Implementation of real-time learning on homomorphically encrypted visual inputs

Abstract

It’s challenging to provide security for cloud-based services, especially for cloud processing services, due to the fact that typical encryption techniques do not al low for calculation on encrypted data. The formation of Homomorphic Encryption techniques shows significant possibilities of incorporating encrypted computation on cloud infrastructures. This enables owners to outsource computation over confiden tial data to cloud vendors. Control and synthesis tasks of sensitive systems like traffic light control, article recommendation for online users and potentially, robot’s action determination can be delegated to a cloud-based Reinforcement Learning agent. In this study, we designed two Deep Reinforcement Learning agents that work on ciphertexts using Homomorphic Encryption. Both agents take encrypted state images and produce encrypted actions. One learns on plain data but evaluates on encrypted inputs, while the other one operates fully on encrypted space. The performance of both agents is compared against plaintext RL agents with identical parameters. The paper also describes possible architectures for such systems.