SA3C-ID: a novel network intrusion detection model using feature selection and adversarial training.

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Network Security

Background:

• Emerging technologies like IoT and 5G introduce complex cybersecurity challenges.
• Traditional intrusion detection systems struggle with feature extraction, complexity, and data imbalance, leading to inefficiencies.
• There is a need for advanced network intrusion detection methods to overcome existing limitations.

Purpose of the Study:

• To propose a novel adversarial intrusion detection model, Soft Adversarial Asynchronous Actor-Critic Intrusion Detection (SA3C-ID), based on reinforcement learning.
• To address limitations of traditional methods, including insufficient feature extraction, high model complexity, and data imbalance.
• To enhance the efficiency and accuracy of network intrusion detection systems.

Main Methods:

• Data preprocessing using one-hot encoding and standardization.
• Feature selection via an improved pigeon-inspired optimizer (PIO) algorithm.
• Modeling intrusion detection as a Markov decision process integrated with Soft Actor-Critic (SAC) reinforcement learning for adversarial training with attacker and defender agents.

Main Results:

• The SA3C-ID model demonstrated superior performance compared to prevalent intrusion detection models.
• Achieved high F1-scores: 92.58% on NSL-KDD and 98.76% on CSE-CIC-IDS2018.
• Ablation experiments and comparative evaluations validated the model's robustness and generalization ability.

Conclusions:

• The SA3C-ID model effectively addresses the limitations of traditional intrusion detection methods.
• The proposed adversarial reinforcement learning approach significantly improves detection efficiency and reduces false alarms.
• SA3C-ID offers a robust and generalized solution for modern network intrusion detection.