Accelerated Intrusion Detection Using Hybrid Feature Optimization With Backward Elimination And Temporal Gradient Framework

Abstract

The exponential growth of web-based infrastructures and the rise in sophisticated cyber threats have placed Intrusion Detection Systems (IDS) under immense performance pressure. Traditional IDS models, while accurate, often suffer from high processing latency due to large and redundant feature spaces. This research introduces a novel hybrid framework to enhance IDS performance by integrating three advanced feature selection algorithms: Hybrid Feature Subset Selection Algorithm (HSSA), Adaptive Mutual Relevance Pruning (AMRP), and Temporal Gradient-Based Feature Pruning (TGBFP). Each algorithm addresses specific bottlenecks related to detection speed, redundancy, and model interpretability. The proposed models are benchmarked using the NSL-KDD dataset, incorporating both static and dynamic attack scenarios. HSSA, based on domain-informed backward elimination, achieves the highest detection accuracy ( 93.24% ) with the lowest response time (2.3s). AMRP optimally balances relevance and redundancy, while TGBFP dynamically filters features during model training via real-time gradient feedback. Together, these methods reduce false positives and processing load without sacrificing classification fidelity. Experimental validation through precision-recall analysis, ROC curves, and performance histograms confirms the efficacy of the proposed hybrid approach. The framework offers a scalable, real-time detection mechanism adaptable to diverse network environments.