TRINITY-IDS: a tri-view hybrid architecture with adaptive drift-evidential learning for robust and continual intrusion detection

The rapid evolution of cyber threats in modern network environments demands intelligent intrusion detection systems. This system should be capable of handling complex attack patterns and heterogeneous data distributions. Existing intrusion detection approaches mainly focus on statistical or temporal flow characteristics. This approach neglects the topological relationships among network entities and the uncertainty associated with detection decisions. In this work, a Tri-View Hybrid Architecture with Adaptive Drift-Evidential Learning (TRINITY-IDS) is proposed for robust intrusion detection. Initially, a Density-Aware Hybrid Resampling strategy based on Cluster-SMOTE and Edited Nearest Neighbor (ENN) cleaning is applied to overcome class imbalance and improve minority attack representation. An adversarial feature selection mechanism with gradient reversal learning is used to reduce dataset-specific feature leakage and enhance cross-domain generalization. The extracted features are represented through three complementary views: a statistical view using convolutional feature extraction with channel attention, a temporal view using speciality-token-augmented BiGRU with self-attention for sequential behavior modeling, and a topological view using Graph Attention Networks (GAT) to capture network relationship patterns. A dynamic cross-view adaptive fusion mechanism automatically learns the contribution of each representation for individual samples. In addition, a Dirichlet evidential learning head is used for uncertainty-aware attack classification. It estimates belief and uncertainty values rather than only deterministic predictions. To support evolving network conditions, an online Page-Hinkley drift detector and experience replay mechanism are incorporated for continual adaptation during streaming intrusion detection. Experimental evaluation demonstrates that the proposed TRINITY-IDS achieves an overall accuracy of 98.36%, precision of 98.26%, recall of 98.32%, and F1-score of 98.29%.