Study of MobileNets Model in Federated Learning

Abstract

The challenges of handling decentralised data lead to the demand for research on secure gathering, efficient pro- cessing, and analysing of the data. In decentralised systems, each node (device) can make independent decisions, reducing the complexity and challenges of dealing with extensive data.Privacy has become a significant concern for our society due to the rise in the number of Edge/IoT devices, the lack of presence of a centralised system, etc. To solve this conundrum, federated learning was proposed. Federated learning works on the sharing of parameter values rather than the data. Worldwide, 10.2 Billion non-IoT and 19.8 billion IoT devices will be active in 2023. These devices lack security when it comes to using traditional machine learning. However, federated learning models solve this problem using techniques such as Secure Aggregation and Differential Privacy, which provide security for the devices and efficient communication between them. The challenges arise from heterogeneous devices, leading to the client selection problem, unbalanced data, and many more problems. The Proposed work focuses on using the MobileNets series of model architecture for federated learning using the FedAvg Strategy. MobileNets architecture has always been robust and reliable when it come to devices with resource constraints. An older generation system is used to show that federated learning is a viable technique for decentralized machine learning.