Diabetes types classification through federated learning and adversarial robustness

Abstract

In many parts of the world, diabetes is a widespread and serious concern affecting both young and older populations. It not only poses a threat to lives but also gives rise to a cascade of additional health complications. What makes it even more challenging is that there are different types of diabetes - type-1 and type-2 - each with its own set of health risks. This underscores the importance of accurate classification for tailored treatment. Despite the many machine learning models used for diabetes classification, the utilization of federated learning and its advancements remains an unexplored frontier in the realm of distinguishing between different types of diabetes. Hence, this research aims to develop a Federated Learning (FL) based diabetes type prediction model and establish it against adversarial attacks by producing adversarial examples using the Fast Gradient Sign Method (FGSM), departing from the traditional methods of Machine Learning. This approach not only keeps medical data private and is good for refining how we classify diabetes but also defends from intentionally manipulating inputted data that can cause incorrect decisions. In our research, we used a custom neural network sequential model to address this. After developing the model, we generated adversarial examples using the FGSM. This method, which is frequently applied to image processing, was modified for our numerical dataset to examine its influence on model stability. We have selected crucial features from our dataset according to the health experts to figure out if it’s type-1 or type-2. Our study’s accuracy rate of 93% is comparable to that of any other traditional centralized model.