Event
Proposes TRACO closed-loop optimization to enhance VLA visual attacks
1 reports · 1 sources
Chunjin Ye et al. propose the TRACO framework, which enhances targeted visual task redirection attacks against vision-language-action (VLA) models through closed-loop optimization. The framework contains two mechanisms: Execution-Driven Supervision reconstructs supervision signals based on the states actually reached under attack to guide subsequent updates; Progress-Guided Selection concentrates updates where task progress stalls while maintaining supervision over the early behaviors required to reach that state. Experiments across multiple VLA models and simulation environments show that TRACO improves the task redirection success rate, highlighting the security implications of closed-loop attack optimization.
Report timeline
- TRACO: Enhancing Visual Task-Redirection Attacks on Vision-Language-Action Models via Closed-Loop Optimization
The TRACO framework enhances targeted visual attacks on vision-language-action (VLA) models through closed-loop optimization. Its Execution-Driven Supervision rebuilds supervision signals based on the states actually reached under attack, while Progress-Guided Selection concentrates updates where task progress stalls and maintains supervision of prior behavior. Experiments across multiple VLA models and simulated environments show that TRACO improves task redirection success rates, highlighting the security implications of closed-loop attack optimization.
arXiv cs.RO ·