SecureMed: Federated learning empowered approach for securing medical data

In today’s era, the world is suffering from an infectious disease COVID-19 which proved to be health predicament. The COVID-19 testing data is being produced by the healthcare center at preeminent scale. The efficient handling of such data by World Health Organization (WHO) team and medical practitioners is the considerable challenge. The researchers have introduced many deep learning approaches to detect COVID-19, but security remained challenging. In this paper, we present federated learning based approach ‘SecureMed,’ which allows to securely process the data. In this approach, two parameters, ‘client selection’ and ‘aggregation method’ are considered. The proposed scheme inherits Markov model for designing the these two parameters. The approach is been validated by using publicly available COVID-19 dataset of size 771, 634 Genomes (source: NCBI, U.S. National Library of Medicine). Taking into account the probability computations, the proposed approach succeeds in client local training by 50%.

I. Introduction

Federated learning mechanism offers training sensitive data (more often, healthcare data) securely as it decentralizes local data model. Local models contain sensitive data and predictive score are shared accordingly with global model. As the large number devices/clients participate in this process, the global model is responsible for collecting the data from all the clients/local models. denotes the data distribution associated with each client c, then the number of samples from each client is , and therefore, we have total sample collection, . Considering medical data, the data collected by the local hospital needs to be preprocessed and then trained by the local model. In brief, the following are the steps followed by the local hospital: a) data preprocessing, b) model broadcasting, c) local model training, d) model aggregating, and e) parameter broadcasting. The patient’s data may contain sensitive information and they need to be preprocessed, and it also includes feature extraction/selection [1].