Intrusion Detection Based on Principal Component Analysis and Cascaded Extreme Learning Machine
| Autor: | ZHEN-YU DU, 杜鎮宇 |
|---|---|
| Rok vydání: | 2019 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 107 Along with the prosperous development of information network, information security has undoubtedly become a hot topic. The Intrusion Detection System (IDS) has played an important role for accurate and timely detection of network attacks from hostile users. In recent years, machine-learning technology has been widely used in IDS. However, a large amount of training samples will require too much training time. The imbalanced amounts of various types of training samples will result in difficulty in identifying the type of intrusion with extremely small amount, from which misjudgments will occur, i.e. the system fails to correctly identify the occurrences of abnormal flow. To solve the two aforementioned problems, we propose a cascade extreme learning machine based on principal component analysis by improving the original Cascaded ELM (CE). First step is the undersampling of training data set to balance the amounts of various types of training samples, using principal component analysis to extract key features from data set, removing redundant features for data dimension reduction, and training binary classifier with respect to every type of intrusion and normal type. Then all classifiers will be connected via the cascade approach such that the test data can be correctly classified layer by layer via our classifiers, and eventually abnormal flow can be accurately judged. As compared to other methods, the method proposed in this paper can provide better effectiveness and identification capability with respect to imbalanced data sets under a huge amount of training samples. We use four different data sets of Ada, Sylva, Gina, and Madelon to experimentally prove the PCE method of this paper is equipped with better classification capability than CE. In the end the PCE method proposed in this paper is compared with other methods during the KDDCUP99 data set experiment, and it shows slight improvement and balanced effectiveness. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|