FedCNNAvg: Federated Learning for Preserving-Privacy of Multi-clients Decentralized Medical Image Classification

Federated Learning (FL) permits the cooperative training of a joint model for several medical facilities while maintaining the decentralization of the data owing to privacy considerations. However, Federated optimizations often struggle with the heterogeneity of data dissemination among medical facilities. Nowadays, the domains of medical image classification, compression, and privacy are particularly difficult for diagnosing disease. The transmission of these medical images through the internet for diagnostic reasons must be protected against cyberattacks. In this proposed method, a Federated Learning approach with a Convolutional Neural Network (FedCNN) and Federated Averaging (FedAVG) is employed for classification problems. This technique adjusts the contribution of each data sample to the local goal during optimization based on knowledge of the client’s label distribution, thereby minimizing the instability caused by data heterogeneity. The model utilizes a hybrid approach to ensure consistency in time-series data. The datasets, namely, COVIDx-19 X-ray and malaria that are freely accessible are the subject of our in-depth investigations. The experimental results have been analyzed by evaluation metrics, namely, accuracy (78.79 and 98.92), precision (73.72 and 95.73), and recall (71.91 and 93.91) for proper validation. The findings demonstrate that FedCNN achieves better convergence performance than the main FL optimization methods under comparison.