A Synergistic Approach to Detect Malicious Behaviour through Correlation-Based Feature Selection with Deep Learning

Malware presents a persistent cybersecurity threat by infiltrating systems to steal sensitive data. Detecting malware traffic is challenging due to its consistent asymmetry, distinguishing it from benign traffic. This paper proposes an advanced malware detection system using deep learning and feature selection techniques. A Bidirectional-GRU approach is introduced, which improves model performance by effectively selecting and removing specific features, outperforming traditional dense-layer methods. The effectiveness of feature selection is demonstrated by enhancing the performance of traditional models such as Naive Bayes and XGBoost, surpassing Gradient Boosting. Our results indicate that at a 0.6 threshold with the selection of 21 features, Bidirectional-GRU achieved a performance of 99.98%, slightly outperforming other models. Additionally, XGBoost, GRU, and Random Forest also demonstrated strong performance in malware detection. The study highlights threshold-specific performance, showing that Bi-GRU excels at lower thresholds, while GRU performs optimally at higher thresholds, showcasing adaptability to varying configurations. Key features used in the model include stime, mm users, gtime, among others, which contributed significantly to the improved detection accuracy.