Saved in:
Bibliographic Details
Main Authors: Escamilla, Jose Efraim Aguilar, Hong, Haoyang, Li, Jiawei, Zhao, Haoyu, Zhang, Xuezhou, Hong, Sanghyun, Wang, Huazheng
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.10062
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915937633959936
author Escamilla, Jose Efraim Aguilar
Hong, Haoyang
Li, Jiawei
Zhao, Haoyu
Zhang, Xuezhou
Hong, Sanghyun
Wang, Huazheng
author_facet Escamilla, Jose Efraim Aguilar
Hong, Haoyang
Li, Jiawei
Zhao, Haoyu
Zhang, Xuezhou
Hong, Sanghyun
Wang, Huazheng
contents We study reward poisoning attacks in reinforcement learning (RL), where an adversary manipulates rewards within constrained budgets to force the target RL agent to adopt a policy that aligns with the attacker's objectives. Prior works on reward poisoning mainly focused on sufficient conditions to design a successful attacker, while only a few studies discussed the infeasibility of targeted attacks. This paper provides the first precise necessity and sufficiency characterization of the attackability of a linear MDP under reward poisoning attacks. Our characterization draws a bright line between the vulnerable RL instances, and the intrinsically robust ones which cannot be attacked without large costs even running vanilla non-robust RL algorithms. Our theory extends beyond linear MDPs -- by approximating deep RL environments as linear MDPs, we show that our theoretical framework effectively distinguishes the attackability and efficiently attacks the vulnerable ones, demonstrating both the theoretical and practical significance of our characterization.
format Preprint
id arxiv_https___arxiv_org_abs_2604_10062
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle When Can You Poison Rewards? A Tight Characterization of Reward Poisoning in Linear MDPs
Escamilla, Jose Efraim Aguilar
Hong, Haoyang
Li, Jiawei
Zhao, Haoyu
Zhang, Xuezhou
Hong, Sanghyun
Wang, Huazheng
Machine Learning
We study reward poisoning attacks in reinforcement learning (RL), where an adversary manipulates rewards within constrained budgets to force the target RL agent to adopt a policy that aligns with the attacker's objectives. Prior works on reward poisoning mainly focused on sufficient conditions to design a successful attacker, while only a few studies discussed the infeasibility of targeted attacks. This paper provides the first precise necessity and sufficiency characterization of the attackability of a linear MDP under reward poisoning attacks. Our characterization draws a bright line between the vulnerable RL instances, and the intrinsically robust ones which cannot be attacked without large costs even running vanilla non-robust RL algorithms. Our theory extends beyond linear MDPs -- by approximating deep RL environments as linear MDPs, we show that our theoretical framework effectively distinguishes the attackability and efficiently attacks the vulnerable ones, demonstrating both the theoretical and practical significance of our characterization.
title When Can You Poison Rewards? A Tight Characterization of Reward Poisoning in Linear MDPs
topic Machine Learning
url https://arxiv.org/abs/2604.10062