Investigating the Effects of Hyperparameter Tuning Process on the Performance of Intrusion Detection Systems

Abstract

As the threat landscape in cybersecurity evolves, the effectiveness of intrusion detection systems becomes paramount. This article delves into the realm of enhancing the performance of machine learning classifiers, specifically Random Forest (RF), XGBoost, and LightGBM (LGBM), by advocating the utilization of hyperparameter tuning processes. Leveraging the Optuna library, the default settings were employed to explore the impact on attack detection efficiency. Results demonstrated a notable improvement in false negative rates, crucial for accurate identification of attacks. While achieving commendable enhancements, the study also sheds light on the observed trade-offs, particularly an increase in false positive values in certain algorithms. This research serves as a critical step toward refining intrusion detection systems, with implications for the ongoing pursuit of robust and balanced security algorithms.