Deep Network Based Classifiers using Regularization Techniques for detecting cyber security threats in IOTs devices

Abstract

Introduction: AI (Artificial intelligence) is now woven into nearly every facet of our lives, sparking innovation, and trans forming industries. Over the last few years, the ML (machine learning) and IoT (Internet of Things) have surged in prominence, becoming essential across numerous fields, ranging from manufacturing and transportation to agriculture and healthcare. Their applications span from fore casting sales to bolstering security measures and devices of monitoring. However, this growing focus on IoT technique has brought about important security concerns that must be addressed to ensure the reliability of these systems.

Objectives: Our study seeks to address a significant gap in the area by presenting a comprehensive strategy. By thoroughly exploring neural networks (NN) along with validating our findings against the MALNET-IMAGE, Virus-MNIST, Malimg and collection datasets, author aspires to push the boundaries of IoT threat detection. This unified framework, shaped by insights from the existing research we reviewed, is developed to provide exceptional resilience against the evolving landscape of threats.

Methods: A new model of deep network based on different classifiers and stacked through RBM strategy (Restricted Boltzmann Machine) for cybersecurity threat detection is suggested and investigated in this paper. Additionally, classifiers layer has been inserted which can classify the images. Authors then ensure the structurer of deep neural network (DNN) against over-fitting due to mighty dropconnect and dropout performance. Training with both techniques of regularization, with  randomly chosen weights / activations subsets have been dropped.

Results: The evaluation over the datasets of MALNET-IMAGE, Virus-MNIST, Malimg and collection datasets to manage threat level displays a classification error rate development once utilizing deep network trained with dropconnect or dropout.

Conclusions: The merits of this strategy over traditional strategies are as follows: (1) experimental outcomes show that the proposed model exceeds traditional classification and other approaches of DNN; (2) the networks of the recommended architectures have deep construction and therefore the competence of extraction of feature is powerful than traditional classifiers.