FedCER - Emotion Recognition Using 2D-CNN in Decentralized Federated Learning Environment

Emotion recognition using physiological signals has received much attention in recent literature. However, current development relies on the use of centralized datasets for training prediction models. But this approach raises a significant risk of privacy violation, especially in cases where the researchers use medically sensitive data like EEG recordings. The following paper proposes a privacy-preserving emotion recognition framework using Federated Learning. It is a decentralized method of training machine learning models. We validate our results by comparing them against a baseline model and discuss the privacy-performance trade-off in Federated Learning. Our proposed model is a convolutional neural network that works upon EEG signal recordings directly and does not rely upon extracted features from the DEAP dataset recordings of each subject. Instead, we have kept the non-IID data in the dataset intact. The proposed architecture achieves 72.22 percent, 70.10 percent, and 66.99 percent accuracy scores for the Dominance, Arousal, and Valence labels on the public DEAP dataset.