentions software components like Open AI Gym, A3C, and DQN algorithms but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes The input to the neural network at time t was the concatenation of the last 4 images. Each of the images was resized to 84 84. The pixel value was rescaled to [0, 1]... We early stopped the optimizer when D(s, s + δ) < ϵ, where ϵ is a small value set to 0.007. The value of temperature T in Equation (4) is set to 1 in the experiments.