Privacy-preserving model training architecture for intelligent edge computing

Abstract

With the rapid development of artificial intelligence and increasing data generated by end devices, the traditional cloud-centric data processing is gradually replaced by intelligent edge computing to achieve faster and nearer service via breaking the limit of network bandwidth and communication delay. However, training machine learning (ML) models on end devices is severely resource-constrained; besides, the privacy protection and continuous improvement of ML models are challenging. To address these problems, we propose an ML model training architecture to achieve intelligent edge computing in a novel cloud-edge-device cooperative manner, which is consisted of two phases: (1) the cooperative federated pre-training phase between the cloud and edge server is inspired by federated learning, coming with an incentive mechanism for fair reward allocation according to the contribution of edge servers for pre-training the model; (2) the privacy-preserving model segmentation training phase between the edge server and device leverages homomorphic encryption to realize model improvement and protection on end devices while transferring a large amount of computation to edge servers. Extensive simulations based on synthetic and real-world data demonstrate the effectiveness and feasibility of our proposed framework.