Cooperative Swarm Learning for Distributed Cyclic Edge Intelligent Computing

Abstract

Emerging technologies and applications including the Internet of Things (IoT), social networking, and crowd-sourcing generate large amounts of data at the network edge. Machine learning models are often built from the collected data, to enable the detection, classification, and prediction of future events. Due to bandwidth, storage, and privacy concerns, it is often impractical to send all the data to a centralized location. Recently, the emerging Federated Learning (FL) systems mitigates some of those concerns, by keeping data private to cater the confidentiality issues. Although FL has drawn the attention of researchers and developers, the involvement of an aggregator server for global model updates, makes the system prone to privacy and security breaches. To resolve this issue Swarm Learning (SL) securely onboards edge clients and dynamically elects the swarm leader, to perform collaborative model training in an extremely decentralized, private and secured manner. However, the conventional blockchain-enabled dynamic selection of leader at every communication round is resource intensive and degrades SL’s effectiveness. In this paper, we propose a lightweight communication and computation efficient cooperative swarm learning framework for cyclic edge intelligent computing. The proposed cooperative swarm learning provides cyclic learning based on Deep Neural Networks procedure for training a global model in a sequential ring topology in one iteration which reduces the computational and communication overhead. The global model update method trains a shared global thermal comfort prediction model. Since there is model updating being performed in each participant of the cyclic edge intelligent computing network, bandwidth and computing power demand are reduced. An empirical investigation on Non-IID thermal comfort dataset demonstrates the superiority of the proposed framework.