Federated learning for automated pain detection

Abstract

This thesis focuses on exploring the performance of automated pain detection systems through the application of federated learning. The study involves the use of two different databases (UNBC-McMaster Shoulder Pain Expression Archive database and the BioVid Heat Pain database) and in- vestigates various aspects of this approach. More specifically, it evaluates performance disparities between individual database training and federated learning methods, with the goal of determining the feasibility and benefits of federated learning. This research also explores the use of federated learn- ing within a single database, treating each patient as an individual ”client” to evaluate the potential benefits in terms of data privacy, while maintain- ing or even improving the accuracy of pain detection. Furthermore, this project evaluates the result of enhancing the privacy using differential pri- vacy. This study aims to provide valuable insights into the application of federated learning in the context of pain detection systems.