Anomaly Detection Method Based on Semi-Supervised Deep Learning for Internet of Things Traffic

Today, the Internet of Things (IoT) is developing rapidly, with IoT traffic and attacks increasing sharply, posing a serious threat to the stability and security of IoT systems. In this condition, it is crucial to safeguard the normal operation of IoT systems, while traditional machine learning methods usually require manual selection of features with low anomaly detection rates. To tackle the issue of significant deviation in anomaly detection results caused by the imbalance of data samples in IoT traffic, an IoT traffic anomaly detection model (DCGAN-BiGRU) was proposed. Deep Convolutional Generative Adversarial Networks (DCGAN) were utilized to learn the inherent feature distribution of known attack sample data for the generation of new attack samples. Therefore, the disadvantage of unbalanced data samples was overcome and the training stability was improved. Bi-Gated Recurrent Unit (BiGRU) consists of two stacked GRU layers. Then the ability of the model to learn more complex patterns in the sequential data was enhanced. The CICIDS2018 and the CICIoT2023 datasets were used as experimental samples to evaluate the model. The DCGAN-BiGRU intrusion detection model enhanced the efficiency and precision of traffic anomaly detection when compared to traditional methods.