Privacy-Preserving Deep Reinforcement Learning Based On Adaptive Noise For Sepsis Treatment

Abstract

In recent years, deep reinforcement learning has made amazing achievements in the field of personalized treatment plan recommendation, especially in sepsis treatment. But deep reinforcement learning has privacy leakage problem. Differential privacy is a reliable privacy-preserving technique that has been proven to be effective in protecting the privacy of machine learning from adversary attacks on sensitive data. However, existing methods add the same amount of noise in the model gradient, which dramatically reduces the usability of the model. In this paper, we propose a privacy-preserving deep reinforcement learning based on relevance analysis, that is used add noise adaptively. This approach aims to better balance the usability and privacy preservation of reinforcement learning models for sepsis treatment. Our experimental results shows that this mechanism can achieve higher model usability with the same privacy-preserving strength.

Ähnliche Arbeiten