A Framework for Anomaly Detection in IoT Networks Using Conditional Generative Adversarial Networks

While anomaly detection and the related concept of intrusion detection are widely studied, detecting anomalies in new operating behavior in environments such as the Internet of Things (IoT) is an active field of research. Anomaly detection models trained on datasets that are likely imbalanced have poor results, but the ability of Generative Adversarial Networks (GANs) to emulate complex high-dimensional distributions seen in real-world data suggests that they may be effective for anomaly detection. This paper proposes a novel framework for detecting anomalies in IoT networks utilizing conditional GANs (cGANs) to build realistic distributions for a given feature set to overcome the issue of data imbalance. To this end, a single class cGAN (ocGAN) was utilized to learn the minority data class to balance the dataset. Then, the binary class cGAN (bcGAN) model generates augmented data for the binary balance dataset. The performance of the ocGAN and bcGAN models in binary and multiclass classification environments were evaluated using a feed-forward neural network (FFN) and tested on two network-based anomaly detection datasets and five IoT network-based anomaly detection datasets. The proposed models outperformed other anomaly detection models in the standard metrics of accuracy, precision, recall, and F1-score.