CAGCNet: generalized contrastive learning for person identification based on channel aggregated EEG features

Person identification method based on electroencephalograms (EEG) signals, or so called brainprint recognition is a novel way to distinguish identities with advantages of high security. However, existing methods neglect the distribution difference between training and test data, and the large distance between projected features in the latent space makes the performance of the model degrade in the unseen domain data. In this paper, we propose channel aggregated based generalized contrastive learning framework, which combines multiple modules to overcome this challenge. To capture features from different granularities, we involve multi-scale convolution with channel attention block. In face of distribution of unseen domain, we introduce feature enhancement-based generalized contrast learning to improve the model generalization ability. In the generalized contrast learning module, taking the difficulty of reconstructing EEG signals into consideration, we augment the source domain data at the feature level to improve the generalization ability of the model on the unseen domain data. Extensive experiments on two multi-session datasets shows that our model outperformed other baseline methods, demonstrating its capability of better generalization performance to unseen domain.