A deep reinforcement based echo state network for network intrusion classification.

Area of Science:

• Cybersecurity
• Machine Learning
• Network Security

Background:

• Network intrusion classification is crucial for identifying suspicious activities.
• Traditional methods struggle with dynamic and evolving attack patterns.
• Class imbalance in datasets poses a significant challenge for intrusion detection systems.

Purpose of the Study:

• To propose a novel deep reinforcement learning (DRL) approach for network intrusion classification.
• To integrate Echo State Networks (ESN) with DRL for adaptive threat detection.
• To evaluate and enhance data balancing techniques for imbalanced network intrusion datasets.

Main Methods:

• Utilized a deep reinforcement learning (DRL) framework combined with Echo State Networks (ESN).
• Evaluated advanced data balancing techniques: Borderline-SMOTE, SMOTE-ENN, ADYSN, and K-means SMOTE.
• Conducted multi-dataset validation on benchmark datasets (NF-BoT-IoT, NF-UNSW-NB15, etc.) and adaptive modeling tests.

Main Results:

• The K-means-based data balancing method demonstrated superior performance over other techniques.
• The DRL-ESN approach showed improved accuracy and reliability in detecting novel and evolving threats.
• Robust performance was validated across multiple diverse network intrusion datasets.

Conclusions:

• The proposed DRL-ESN approach offers a robust and adaptive solution for network intrusion classification.
• Effective data balancing, particularly K-means SMOTE, is critical for enhancing intrusion detection performance.
• The approach provides a viable solution for securing modern network infrastructures against sophisticated cyber threats.