Privacy-preserving image multi-classification deep learning model in robot system of industrial IoT

Deep learning in robot systems is a popular application that can learn and train the results per requirements, but that collects sensitive information in the training process, easily causing leakage of users' private information. To date, privacy-preserving deep learning models in robot systems have been sparsely researched. To solve the privacy leakage problem of deep learning in robot systems and fill the gap in robotics deep learning privacy research, in this paper a novel privacy-preserving image multi-classification deep-learning (PIDL) model in robot systems is presented. In PIDL, two schemes are proposed that adopt two groups of encrypted activation and cost functions-sigmoid plus cross-entropy function (PIDLSC) and softmax plus log-likelihood function (PIDLSL)-with secure calculation protocols, which are applied in a fog control center (FCC) with a non-colluding honest server by homomorphic encryption to improve the training efficiency, solve the encryption computation questions, and protect data and model privacy in robot systems. Security analysis and performance evaluation demonstrate that the proposed schemes realize security, correctness, and efficiency with low communication and computational costs.

Automated summary

A novel privacy-preserving image multi-classification deep-learning (PIDL) model is presented in this paper, with two schemes proposed that adopt secure calculation protocols applied in a fog control center (FCC) with a non-colluding honest server to protect data and model privacy in robot systems. The proposed schemes realize security, correctness, and efficiency with low communication and computational costs, as demonstrated in security analysis and performance evaluation.