TRUST AND HETEROGENEITY AWARE DECENTRALIZED FEDERATED-LEARNING FRAMEWORK FOR PRIVACY PRESERVED HEALTHCARE CLASSIFICATION

Abstract

Healthcare centers are becoming essential hubs for processing vast amounts of data used in disease diagnosis. Despite this, practical implementations face issues, like data privacy concerns, insecure storage, and limited sharing efficiency. To tackle such complexity, Trust Heterogeneity aware-based Fractional Football Optimization Algorithm enabled Deep High-order Attention Neural Network (TrustHet aware-based FFboA_DHA-Net) for classifying privacy preserved healthcare system, is proposed. Initially, Local training is conducted on every node using local data, followed by server-side model aggregation, where nodes download global model, update it with local models. In training model, input image is pre-processed, then lesions are segmented, and feature are extracted. Using Deep High-order Attention Neural Network (DHA-Net) health care classification is done, DHA-Net is trained by FFbOA. FFbOA is an integration of Football Optimization Algorithm (FbOA) with Fractional Calculus (FC). Here, Heterogeneity-aware FL allows devices with diverse data to contribute effectively to a shared model. A decentralized aggregation strategy is used, trust establishment mechanism between server and nodes is designed by considering trust factors. Aggregation is enhanced through harmonic analysis, and both local updates and server-side aggregation are extracted using an averaging method. Additionally, optimal results attained are 96.981% of F1-score, 0.020 of Loss function, 97.817% of Mean Average Precision (MAP), 0.111 of Normalized Mean Square Error (MSE), and 0.333 of Normalized Root Mean Squared Error (RMSE).