Privacy-Preserving Reinforcement Learning Using Homomorphic Encryption in Cloud Computing Infrastructures

Reinforcement learning (RL) is a learning technique that enables state-dependent learning through feedback from an environment and makes an action decision for maximizing a reward without prior knowledge of the environment. If these RL techniques are used for data-centric services running on cloud computing, serious data privacy issues may occur because it is required to exchange privacy-related user data for RL-based services between the users and the cloud computing platform. We consider using homomorphic encryption (HE) scheme, which enables cloud computing platforms to perform arithmetic operations without decrypting ciphertexts. Using the HE scheme, users are allowed to deliver only ciphertexts to the cloud computing platform for using RL-based services. We propose a privacy-preserving reinforcement learning (PPRL) framework for the cloud computing platform. The proposed framework exploits a cryptosystem based on learning with errors (LWE) for fully homomorphic encryption (FHE). Performance analysis and evaluation for the proposed PPRL framework are conducted in a variety of cloud computing-based intelligent service scenarios.