Intelligent web application attack detection system

Nowadays, it is becoming increasingly difficult to ensure the security of the web applications we use in every aspect of life, whether in the infrastructure or directly. As more valuable data becomes available to people through the Internet and the network, attacks are advancing and diversifies comparatively, and therefore they become more difficult to deal with. While signature-based intrusion detection systems are quite successful in preventing known attacks and their close derivatives, they are incapable of preventing zero-day attacks that are discovered for the first time, and periodically data leaks from the systems or they become unserviceable. Artificial intelligence studies have also accelerated in recent years in order to detect web attacks with higher accuracy. The biggest shortcoming of artificial intelligence studies is the lack of a carefully created, numerically rich or balanced dataset where each class is represented with sufficient data for multi-label classification. In this study, fort the purpose of detect web attacks, a balanced dataset containing multiple attack data types, which is larger in number than other datasets, was created in order to perform multi-label classification. The quality of the dataset was increased by running the newly developed pre-processing phase on the dataset several times, and the same pre-processing phase was used to increase the classification capacity. The purpose of multi-label classification is to provide an environment for cyber security researchers to see attack types before they happen and to carry out preventive work without being damaged. The developed dataset and pre-processing technique are the direct contribution of this study to the literature. The created dataset contains 274,242 lines of attack data, 24,962 lines of normal web requests and a total of 299,205 lines of data. For binary classification, normal web request entries in the dataset were increased to 170,547 lines. The highest hit results of 99.872% in binary classification and 94.801% in multiple classification were obtained from three deep learning and 14 machine learning classifiers for which hyperparameter optimizations were also performed.