Veri madenciliği ve makine öğrenme temellisaldırı tespit modeli

Recently, the use of online services has grown rapidly, which imposes the need to protect servers that provide these services without affecting the quality of these services. Traditional network protection techniques are no longer applicable, according to the development of the intrusion techniques being used by intruders. Thus, more complex techniques are being used to provide better protection to these networks. Data mining is one of the machine learning fields that can be used to extract relations between packets information and the labels given to them. Thus, in this study, three different data mining classification techniques, which are the Support Vector Machine, Random Forest and Feed-Forward Neural Networks are evaluated to detect anomalies in the packets incoming to the network. Then, the type of attack being executed is also detected by these classifiers, in case an intrusion is detected. The results show that the feed-forward deep neural network classifier, with only three hidden layers of 32 neurons each, has the best overall performance with a predictions accuracy of 99.27% in binary classification with an average prediction time of 0.7 uSec per each prediction, while the Random forest classifier, with 100 trees in the forest, has scored an accuracy of 99.60% but consumes an average of 8.54 uSec per each prediction, which is extremely high time compared to the deep learning model. Moreover, the support vector machine classifier has scored an accuracy of 98.70% and an average execution time of 218.3 uSec per each prediction. Moreover, in multi-class classification, the deep learning model with the same hidden layers has shown the best prediction accuracy and time with 90.82% accuracy and 0.89 uSec average prediction time, while the random forest classifier achieved an accuracy of only 87.92% consuming an average of 17.28 uSec per prediction and the support vector machine classifier has a prediction accuracy of 70.43% and consumes an average of 709.65 uSec per prediction. These results show that the feed-forward deep neural network is the best choice to be employed in an intrusion detection system.