New TrojanTO attack targets trajectory optimization models in RL

Researchers have developed TrojanTO, a novel method for executing action-level backdoor attacks against trajectory optimization (TO) models used in offline reinforcement learning. Unlike previous reward-manipulation attacks, TrojanTO targets the sequence modeling nature of TO models and addresses challenges posed by high-dimensional action spaces. The attack enhances trigger-action connections through alternating training and uses precise poisoning via trajectory filtering for stealth, achieving effectiveness with a low poisoning budget. AI

IMPACT This research highlights potential security vulnerabilities in trajectory optimization models, necessitating the development of more robust defenses against sophisticated backdoor attacks.

RANK_REASON The cluster contains a research paper detailing a novel attack method against AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• offline reinforcement learning
• Trajectory Optimization models
• TrojanTO
• Yang Dai

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →