TMK-TRAF: A Multi-Task Framework with Threat-Risk Adaptive Fusion for Integrated Banking Intrusion and Fraud Detection

Abstract

Banking security requires simultaneous detection of network-level intrusions and transaction-level fraud, yet existing approaches typically address these threats independently and fail to exploit their shared security characteristics. This paper proposes TMK-TRAF, a multi-task deep learning framework that integrates Temporal Convolutional Network (TCN), Mamba, and Kolmogorov–Arnold Network (KAN) to learn complementary local-temporal, long-range, and nonlinear representations. The proposed Threat-Risk Adaptive Fusion (TRAF) mechanism performs task-specific fusion of these representations, while the Threat-Risk Consistency (TRC) loss promotes beneficial cross-task knowledge while preserving task-specific decision boundaries. A downstream BiLSTM–GRU–MLP ensemble performs intrusion and fraud classification, with SHAP providing feature-level interpretability. Network-flow data are generated using OMNeT++/INET, while banking fraud data are used for transaction-level evaluation, and CloudSim is employed for simulated scalability analysis. TMK-TRAF achieves 97.68% intrusion-detection accuracy with 97.26% F1-score and 98.30% fraud-detection accuracy with 93.12% F1-score. The results demonstrate the effectiveness of unified multi-task representation learning, adaptive fusion, and cross-task consistency for integrated banking-security analytics.