Publications

Preprints

Practical Adversarial Attacks on Stochastic Bandits via Fake Data Injection adversarial learning bandits online learning Qirun Zeng, Eric He, Richard Hoffmann, Xuchuang Wang, Jinhang Zuo arXiv preprint, 2025 [paper] [website] [code] Existing bandit attack models rely on unrealistic assumptions like unrestricted reward manipulation. We propose *Fake Data Injection*, a practical threat model where attackers inject bounded fake feedback to mislead UCB and Thompson Sampling with sublinear effort, exposing real-world vulnerabilities in stochastic bandit algorithms.

Object-Pose Estimation With Neural Population Codes computer vision population code Heiko Hoffmann, Richard Hoffmann arXiv preprint, 2025 [paper] [website] [code] Robotic assembly tasks require precise object-pose estimation, but object symmetry makes direct rotation prediction ambiguous. Could using a *neural population code* for object rotation enable faster and more accurate pose estimation, achieving a higher accuracy on the T-LESS dataset in less time compared to direct pose mapping?