An Adaptive Model Aggregation Framework for Enhanced Swarm Learning

Abstract

Swarm Learning (SL) offers a transformative solution to the challenges posed by growing data security regulations and privacy concerns. It creates new opportunities for research in fields such as healthcare, finance, and smart technologies. This decentralized machine learning framework harnesses the collective intelligence of distributed nodes, each holding private data, and uses blockchain technology to ensure data privacy. The framework constructs a shared model by aggregating insights from each node without compromising the security of local data. Motivated by the goals of enhancing model performance and deepening the understanding of model aggregation, this study systematically tested various merging strategies on three datasets—MNIST, BloodMNIST, and Blood Cell Cancer (ALL)—within a simulated Swarm Learning environment. As a result, we developed the Adaptive Performance-Based Merge Strategy (AP-BMS), a novel method that dynamically selects the optimal merging algorithm within the Swarm network based on continuous model evaluations. This strategy improved performance by approximately 1% on the MNIST dataset, 6% on BloodMNIST and 4% on the Blood Cell Cancer (ALL) dataset. The AP-BMS marks a significant advancement in local model aggregation and further accelerates the evolution of Swarm Learning and its application in secure, decentralized machine learning environments.