Disentangled Adversarial Generalization Network for Cross-Session Task-Independent Brainprint Recognition

Capitalizing on the invisibility of electroencephalography (EEG) signals, brainprint has emerged as a promising EEG-based biometric technology that can meet scenarios with high-security requirements. Most studies have focused on single cognitive tasks while spontaneous brain activity would be affected by mental processes, making identity and task information spuriously correlated. Furthermore, the temporal variability of EEG signals may lead to differences in data distribution of multiple sessions. Then, the training model fails to meet the unseen cross-session and cross-task data, limiting the use of brainprint recognition in reality. In this article, we proposed a disentangled adversarial generalization network (DAGN) for stable task-independent brainprint recognition across sessions. Specifically, we first facilitate disentangling task-relevant and identity-relevant features via decorrelation to get rid of spurious correlations. An adversarial self-challenging strategy is then designed to challenge the activation of remaining label-related features in addition to dominant features. This naturally balances the contribution of dominant and inferior dimensions of discriminative features to alleviate the poor robustness of model in different unseen data distributions. Extensive experiments on the representative multitask benchmarks with challenging leave-one-task-out and leave-test-session-out validation are carried out to demonstrate that our method performs favorably against the state-of-the-art approaches.