notesum.ai

Published at November 20

Provably Efficient Action-Manipulation Attack Against Continuous Reinforcement Learning

cs.AI

Released Date: November 20, 2024

Authors: Zhi Luo¹, Xiyuan Yang¹, Pan Zhou¹, Di Wang²

Aff.: ¹School of Cyber Science and Engineering, Huazhong University of Science and Technology; ²King Abdullah University of Science and Technology

Arxiv: http://arxiv.org/abs/2411.13116v1

Refer to caption

Environment 1	Training steps	DDPG		PPO
	Training steps	Oracle attack	LCBT attack	Oracle attack	LCBT attack
	3e5	99.816%	99.065%	98.581%	98.642%
	4e5	99.910%	99.146%	98.640%	99.501%