Differentially Private Collaborative Learning for the IoT Edge

Linshan Jiang, Xin Lou, Rui Tan, Jun Zhao

Abstract: Collaborative learning based on training data contributed by many edge devices is a promising paradigm for implementing crowd intelligence. The collaboratively trained model generally provides superior classification performance due to the increased volume and expanded coverage of the training data. However, the data contribution may incur the concern of privacy breach. This paper presents the design of a privacy-preserving collaborative learning approach, in which the edge devices and the cloud train different stages of a deep neural network, and the data transmitted from an edge device to the honest-but-curious cloud is perturbed by Laplacian random noises to achieve ε-differential privacy. We apply the proposed approach to a case study of collaboratively training a convolutional neural network for handwritten digit recognition. The results show that our approach maintains 99% and 96% classification accuracy in implementing privacy loss levels of ε = 5 and ε = 2, respectively.