Efficient Privacy-Preserving Federated Learning for Resource-Constrained Edge Devices

A large volume of data is generated by ubiquitous Internet-of-Things (IoT) devices and utilized to train machine learning models by IoT manufacturers to provide users with better services. Many deep learning systems for IoT data are required to perform all computation locally on small devices, which is not suitable for these resource-constrained devices. The devices can also send all the collected data to a server for costly model training by ignoring privacy concerns. To design an efficient and secure deep learning model training system, in this paper, we propose a federated learning system on the edge using the differential privacy mechanism to protect sensitive information and offload computation work from edge devices to edge servers, with consideration of communication reduction. In our system, a large-scale deep learning model is partitioned onto edge devices and edge servers, and trained in a distributed manner, in which all untrusted components are prevented from retrieving protected information from the training and inference process. We evaluate the proposed approach with respect to computation, communication, and privacy protection. The experiment results show that the proposed approach can preserve users’ privacy while significantly reducing computation and communication costs.