A federated learning privacy framework for missing data inference in environmental crowd sensing

Abstract

AbstractFederated learning (FL) is an attractive solution which holds promise to efficiently realize intelligent privacy‐preserving IoT systems. It does so by training local models at the IoT devices using locally collected data and aggregating these models at the server. First, user privacy is achieved as no data is shared with the server. Second, the traffic generated reduces significantly as only the model updates are exchanged between the server and the devices. The first part of this study examines the validity of using FL for provisioning user data privacy in the context of missing data inference in environmental MCS. Using a representative machine language (ML) model, namely, neural network (NN), it is found through simulations with the help of an existing dataset that FL performs as good as traditional ML, while maintaining user privacy. The second part of the study is the design of a privacy‐aware framework for FL sharing (sharing of model updates) so as to ensure privacy of model updates. The proposed framework is economical w.r.t. computational and communication overhead as it uses: (a) a single asymmetric key shared among the FL clients with the help of threshold cryptography, and (b) partial homomorphic encryption (PHE) for sharing the model updates. Additionally, security analysis of the framework supports its validity.