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

With the continuous proliferation of emerging technologies such as cloud computing, 5G networks, and the Internet of Things, the field of cybersecurity is facing an increasing number of complex challenges. Network intrusion detection systems, as a fundamental part of network security, have become increasingly significant. However, traditional intrusion detection methods exhibit several limitations, including insufficient feature extraction from network data, high model complexity, and data imbalance, which result in issues like low detection efficiency, as well as frequent false positives and missed alarms. To address the above issues, this article proposed an adversarial intrusion detection model (Soft Adversarial Asynchronous Actor-Critic Intrusion Detection, SA3C-ID) based on reinforcement learning. Firstly, the raw dataset is preprocessed via one-hot encoding and standardization. Subsequently, the refined data undergoes feature selection employing an improved pigeon-inspired optimizer (PIO) algorithm. This operation eliminates redundant and irrelevant features, consequently reducing data dimensionality while maintaining critical information. Next, the network intrusion detection process is modeled as a Markov decision process and integrated with the Soft Actor-Critic (SAC) reinforcement learning algorithm, with a view to constructing agents; In the context of adversarial training, two agents, designated as the attacker and the defender, are defined to perform asynchronous adversarial training. During this training process, both agents calculate the reward value, update their respective strategies, and transfer parameters based on the classification results. Finally, to verify the robustness and generalization ability of the SA3C-ID model, ablation experiments and comparative evaluations are conducted on two benchmark datasets, NSL-KDD and CSE-CIC-IDS2018. The experimental results demonstrate that SA3C-ID exhibits superior performance in comparison to other prevalent intrusion detection models. The F1-score attained by SA3C-ID was 92.58% and 98.76% on the NSL-KDD and CSE-CIC-IDS2018 datasets, respectively.