Analyzing the security differential privacy provides and the trade-off between performance and privacy in medical image classification

Abstract

One of machine learning’s main purposes is to draw out functional and practical information from a set of data while perpetuating the entire privacy by protecting all information. While it might seem a bit hard to maintain, privacy does play a vital role in every sector, and thus, the information must be frequently balanced, especially when extracting sensitive datasets. For instance, medical research or image classification can be considered an important application where patient privacy, as well as the extraction of information, are both of utmost importance [12]. Medical images are details that consist of a patient’s private information and are collected from various hospitals, nursing homes, and research institutes. Later on, these images are utilized to infer a patient’s physical condition, ultimately leading to an invasion of privacy[10]. In recent years, medical images have become a prominent research and analysis subject, and therefore more and more people are getting affected as their private information is being shared. Thus, in our research, we are going to showcase different ways to defend against information leakage. Differential privacy is considered one of the strongest forms of privacy because we work with privacy-preserving algorithms and learning-based mechanisms. Apart from that, federated learning and image watermarking can also help in preserving privacy. Deep learning techniques that can be utilized to preserve data utilizing Conditional GANs also face particular difficulties when used with medical images. In order to show the optimal method of data preservation, we will attempt to collect a dataset.