Ameliorating clustered federated learning using grey wolf optimization algorithm for healthcare IoT devicesAmeliorating clustered federated learning...R. K. Chaudhary et al.

The rapid growth of healthcare IoT devices has created a substantial demand for effective and secure data processing approaches that prioritize privacy while ensuring optimal performance in model training. Federated learning stands out as a significant solution for preserving data privacy while ensuring high-performance model training. Its application can confront problems like non-independent and identically distributed (non-IID) data, communication efficiency, and privacy issues. However, recent research has revealed that training models using non-IID data have a detrimental influence on performance, convergence, and overall model accuracy in federated learning. Traditional federated learning approaches, including clustered federated learning, are suffering from different problems like inefficient client training and fixed hyperparameter use. Thus, to overcome these issues, in this paper, a dynamic clustered federated learning approach for healthcare IoT devices is proposed using DBSCAN clustering and a grey wolf optimization algorithm. The proposed approach is named “GWO-CFL: Grey Wolf Optimized Clustered Federated Learning”. The GWO-CFL method utilizes DBSCAN clustering to cluster the healthcare IoT devices into distinct clusters based on learning rates to significantly improve the model aggregation efficiency and facilitate more effective model training across the healthcare IoT devices. After clustering the IoT devices based on learning rate similarities, grey wolf optimization is then employed to optimize or refine the learning rates at the cluster level using different phases to enhance the overall model convergence and accuracy across the grouped devices. This approach is primarily designed to handle the above-stated issues and the problem of static hyperparameter utilization in clustered federated learning. Experimental results on two healthcare datasets show that the proposed approach significantly improves the model performance in terms of accuracy, precision, recall, and F1-score and majorly enhances communication efficiency in clustered federated learning. Also, it is found that the proposed approach exhibits superior performance than other federated learning or clustered federated learning methods.