Abstract:
Excessive and exponential deployment of the Internet of Things (IoT) devices as wireless network makes it difficult to secure
data. It is a high time to make IoT network vulnerable against various attacks. Since multiple vulnerabilities are simultaneously available
in the IoT network, therefore, classifying the IoT vulnerabilities is multi-class problem. The paper contributed to design and develop a deep
learning (DL) IoT security framework to classify and predict vulnerabilities. The principal component analysis (PCA) is used for feature
set reduction and then rectified linear unit (ReLU) activation is used for eliminating vanishing gradient problem for binary classification.
Further, the binary classification performance is compared with multi class classification models. Determining benignity or maliciousness
of a sample is the main objective of detection models. The classification algorithms' objective is to assign each sample into one of the
subsequent classes: beneficent, tsunami, Mirai, or Gafgyt. The classification of attacks is implemented using the deep neural network (DNN)
based learning model. The N-BaIoT multiple vulnerabilities dataset is used which is of unbalanced nature. The paper suggested employing
the oversampling approach Synthetic-Minority Oversampling Technique (SMOTE) for balancing the data set and compared performance
with Random Oversampling and Undersampling (ROU) approach. Performance is compared with conventional decision tree, Random
Forest (RF). The accuracy with proposed DNN approach is more than 99.99% in all cases.
