IoT network traffic classification: a deep learning method with Fourier transform-assisted hyperparameter optimization

Abstract

There has been a significant increase in IoT-related network traffic in recent years. The surge in IoT has resulted in a more complex network environment than ever before. In light of this, deep learning (DL)-based network traffic classification has gained prominence, because of its powerful feature extraction capabilities for complex problems. However, selecting hyperparameters for DL models, such as network depth, lacks a theoretical basis and costs a lot of time. Often, the setting of hyperparameters is not directly related to the inherent characteristics of the data but relies on empirical knowledge. Traditionally, hyperparameters are adjusted based on performance during model training, leading to a significant amount of tuning work. To address these problems, this paper proposes a novel DL-based anomaly network traffic classification algorithm. This algorithm estimates the required hyperparameters by analyzing the spectrum obtained from Fourier Transform of the input samples in advance, enhancing the efficiency of model training for IoT network traffic classification. Our experiments reveal that the complexity of the neural network is directly proportional to the spectrum of the data being trained. As the presence of high-frequency components increases, the complexity needed for the neural network parameters also rises. Based on the conclusions drawn from our experiments, we can pre-determine appropriate hyperparameters for the neural network, thereby saving over 70% of the time in neural network parameter tuning.