Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning
Policy gradients can leak whole embodied-agent trajectories, not just isolated observations.
The paper introduces TRACE, a temporal gradient-inversion attack that reconstructs private observation-action sequences from per-step policy-learning gradients. It uses correlations across consecutive gradients and recovers actions from policy-head gradient structure under small entropy regularization. On held-out embodied scenes, it reports 18.8 dB PSNR, near-perfect action recovery, and 3-4.5 ms reconstruction per frame. The authors say the results hold across several victim architectures, input types, and larger discrete action spaces, and argue defenses need to treat gradient streams as sequences. ArXiv · AI/CL/LG's note
The paper introduces TRACE, a temporal gradient-inversion attack that reconstructs private observation-action sequences from per-step policy-learning gradients. It uses correlations across consecutive gradients and recovers actions from policy-head gradient structure under small entropy regularization. On held-out embodied scenes, it reports 18.8 dB PSNR, near-perfect action recovery, and 3-4.5 ms reconstruction per frame. The authors say the results hold across several victim architectures, input types, and larger discrete action spaces, and argue defenses need to treat gradient streams as sequences. ArXiv · AI/CL/LG's note
score 4