EEG-based brainprint technology, capitalizing on the inherent invisibility of EEG signals, has garnered significant attention as a promising biometric approach capable of meeting the stringent security requirements of various high-stakes applications. Traditional studies in this domain have largely focused on identity recognition based on single mental tasks, yet such tasks often introduce interference from spontaneous brain activity, which in turn causes spurious correlations between identity and task-related information. This issue is further compounded by the inherent temporal variability of EEG signals, which leads to significant data distribution differences across sessions. Consequently, existing models frequently fail to generalize unseen data from different sessions or tasks, severely limiting the practicality and scalability of brain fingerprint identification systems in real-world scenarios. In this chapter, we propose the Disentangled Adversarial Generalization Network (DAGN), a novel deep learning framework aimed at achieving stable and robust task-independent brain fingerprint identification across sessions. The core innovation of the DAGN lies in its ability to disentangle identity-relevant and task-relevant features through a decorrelation mechanism. This process effectively eliminates the spurious correlations between the two, ensuring that identity information is isolated from task-induced variability. To further enhance the generalization ability of the model, we introduce an adversarial self-challenging strategy that penalizes the activation of task-related features, forcing the network to focus exclusively on identity-specific information. This approach significantly improves the robustness of the learned features, enabling the DAGN to maintain high performance when applied to unseen data across different tasks and sessions. Extensive experiments, conducted on representative multi-task benchmarks with challenging leave-one-task-out and leave-one-session-out cross-validation protocols, demonstrate the superiority of our approach over state-of-the-art methods in terms of generalization performance.