An Online Intrusion Detection Method using Adaptive Multi-Level Classifier Network and PCA-Guided Model Reuse Mechanism
Haizhaoyang Huang, Hongpo Zhang · 2025
The network era caused the network security concerns to escalate dramatically, and network intrusion detection is shown as an important technology to guarantee network security. During the past few years, artificial intelligence technology has improved remarkably, and also involve advances from various studies, which started to adapt deep learning with intrusion detection models and obtained some salient detection results. However, the Internet is a developing environment, concept drift will unavoidably be produced which may result in model degeneration. Meanwhile, to tackle these difficulties, we propose an online adaptive intrusion detection method and the key point is a model reuse mechanism and an adaptive multi-level classifier network. The reuse of model exploits the principal component analysis to identify the most similar historical distribution to the current data and adopts different strategies with respect to different degree of drift to reduce the computational overhead avoid the model retraining. Adaptive multi-level classifier network consists of several classifiers with different depth layer. These classifiers dynamically adapt their adaptive weights based on their classification losses, and the final classification is calculating the weighted output of multiple layers. Experiments conducted on CIC-IDS2017 and UNSW-NB15 datasets and demonstrate that the proposed method achieves an average accuracy of 97.69% and 95.96%, respectively, which superior to that of state-of-the-art methods: