FedEmo: A Privacy-Preserving Framework for Emotion Recognition using EEG Physiological Data

Emotions are intricate mental states triggered by neurophysiological adjustments linked to ideas, sensations, behavioral reactions, and a level of pleasure or annoyance. These changes are best traced with the physiological signal Electroencephalogram (EEG), as it records the direct sensations sent by the brain. Recent research on emotion classification methods employs conventional machine learning classifiers to access human emotions and perform automatic emotion recognition tasks. However, they lack in securing users' privacy and sensitive information because they need access to all data. A newly introduced framework Federated Learning (FL), can resolve this problem. It is an approach that aims to create a global model classifier without requiring access to users' local data. This study proposes a novel FL framework, Federated learning for Emotion recognition (FedEmo), for emotion state classification from physiological signal EEG while preserving users' data privacy. It uses Artificial Neural Network (ANN) as a baseline model for classifying emotional states: Arousal, Valence, and Dominance. Adding the concept of federated learning to build a framework FedEmo prevents loss of privacy as it enables the local training on the client's end with an updated model from the global server without compromising privacy. The proposed FedEmo framework approach achieves accuracies of 63.3%, 56.7%, and 52.2% for Valence, Arousal, and Dominance, respectively, using the well-known DREAMER dataset. These results are comparable to the basic centralized ANN model with the additional development of privacy preservation.