aiqbench analysis
SEBA introduces a novel framework for launching sample-efficient black-box adversarial attacks specifically against visual reinforcement learning agents, addressing critical security gaps in robotics and autonomous control systems. By integrating shadow Q models and world models to minimize environment queries, this research highlights the inherent vulnerability of high-performance visual RL systems to imperceptible perturbations.
SEBA: Sample-Efficient Black-Box Attacks on Visual Reinforcement Learning
arXiv:2511.09681v2 Announce Type: replace Abstract: Visual reinforcement learning has achieved remarkable progress in visual control and robotics, but its vulnerability to adversarial perturbations remains underexplored. Most existing black-box attacks focus on vector-based or discrete-action RL, and their effectiveness on image-based continuous control is limited by the large action space and excessive environment queries. We propose SEBA, a sample-efficient framework for black-box adversarial attacks on visual RL agents. SEBA integrates a shadow Q model that estimates cumulative rewards under adversarial conditions, a generative adversarial network that produces visually imperceptible perturbations, and a world model that simulates environment dynamics to reduce real-world queries. Through a two-stage iterative training procedure that alternates between learning the shadow model and refining the generator, SEBA achieves strong attack performance while maintaining efficiency. Experiments on MuJoCo and Atari benchmarks show that SEBA significantly reduces cumulative rewards, preserves visual fidelity, and greatly decreases environment interactions compared to prior black-box and white-box methods.
Abstract:Visual reinforcement learning has achieved remarkable progress in visual control and robotics, but its vulnerability to adversarial perturbations remains underexplored. Most existing black-box attacks focus on vector-based or discrete-action RL, and their effectiveness on image-based continuous control is limited by the large action space and excessive environment queries. We propose SEBA, a sample-efficient framework for black-box adversarial attacks on visual RL agents. SEBA integrates a shadow Q model that estimates cumulative rewards under adversarial conditions, a generative adversarial network that produces visually imperceptible perturbations, and a world model that simulates environment dynamics to reduce real-world queries. Through a two-stage iterative training procedure that alternates between learning the shadow model and refining the generator, SEBA achieves strong attack performance while maintaining efficiency. Experiments on MuJoCo and Atari benchmarks show that SEBA significantly reduces cumulative rewards, preserves visual fidelity, and greatly decreases environment interactions compared to prior black-box and white-box methods.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2511.09681 [cs.LG]
(or arXiv:2511.09681v2 [cs.LG] for this version)
Huang T, Jin Y, Liu J, et al. SEBA: Sample-Efficient Black-Box Attacks on Visual Reinforcement Learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026: 27861-27871
Full text under the publisher's open licence · 191 words
Related coverage
Returning to ARC
AI Alignment Forum · Aug 4, 2026
Adversarial Agents: Black-Box Evasion Attacks with Reinforcement Learning
arXiv · Jun 6, 2026
Robustness of Vision Language Models Against Split-Image Harmful Input Attacks
arXiv · Sep 1, 2026
REPLICANT: Learning Policies for Evading and Hardening Malware Detectors
arXiv · Aug 31, 2026