Internet of Things (IoTs) Security: Intrusion Detection using Deep Learning

Sahingoz O. K. , Cekmez U., BULDU A.

JOURNAL OF WEB ENGINEERING, vol.20, no.6, pp.1721-1760, 2021 (Journal Indexed in SCI) identifier identifier

  • Publication Type: Article / Article
  • Volume: 20 Issue: 6
  • Publication Date: 2021
  • Doi Number: 10.13052/jwe1540-9589.2062
  • Page Numbers: pp.1721-1760
  • Keywords: Convolutional neural networks, deep learning, imbalanced datasets, Internet of Things, IoTs, web security


With the development of sensor and communication technologies, the use of connected devices in industrial applications has been common for a long time. Reduction of costs during this period and the definition of Internet of Things (IoTs) concept have expanded the application area of small connected devices to the level of end-users. This paved the way for IoT technology to provide a wide variety of application alternative and become a part of daily life. Therefore, a poorly protected IoT network is not sustainable and has a negative effect on not only devices but also the users of the system. In this case, protection mechanisms which use conventional intrusion detection approaches become inadequate. As the intruders' level of expertise increases, identification and prevention of new kinds of attacks are becoming more challenging. Thus, intelligent algorithms, which are capable of learning from the natural flow of data, are necessary to overcome possible security breaches. Many studies suggesting models on individual attack types have been successful up to a point in recent literature. However, it is seen that most of the studies aiming to detect multiple attack types cannot successfully detect all of these attacks with a single model. In this study, it is aimed to suggest an all-in-one intrusion detection mechanism for detecting multiple intrusive behaviors and given network attacks. For this aim, a custom deep neural network is designed and implemented to classify a number of different types of network attacks in IoT systems with high accuracy and F-1-score. As a test-bed for comparable results, one of the up-to-date dataset (CICIDS2017), which is highly imbalanced, is used and the reached results are compared with the recent literature. While the initial propose was successful for most of the classes in the dataset, it was noted that achievement was low in classes with a small number of samples. To overcome imbalanced data problem, we proposed a number of augmentation techniques and compared all the results. Experimental results showed that the proposed methods yield highest efficiency among observed literature.