Reproduce “Privacy Preserving Reinforcement Learning for Population Processes” on OpenResearch

About this paper

Samuel Yang-Zhao, Kee Siong Ng

A privacy-preserving reinforcement learning framework is developed for controlling population processes, where an agent learns from aggregate, sensitive data. The work formally defines individual privacy for correlated data in such systems using Pufferfish Privacy, demonstrating that favorable privacy-utility trade-offs are achievable, especially with larger population sizes, through a novel meta-algorithm.

Key insights

• Individual privacy in population processes with correlated data can be formally defined as plausible deniability of an individual's participation in sampled datasets over time, rather than their specific status.
• Favorable privacy-utility trade-offs are theoretically and empirically achievable for differentially private RL in population processes, contrasting with negative results in other privacy-preserving RL settings.
• Increasing the sampled population size (N) significantly mitigates utility loss due to privacy noise, allowing for effective control even under stringent privacy guarantees.

Summary by alphaXiv.