FedRSMax: An Effective Aggregation Technique for Federated Learning with Medical Images

The traditional deep learning framework faces two critical challenges: limited data available for successful model training and concerns regarding user data privacy. Federated learning, which operates in a decentralised paradigm, offers a promising solution to these challenges. Federated averaging (FedAvg) is a common aggregation procedure in federated contexts. FedAvg, however, experiences convergence issues, especially when there is significant diversity in the data distributions among clients. To address this problem, we explore two effective aggregation techniques, namely random-sampling federated maximum (FedRSMax) and random-sampling federated median (FedRSMed) with adaptive moment estimation (Adam), and compare their performance characteristics with FedAvg. In this study, we use a well-established convolutional neural network (LeNet) as a global model for federated learning, and the HAM10000 dermatoscopic image dataset is used as the primary data source. We balance the dataset and generate random subsets to induce data heterogeneity for different simulated clients and evaluate the performance of the proposed techniques. Our findings demonstrate that FedRSMax outperforms FedRSMed and FedAvg in terms of accuracy, recall, and precision and can therefore serve as an effective alternative for aggregation in federated learning.