Adversarial AI Based Framework for Enhancing Security of IoT Networks

The development of Internet of Things (IoT) technologies results in the widespread adoption and extensive installation of IoT systems throughout the globe. IoT systems can offer sophisticated products and solutions, but they also pose significant security risks due to the sheer amount of personally identifiable information they gather and process. Intelligent network security system design has received a lot of attention as a means of preventing IoT data exploitation in advanced applications. However, while training the detection model, existing systems may have insufficient and uneven attack data, which leaves the system vulnerable, particularly to attacks that are undetermined in nature. In this regard, the present research is focused on identifying the adversarial attacks in IoT devices using Artificial Intelligence (AI) algorithms. The adversarial samples are generated using Iterative Fast Gradient Sign Method, Jacobian-based Saliency Map Method, Limited-memory Broyden-Fletcher-Goldfarb-Shanno Method and Projected Gradient Descent Method. The dimensions of the features are minimized using Variational Generative Adversarial Networks (VGAN). The adversarial attacks in the IoT systems are detected effectively using Fully Deep Connected Convolutional Neural Networks (FDCCNN). The efficacy of the model is evaluated on DS2OS dataset and achieved an accuracy of 98.5%.