PRIVACY-ENHANCED FEDERATED LEARNING FOR BRAIN TUMOR CLASSIFICATION: AN EVALUATION OF ACCURACY AND PRIVACY

Abstract

 Federated Learning (FL) is increasingly recognized as a valuable approach in medical imaging. By allowing multiple healthcare institutions to collaboratively train machine learning models on their local data, FL addresses the challenges of data sharing while preserving patient privacy. This decentralized method not only mitigates privacy concerns but also enhances the diversity and robustness of models by leveraging heterogeneous data from different sources, leading to improved performance in critical applications such as disease diagnosis and treatment planning. This work implements FL to classify brain tumor images, enabling model training across multiple sites without direct data exchange. To further enhance privacy, FL is integrated with Differential Privacy (DP), which introduces controlled noise to protect individual data points. The study evaluates three configurations: centralized deep learning (as a benchmark), federated learning, and federated learning integrated with the DP-SGD technique. The dataset used to train the model is brain tumor MRI images, consisting of four distinct classes. The results indicate that the centralized DL model achieved a 97% testing accuracy, comparable to the FL model, which also reached 97% accuracy at the 18th round—confirming the efficiency of collaborative training without data sharing. However, integrating differential privacy (FL+DP) reduced the testing accuracy to 64%, reflecting the expected trade-off between privacy preservation and model performance.