Sequence-aware defenses against temporal gradient inversion

Develop lightweight, sequence-aware privacy defenses that protect ordered policy-gradient streams against amortized temporal reconstruction attacks while balancing privacy, computational cost, and training efficiency in resource-constrained embodied reinforcement-learning deployments.

Background

The paper evaluates quantization, pruning, additive noise, and DP-SGD against TRACE, an attack that reconstructs embodied observation-action trajectories from ordered per-step policy gradients. The results indicate that pruning and mild quantization provide little protection, whereas stronger perturbations such as DP-SGD substantially reduce reconstruction quality but may impose utility and computational costs.

The authors therefore identify a need for defenses that operate at the sequence level rather than treating gradient updates independently. Such defenses should disrupt the cross-step correlations exploited by TRACE without imposing prohibitive overhead on embodied agents that rely on real-time or resource-constrained collaborative training. The related-work discussion also states that the effectiveness of existing defenses against amortized temporal attacks on sequential trajectory data remains unresolved.

References

While these defenses reduce leakage in static settings, their effectiveness against amortized temporal attacks on sequential trajectory data remains open.

— Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning  (2609.30258 - Bhujel et al., 24 Sep 2026) in Section 3, Related Work; Appendix, Section Broader Impact

However, such mechanisms often incur computational overhead prohibitive for resource-constrained embodied deployments reliant on real-time training, so developing lightweight, sequence-aware defenses that balance privacy and efficiency remains an open challenge.

— Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning  (2609.30258 - Bhujel et al., 24 Sep 2026) in Appendix, Section Broader Impact