Noise-Resilient Attention-Guided Deep Learning Framework for Network Intrusion Detection

Abstract

The rising complexity of the state of the art network infrastructures and the swift advancement of cyber-attacks have rendered the correct and sound detection of intrusion to be a vital security need. Conventional intrusion detection solutions tend to collapse with noisy, high-dimensional network traffic, temporally changing traffic and encourage superior Deep Learning (DL)-based options. The main goal of this work is to improve the intrusion detection reliability by resolving the issue of noise sensitivity, over-representation of redundant features and inadequate temporal modelling. This is accomplished by introducing a noise-resilient attentionguided DL architecture which are Adaptive Noise-Resilient Data Preprocessing (ANR-DP), Hybrid Recursive Attention-Based Feature Selection (HRAFS), and a Long Short-Term Memory (LSTM)-based temporal classifier. ANR-DP acts dynamically on traffic noise and maintains the characteristics of the attacks, but HRAFS is expected to select stable and security-critical features through recursive elimination and attentiondriven relevance learning. The optimized features are fed through an LSTM network in order to detect sequential and long-term intrusion patterns. Experimental analysis of the CICIDS2017 dataset provides a better performance with Proposed ANR-DP having low Missing Value Ratio of 1.2, High Outlier Reduction Rate of 91.5, HRAFS achieves a Feature Reduction Ratio of 68.7% and 9.1% accuracy improvement. The LSTM classifier also has an accuracy of 98.6%, and a low false positive rate of 3.1%. These findings establish that the proposed framework is very effective in enhancing robustness, precision, and dependability, it can be applicable in the environment of network intrusion detection.