PDLHR: Privacy-Preserving Deep Learning Model With Homomorphic Re-Encryption in Robot System

Abstract

The robot system is a significant application that has attracted great attention, and deep learning is a powerful feature extraction technology that has achieved significant breakthroughs in many fields, especially in robot systems. However, the required massive dataset for a deep learning model in a robot system easily leads to privacy leakage. There have been few reports on privacy-preserving deep learning models, and none on multikeys, in robot systems. Existing privacy-preserving deep learning schemes in multiple keys have low efficiency and high interactions in non-robotic environments. To address these issues, this article proposes a privacy-preserving deep learning model with homomorphic re-encryption (PDLHR) and secure calculation tools in a robot system. The proposed re-encryption scheme is based on the Bresson-Catalano-Pointcheval (BCP) cryptosystem, which solves the multiple keys question, keeps the homomorphic nature, and is more simplified than the existing re-encryption scheme based on the BCP cryptosystem. The secure calculation tools are designed to realize efficient ciphertext computations. Compared to the previous work, PDLHR decreases the interactions in the decryption process, improves the ciphertext training efficiency, and preserves the privacy of input data, training model, and inference results. Security analysis and performance evaluations demonstrate that the proposed scheme realizes security, efficiency, and effectiveness with low communication and computation costs.