Date of Award
Master of Science
Electrical and Computer Engineering
Intrusion detection systems have been widely used as "burglar alarms" in the computer security field. There are two major types of detection techniques: misuse detection and anomaly detection. Although misuse detection can detect known attacks with lower false positive rate, anomaly detection is capable of detecting any new or varied attempted intrusion as long as the attempted intrusions disturb the normal states of the systems. The network anomaly detector is employed to monitor a segment of network for any suspicious activities based on the sniffered network traffic. The fast speed of network and wide use of encryption techniques make it almost unpractical to read payload information for the network anomaly detector. This work tries to answer the question: What are the best features for network anomaly detector? The main experiment data sets are from 1999 DARPA Lincoln Library off-line intrusion evaluation project since it is still the most comprehensive public benchmark data up to today. Firstly, 43 features of different levels and protocols are defined. Using the first three weeks as training data and last two weeks as testing data, the performance of the features are testified by using 5 different classifiers. Secondly, the feasibility of feature selection is investigated by employing some filter and wrapper techniques such as Correlation Feature Selection, etc. Thirdly, the effect of changing overlap and time window for the network anomaly detector is investigated. At last, GGobi and Mineset are utilized to visualize intrusion detections to save time and effort for system administrators. The results show the capability of our features is not limited to probing attacks and denial of service attacks. They can also detect remote to local attacks and backdoors. The feature selection techniques successfully reduce the dimensionality of the features from 43 to 10 without performance degrading. The three dimensional visualization pictures provide a straightforward view of normal network traffic and malicious attacks. The time plot of key features can be used to aid system administrators to quickly locate the possible intrusions.
Xin, Jianqiang, "Feature selection and visualization techniques for network anomaly detector" (2003). Retrospective Theses and Dissertations. 20092.