Problems of Privacy and Heterogeneity for Federated Learning Applications in Medical Image Analysis

Abstract

Recently, machine learning has become one of the most promising directions in working with medical data. Deep neural network models are the most eff ective and accurate, but they require large volumes of information for training. This is a common problem in the case of medical data, especially images, as their creation involves signifi cant costs. One solution to improve the quality of deep neural network models without increasing the training dataset is model aggregation. However, a problem arises with preserving the confi dentiality of medical images. For example, if one model is trained on an image containing information about a specifi c patient, other models participating in the aggregation may also gain access to this information. As a result, information about a specifi c patient may be disclosed. In an attempt to address the problem described above, this work aims to research and develop methods for aggregating machine learning models while preserving the privacy of medical images, particularly federated learning methods