Privacy-preserving and communication-efficient federated learning in Internet of Things

Abstract

Aimed at the privacy leakage caused by collecting data from numerous Internet of Things (IoT) devices for centralized training, a novel distributed learning framework, namely federated learning, came into being, where devices train models collaboratively while leaving their private datasets locally. Although many schemes have been proposed about federated learning, they are still short in communications and privacy due to the limited network bandwidth and advanced privacy attacks. To address these challenges, we develop PCFL, a privacy-preserving and communication-efficient scheme for federated learning in IoT. PCFL is composed of three key components: (1) gradient spatial sparsification where irrelevant local updates that deviate from the collaborative convergence tendency are prevented from being uploaded; (2) bidirectional compression where computation-less compression operators are used to quantize the gradients both at the device-side and server-side; and (3) privacy-preserving protocol which integrates secret sharing with lightweight homomorphic encryption to protect the data privacy and resist against various collusion scenarios. We analyze the correctness and privacy of our scheme, and carry out theoretical and experimental comparison on two real-world datasets. Results show that PCFL outperforms the state-of-the-art methods by more than 2× in terms of communication efficiency, along with high model accuracy and marginal decreases in convergence rate.