Using Federated Learning to construct a Robust Predictive Model in case of Limited Availability of Data

Abstract

        The traditional methods of building machine learning models involve centralizing sensitive user data. Federated learning, is a de-centralized method, born from an intersection of artificial intelligence and edge computing on-device. This form of learning allows for smarter models, lower latency and lower power consumption, all while guaranteeing privacy. Healthcare is one such domain where the data must be strongly secure and privately owned. Federated healthcare learning can be proven highly successful as it helps AI models to learn on private data without violating privacy. In this paper we show the accuracy and learning of individual models (Alice, Bob, Chips) that have less data and their relation to federated model. The chosen dataset is the popular Wisconsin breast cancer dataset which Dr. Willaim. H.Wolberg creates using fluid samples from patients with solid breast masses. An experimentation is carried out using this dataset, from which the 80% training data is split equally among the three workers giving 33% of each. An accuracy of 93.42%, 90.78%, 90.78% was achieved in testing with 20 percent test data by Alice, Bob, Chips models respectively. The federated model developed using average weights of individual models proved the principle with an accuracy of 98.68%.

Keywords: Federated Learning, Privacy, Data Security