Two-Level Feature Selection for Enhanced Accuracy and Reduced Computational Overhead in Intrusion Detection Systems Using Rough Set Theory and Binary Particle Swarm Optimization

Intrusion Detection Systems (IDS) are essential for safeguarding network infrastructures by detecting and mitigating malicious activities. This study introduces a two-level feature selection approach (TLFSA) designed to enhance classification accuracy and reduce computational overhead. The first phase employs Rough Set Theory (RST) to filter out irrelevant features, while the second phase uses Binary Particle Swarm Optimization (BPSO) to refine the feature subset based on their discriminative power. Experiments conducted on the NSL-KDD dataset show that the TLFSA approach outperforms traditional algorithms such as Genetic Algorithm (GA) and Gravitational Search Algorithm (GSA), achieving a notable improvement of 0.99% in classification accuracy. Furthermore, class-specific feature subsets produced by the method demonstrate superior detection rates across all network traffic classes, with an average accuracy of 97.22%, compared to 91.11% for alternative methods. The proposed method effectively reduces the feature set to approximately 15% of the original features, streamlining the IDS model and improving both operational efficiency and real-time applicability.