DLDP-FL: Dynamic local differential privacy federated learning method based on mesh network edge devices

Abstract

With increasing concerns about data security, privacy protection has become a necessary consideration in machine learning. Federated Learning (FL) solves the problem of data islands in machine learning and protects the privacy of participants by enabling further encryption through Differential Privacy (DP). In Artificial Intelligence of Things (AIoT), edge devices also suffer from data islands, which are addressed by applying FL to AIoT. However, the privacy and efficiency of existing FL needs to be further improved, and the integration with edge computing is not yet high enough. In this paper, first, we propose a novel edge FL architecture based on edge devices in mesh network architecture; next, we exploit the mesh networking features to address the problem of possible internal attacks from edge devices and design a Dynamic Local Differential Privacy (DLDP) algorithm; then, according to the communication characteristics of mesh network, we design Edge-FedAvg algorithm to reduce the communication cost; finally, to enhance the response to untrusted center servers, embed watermark in the model to further enhance the privacy protection capability. The Dynamic Local Differential Privacy Federated Learning (DLDP-FL) framework designed in this paper is used for FL of edge devices under mesh network, which can improve communication efficiency and enhance privacy protection capabilities at the same time.