High-Performance Federated Learning Techniques for Deep Learning on Medical Dataset

In recent years, a gigantic amount of data has been generated by varied sources such as smartphones, sensors, IoT devices, and many smart applications. This rapid generation of data leads to processing delays, communication overheads, security, and privacy issues where all the sensitive data is stored at a central location. To provide accurate decisions using a machine learning technique, the data needs to be collected at a central location, which increases the risk of data breaches or privacy violations. Federated learning has emerged as a promising solution to this challenge, enabling collaborative model training across distributed data sources without consolidating raw data. This led to an increasing demand for decentralized machine learning techniques that can leverage this distributed data while respecting privacy constraints. In this paper, we proposed the evaluation of the model training for the participating nodes of high-performance computing using cross-silo federated learning. Nodes are communicating with the master node using cross-silo federated learning. The local model training was performed at four parallel compute nodes, and the global model was placed at the master node of the computational facility to aggregate the updates from the local training nodes. The study is experimented with deep learning models on the healthcare dataset, and for federated optimization, the FedAvg and FedSGD were applied. With the proposed high-performance federated learning, the model training has been optimized and outperformed with the FedAvg strategy.