FTL-Emo: Federated Transfer Learning for Privacy Preserved Biomarker-Based Automatic Emotion Recognition

Advancements in IOT has revolutionized remote patient monitoring, however, privacy is still the major challenge faced by researchers. We put forward a Federated learning-based technique to handle the issue of privacy, and to overcome the issue of requirement of large dataset we have employed a transfer learning approach. Federated transfer learning (FTL) model analyze electronic health records of the user to detect their emotional state. Emotion analysis has been observed by monitoring physiological changes of human body, measured using EEG. Convolution network has been used at the server and at each client node in FTL. The model is pre-trained on publicly available dataset DEAP on Centralized Machine and is fine-tuned on the K-EmoCon dataset on each client device, without sharing the data of any subject with the centralized model. Valance and Arousal are detected using FTL. On both emotions, the state-of-the-art average F1 score has been achieved.