Performance Analysis of Aggregation Algorithms in Privacy Preserving Federated Learning

Countless individuals around the world are impacted by breast cancer, making it a substantial global health challenge. Detecting this disease at an early stage is essential for effective treatment, and recent technological progress has provided fresh opportunities to enhance the precision of breast cancer diagnosis. Decentralized machine learning commonly known as federated learning is one of the privacy preservation technique, where machine learning models are trained locally and tested globally [1]. In our research we employed it for breast cancer detection while taking into account the privacy considerations of healthcare institutions. Decentralized machine learning (DML) commonly known as Federated Learning (FL) surpasses centralized machine learning (ML) in preserving privacy. In centralized ML, raw patient data is gathered and stored in a central location, posing privacy risks. Decentralised Machine Learning (DML), however, ensures privacy by sharing model insights and updates derived from local datasets, avoiding the transfer of identifiable patient information. Aggregation of the parameters collected by individual nodes plays crucial role in the performance of the global machine learning model. In this research we analysed the performance of commonly adopted aggregation algorithms FedAvg and FedProx under different parametric conditions. Comprehensive analysis of the performance metrics, includes accuracy and sensitivity. This comparison helps determine which approach offers superior diagnostic capabilities.