Development of artificial ıntelligence based cyber security approaches in ındustry 4.0 systems

The concept of Industry 4.0 is emerged with the increase in production in the industrial area, the need for high productivity in production and the idea of making use of technology in the industrial field. Internet technology is used in the communication of industrial systems with each other in Industry 4.0. The data obtained from the systems with the Internet can be used in real time. Thanks to the internet technology, the operation of the systems are automated. Thus, human errors are minimized. Remote control and management of systems are begun to be provided but internet technology enables critical systems to communicate with each other while also causing these systems to be exposed to cyber attacks. Protecting the confidentiality, integrity and accessibility of the system network is extremely important because a cyberattack that may occur can cause more destruction than expected. For this purpose, a study has been carried out to protect systems against cyber attacks, which increase in proportion to the developments in the industry, by using artificial intelligence-based learning algorithms. In this thesis, it was mentioned that intrusion detection systems can be developed by using artificial intelligence-based learning algorithms against increasing cyber security threats. Traditional signature-based intrusion detection systems and learning-based intrusion detection systems is compared. Support vector machine, Naive bayes, Random forest, Decision trees, K nearest neighbor, Logistic regression algorithms are used in the prediction of attacks that can be made to networks used in the industrial field. The algorithms used in this thesis study were trained with a part of the data set obtained from the traffic of the network used in the communication of industrial systems and tested with the remaining data set. The performance of the algorithms are compared with performance criteria such as precision, sensitivity, F1 score and accuracy. Precision, sensitivity and F1 score metrics are also calculated as the accuracy criterion alone isn't sufficient in performance criteria. Using the same data sets in the training and testing phases of the algorithms in this study may cause the algorithms to memorize the data set after a while. In order to prevent this situation, 10 training/test data sets were obtained randomly from the data set by using the k-fold cross validation method. A general accuracy performance value was obtained by averaging the accuracies of the obtained data sets. One training / test data accuracy criterion was compared with the accuracy criterion obtained as a result of k-fold cross validation. When the results of all performance criteria are examined, it was seen that the algorithms that achieve the best results in attack prediction are decision tree and random forest algorithms.