Enhanced IoT Intrusion Detection Using an Improved Autoencoder and Adversarial Convolutional Encoders

Abstract

To develop an efficient and intelligent automated intrusion detection system for IoT, this study proposes a malicious network traffic recognition model based on an improved autoencoder and adversarial convolutional encoder (AECE). The model first uses mixed sampling and improved autoencoder for data augmentation. Then, convolutional neural networks and gated recurrent units are used to extract spatial and temporal features. AECE combines the idea of generative adversarial networks to enhance the model's adaptability to complex attack patterns. Finally, experimental validation was conducted on the NSL-KDD, UNSW-NB15, IoT-23, and CSE-CIC-IDS2018 datasets. The results showed that the designed data augmentation algorithm could effectively improve the clustering and classification performance of the dataset, with a minimum Xie Beni value of 0.259, a maximum decrease of 15.88% in Davidson Boudin index, and a maximum improvement of 0.214 in classification accuracy. In the IoT-23 dataset, the highest detection rate of the baseline model was 0.882, while the detection rate of the proposed intrusion detection model was 0.949, with an increase of about 7.6%. At the same time, the model had a minimum loss convergence value of 0.08, a response time of 368.16 ms, and the values of false alarm rate fluctuated between 0.10 and 0.20. The comprehensive values of data traffic per second and packet capture per second confirmed that the model had strong detection ability and efficiency for attack behavior. This study expands the application scope of deep learning in anomaly detection, providing new ideas and methods for improving the security and stability of Internet of Things systems.