Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning
TRACE is introduced, an amortized temporal gradient-inversion attack that autoregressively reconstructs the sequence of private observation-action trajectories from per-step policy-learning gradients, dominating the learning-based baseline across all reconstruction metrics and exceeding optimization attacks while runni...
Sudip Bhujel, Shang-Hao Shi, Rui-Quan Huang et al.
· 0 citations