Virtual Trust Algorithm to Mitigate Cyberattacks in Iot Networks

Abstract

There is drastic increase in needs of networking and data sharing in today’s world. Such
globalization of increased information technology and development there exists need of
network security. Firewalls may provide some level of security but they never alert
administrator for upcoming attacks. In order to find such abnormal behavior of network
packets there is need of reliable detection system for improvement of efficiency and
accuracy. As in today’s developing network environment there is threat of new type of
attacks daily in the network. So, the network administration system is also needed to be
updated regularly for upgradation of security level. One of the network packet monitoring
system is Anomaly Detection Systems (ADS). Many research works focused on machine
learning approach for enhancing the efficiency of the Anomaly detection system and to
detect malicious network activity automatically on the basis of network packet behaviors.
The proposed model is designed using machine learning approach for detection of malicious
activities of the network packets. For that KDD-99 dataset is used. First of all the dataset is
normalized for reducing calculation complexity, further features are reduced using corelation
algorithm. The reduced features determine that only efficient features can be used
for malicious behavior detection. From result analysis it is seen that while selecting more
than 15 features PSO outperforms better whereas below 15 features co-relation outperforms
best. After feature reduction data clustering is performed using k-mean clustering algorithm.
By using clustering, small datasets is built that represents the entire original dataset which
can expressively reduce the training time of classifiers and improve the efficiency.In final
step of proposed algorithm multilevel hybrid classifiers, based on DNN, are designed for
classification of dataset into five attack categories i.e. DOS, U2R, R2L, Probe and Normal.
As compared to some other multilevel classifier work the proposed algorithm proves its
efficiency in terms of high accuracy, high detection rate and False Alarm Rate (FAR).