Anomaly Detection in IoT Networks: A Deep Learning Approach Using Autoencoders

The current fast proliferation of the Internet of Things (IoT) networks has made anomaly detection and security more difficult. Traditional methods are not able to detect hostile activities because they cannot understand dynamic, complex, and high-dimensional data. In this paper, a deep learning framework-based autoencoder is designed to efficiently identify the anomalies primarily focusing on attack types in IoT environments. This model identifies the deviations through reconstruction error metrics, dimensionality reduction and studies normal behavior. The study experiments and observes how the performance of machine learning algorithms varies according to the way of distribution of the datasets, and it was found that the Random forest and the Decision tree are efficient in most of the distributions. The analysis of various loss function types also revealed that, although the Arcface and AM Softmax loss functions are noticeably more accurate and efficient, the All Combined Loss Function does remarkably well in clustering. Moreover, while keeping low false positive rates, the proposed design exhibits sensitivity against significant attack types. Considering its efficiency, this framework can be used for real-time applications in a range of IoT scenarios. Furthermore, we compared the proposed model with that of Isolation Forest, and show that, the proposed model outperforms it significantly.