Research map / Adversarial robustness and security / Attacks on text and graphs
Spatiotemporally Constrained Action Space Attacks on Deep Reinforcement Learning Agents
Xian Yeow Lee, Sambit Ghadai, Kai Liang Tan, Chinmay Hegde, Soumik Sarkar
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add look-ahead planning that allocates attack budget across time, rather than using myopic per-step attacks | PPO · Lunar Lander | cumulative reward | Improved |
| Add look-ahead planning that allocates attack budget across time, rather than using myopic per-step attacks | PPO · Bipedal Walker | cumulative reward | Improved |
| Add look-ahead planning that allocates attack budget across time, rather than using myopic per-step attacks | DoubleDQN · Lunar Lander and Bipedal Walker | cumulative reward | Improved |
| Add look-ahead planning that allocates attack budget across time, rather than using myopic per-step attacks | PPO · Hopper, Half-Cheetah, and Walker | cumulative reward | Improved |
| Increase LAS budget (ablation) | PPO · Lunar Lander | cumulative reward | Hurt |
| Increase LAS budget beyond the point where attacks saturate (ablation) | LAS · Lunar Lander | cumulative reward reduction versus MAS | No clear effect |
| Use a longer horizon with the same overall budget (ablation) | LAS · Lunar Lander | cumulative reward | Mixed |
| Use ℓ2 rather than ℓ1 temporal projection to distribute attacks more frequently (ablation) | LAS · Lunar Lander | cumulative reward | Improved |
| Use ℓ1 rather than ℓ2 temporal projection to concentrate attacks at fewer time points (ablation) | LAS · Lunar Lander | cumulative reward | Hurt |
| Increase budget with ℓ1 spatial projection at horizon 5 beyond B=3 (ablation) | LAS · Lunar Lander | cumulative reward | No clear effect |
| Use action-dimension-aware LAS perturbations, concentrating more attack on the vulnerable Up-Down action (ablation) | PPO · Lunar Lander | cumulative reward | Improved |
Closest papers by meaning
- Optimal Attack and Defense for Reinforcement Learning (AAAI 2024 · similarity 0.85)
- Adversarial Inception Backdoor Attacks against Reinforcement Learning (ICML 2025 · similarity 0.84)
- There are Levels to It: Red Teaming LLMs with Hierarchical Reinforcement Learning (NeurIPS 2026 · similarity 0.84)
- Rethinking Adversarial Policies: A Generalized Attack Formulation and Provable Defense in RL (ICLR 2024 · similarity 0.84)
- Exploring the Efficacy of Multi-Agent Reinforcement Learning for Autonomous Cyber Defence: A CAGE Challenge 4 Perspective (AAAI 2025 · similarity 0.84)
- PolicyGuard: Towards Test-time and Step-level Adversary Defense for Reinforcement Learning Agent (ICML 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
