202111 bcwh

Intrusion detection systems are widely implemented to protect computer networks from threats. To identify unknown attacks, many machine learning algorithms like neural networks have been explored for anomaly based detection. However, in real-world applications, the performance of classifiers might be fluctuant with different data sets, while one main reason is due to some redundant or ineffective features. To mitigate this issue, this study investigates some feature selection methods and introduces an ensemble of Neural Networks and Random Forest to improve the detection performance. In particular, we design an intelligent system that can choose an appropriate algorithm in an adaptive way. In the evaluation, we study the feasibility of our approach with KDD99 data set and evaluate its practical performance with a real data set collected from a Honeynet environment. The experimental results indicate that as compared with similar approaches, our approach can overall provide a better result, through identifying important and closely related features.