Decentralized Gossip-Assisted Deep Learning Model Training for Resource-Constraint Edge Devices

There is significant interest in edge computing (EC) for computational social systems to process and store data at the edge of the network. One of the key applications of EC is to analyze the large-scale social data, received from multiple sources using machine learning/deep learning (ML/DL) models with minimum delay and higher accuracy. However, traditional models are often large and require significant computational resources, posing a challenge in resource-constrained edge networks. Besides that, traditional centralized ML/DL methods including collaborative learning have limitations such as data privacy and communication overhead. Federated learning (FL) is an alternative solution to overcome some of these limitations by allowing model training across multiple decentralized devices without sharing the actual data. However, the standard FL approaches face some challenges, including extended training times due to the heterogeneous devices and the risk of single-point failure. To address these challenges, in this article, we propose a novel Gossip-assisted DL model for resource-constraint edge devices problem by enabling decentralized and serverless training while mitigating the risk of single-point failure. Besides that, we develop a lightweight model extractor for local edge devices to train a DL model with the collaboration of neighboring devices that improves knowledge discovery with higher prediction accuracy. Extensive simulation results over two publicly available large-scale datasets demonstrate the effectiveness of the proposed approach over the state-of-the-art techniques.