A Multi-level Features Fusion Model for Network Communication based on Machine Learning

Abstract

Today's societies couldn't function without elaborate networks of communication. Many problems remain unresolved, but novel approaches to these problems are constantly being offered. Many of the problems plaguing existing works, such as high characteristic design cost, challenging feature selection, poor real-time performance, etc., stem from their focus on a wide range of characteristics. Worse still, the difficulty in training models due to data imbalance results in a poor detection rate for aberrant samples. To achieve a more effective and robust model, we present a multi-level feature fusion (MFFusion) model that utilizes a combination of data temporal, byte, and statistical characteristics to extract relevant information from different angles. Too far, MFFusion has outperformed the state-of-the-art on several real-world network datasets in terms of prediction performance and false alarm rate. We also use MFFusion for anomaly detection in an IoT network, using the most recent IoT malicious traffic information. The experimental results demonstrate the adaptability of MFFusion and its suitability for identifying network anomalies in an IoT context with system performance.