NDSS 2025 – TrajDeleter: Enabling Trajectory Forgetting In Offline Reinforcement Learning Agents

A blog post summarizing an NDSS 2025 session presents TrajDeleter, a method to unlearn specific trajectories in offline reinforcement learning by degrading performance on targeted states while maintaining performance elsewhere, and TrajAuditor for verifying successful unlearning. Experiments across six offline RL algorithms and three tasks show effective unlearning (~94.8%) at about 1.5% of full retraining time, with replication materials available.