PROJECT TITLE :

Learning Latent Representation for IoT Anomaly Detection

ABSTRACT:

The Internet of Things (IoT) has recently emerged as a cutting-edge technology that is transforming everyday life for people. However, the rapid and widespread applications of IoT make cyberspace more vulnerable, particularly to IoT-based attacks, in which IoT devices are used to launch attacks on cyber-physical systems. These kinds of attacks are becoming increasingly common. Considering the enormous number of Internet of Things devices, on the order of billions, it is essential to detect and prevent attacks that are based on IoT. However, because of the limited energy and computing capabilities of IoT devices, as well as the continuous and rapid evolution of attackers, this task is extremely difficult to accomplish. When it comes to attacks based on the Internet of Things, the unknown ones are by far the most damaging because they have the potential to outsmart the majority of the existing security systems, and it takes time to discover them and "cure" the systems. In this article, we propose a novel representation learning method to better predictively "describe" unknown attacks, which will make it easier to implement anomaly detection methods that are based on supervised learning. This will allow for more effective detection of new or previously unknown attacks. To be more specific, in order to learn a latent representation from the input data, we develop three regularized versions of autoencoders (AEs). The bottleneck layers of these regularized AEs will then be used as the new input features for classification algorithms after being trained in a supervised manner using normal data and known IoT attacks. This training will be done using normal data. In order to assess how well the proposed models work, we conduct in-depth experiments on nine contemporary Internet of Things datasets. The results of the experiments show that the new latent representation has the potential to significantly improve the performance of supervised learning methods in the detection of unknown attacks against the internet of things. In addition, we carry out experiments to investigate the properties of the proposed models and the impact that hyperparameters have on the performance of the models. The execution time of these models is approximately 1.3 milliseconds, which is sufficient for the majority of applications.