A fuzzy feature evaluation framework for network intrusion detection

The design of a Network Intrusion Detection System (NIDS) is a delicate process which requires the successful completion of numerous design stages. The feature selection stage is one of the first steps that needs to be addressed, and can be considered among the top most important ones. If this step is not carefully considered the overall performance of the NIDS will greatly suffer, regardless of the detection technique, or any other algorithms that the NIDS is using. The most common approach for selecting the network features is to use expert knowledge to reason about the selection process. However, this approach is not deterministic, thus, in most cases researchers end-up with completely different sets of important features for the detection process. Furthermore, the lack of a generally accepted feature classification schema forces different researchers to use different names for the same (subsets of) features, or the same name for completely different ones. It is our belief that these issues are not sufficiently studied and explored by the network security research community. This thesis focuses on mining the most useful network features for attack detection. Accordingly, we propose a new network feature classification schema as well as a mathematical feature evaluation procedure that helps us identify the most useful features that can be extracted from network packets. The network feature classification schema is intended to provide a better understanding, and enforce a new standard, upon the features that can be extracted from network packets, and their relationships. The classification has a set of 27 feature categories based on the network abstractions that they refer to (e.g., host, network, connection, etc). We use our feature classification schema to select a comprehensive set of 671 features for conducting and reporting our experimental findings. The feature evaluation procedure provides a deterministic approach for pinpointing those network features that are indeed useful in the attack detection process. The procedure uses mathematical, statistical and fuzzy logic techniques to rank the participation of individual features into the detection process. In particular, we propose a new feature dependency measure for independent evaluation criteria that is, to our knowledge, a pioneer method designed for intrusion detection. In our research we have identified several tuning parameters that directly influence the detection performance of each individual feature. To address this issue, our method takes into account the performance of each feature while using multiple tunings, making the evaluation process more robust to biases that could be accidentally introduced by a poor tuning combination. The experimental results, conducted on three different real-world network datasets, empirically confirm that our feature evaluation model can successfully be applied to mine the importance of a feature in the detection process.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.