Multinomial classification of web attacks using improved fuzzy rules learning by Neuro-Fuzzy

Neural Networks are used together with fuzzy inference systems in Neuro-Fuzzy, a prominent synergy of rules parameters unsupervised discovery and supervised tuning of classification model. The binary classification task in Network Forensics applications are the most widely used and applied for detection ``benign'' and ``malicious'' activities. However, in many areas it is not enough to distinguish between those two classes, yet also important to provide a more specific determination of what exactly ``malicious'' sub-class some action belongs to. Despite the inherited properties and limitations of Neural Networks, the Neuro-Fuzzy may be tuned to handle non-linear data in multinomial classification problems, which is not a simple addition to a binary classification model. This work targets the optimization of the Neuro-Fuzzy output layer construction and rules tuning in multinomial classification problems as well as solving accompanying challenges. Moreover, we performed extensive study of ML methods designed for binary- and multinomial classification problems. We believe that our approach will help to derive more accurate fuzzy rules multinomial model to be used for web attacks identification.