Homomorphic Advantage Operator: Stabilizing Reinforcement Learning Under Fully Homomorphic Encryption Constraints

Source facts

The preprint proposes the Homomorphic Advantage Operator to stabilize deep reinforcement learning when nonlinear operations are approximated under fully homomorphic encryption. It applies a zero-mean centering projection to temporal-difference targets to control recursive approximation drift without adding nonlinear multiplicative depth. The abstract reports zero boundary breaches across tested seeds and improved tabular policy accuracy in its experiments.

Analysis

The work connects encrypted computation with reinforcement-learning stability. If independently replicated beyond the reported tabular, CartPole, and logistics settings, this could support more practical privacy-preserving control and decision systems.

Uncertainty/Doubt

• This is an arXiv preprint. The cryptographic overhead, scalability, and reported stability results were not independently reproduced here.