Generative Adversarial Network and Auto Encoder based Anomaly Detection in Distributed IoT Networks

Abstract

With the advances in modern communication technologies, the application scale of Internet of Things (IoT) has evolved at an unprecedented level, which on the other hand poses threats to the IoT ecosystem. As the intrusions and malicious actions are becoming more complex and unpredictable, developing an effective anomaly detection system, considering the distributed nature of IoT networks, remains a challenge. Moreover, the lack of sufficiently large amount of data samples of IoT traffic and data privacy pose further challenges in developing a behavior-based anomaly detection system. To address these issues, we present an unsupervised hierarchical approach for anomaly detection through cooperation between generative adversarial network (GAN) and auto-encoder (AE). The problems of data aggregation and privacy preservation are addressed by reconstructing a sampling pool at a centralized controller using a collection of generators from the individual IoT networks. Then, a centralized global AE is trained and passed to individual local networks for anomaly detection after a final adaptation with the local raw data from the IoT nodes. The performance is evaluated using the UNSW Bot-IoT dataset and the results demonstrate the effectiveness of our proposed approach which outperforms other approaches.