An Empirical Study of Federated Brain Tumor Classification Under IID and Non-IID Client Distributions with DP-Inspired Local Perturbation

Abstract

Federated learning (FL) is an approach to protect privacy during medical image analysis due to the possibility to jointly train the model without sharing data between institutions. In this study, we provide a comparative analysis of federated brain tumor classification performance in case of IID and non-IID clients with and without differential privacy (DP)-based local perturbation. The performance metrics are calculated according to the accuracy, F1-score, and AUC measures for the multi-client FL convolutional neural network (CNN) based binary classifier. We demonstrate that the best results were obtained for the IID client configuration without DP application, with 97.30% test accuracy, 0.9754 F1-score, and 0.9947 AUC values, while the best configuration among the non-IID ones resulted in 93.30% test accuracy, 0.9401 F1-score, and 0.9761 AUC values. The application of DP caused significant degradation of classification performance with IID and non-IID configurations leading to 84.25% and 78.79% test accuracies, respectively. This demonstrates the presence of a privacy-utility tradeoff. Overall, the results suggest that federated brain tumor classification is achievable and works effectively; however, the effectiveness of federated brain tumor classification depends greatly on the client data heterogeneity and DP application.