A Novel Hybrid Deep Learning Framework with Multi-Scale Temporal Convolutions, Attention, and Uncertainty Quantification for Network Intrusion Detection

Abstract

This paper presents a novel hybrid deep learning framework for network intrusion detection that addresses the limitations of existing systems in detecting both known and novel attacks. The proposed architecture integrates six sequential modules: data preprocessing, adaptive feature selection, Multi-Scale Temporal Convolutional Network (MS-TCN), Bidirectional LSTM with attention mechanism, ensemble anomaly detector, and uncertainty-aware classifier. The framework employs parallel 1D convolution layers with varying kernel sizes (3, 5, 7) to capture temporal patterns of different complexities, while BiLSTM processes sequential dependencies bidirectionally. An ensemble of three anomaly detectors (Isolation Forest, DBSCAN, One-Class SVM) handles zero-day attacks through majority voting. Experimental results demonstrate superior performance with 91% accuracy, 0.9783 ROC-AUC, and 0.9895 average precision. The adaptive feature selection reduces dimensionality by 50% while maintaining discriminative power. The model effectively balances precision (0.99 for attacks) and recall (0.81 for attacks), showing robust generalization without overfitting across ten training epochs.