期刊名称：TELKOMNIKA (Telecommunication Computing Electronics and Control)
印刷版ISSN：2302-9293
出版年度：2016
卷号：14
期号：3A
页码：321-325
DOI：10.12928/telkomnika.v14i3A.4387
语种：English
出版社：Universitas Ahmad Dahlan
摘要：There are increasing demands for accessing information over the Internet, more and more networks are designed and deployed. The information and network security becomes a key issue for us to study. Neural network is effective to detect network intrusion. Much effort has been taken in this field. Stolfo et al put forward 41 higher-level derived features to distinguish normal connections from abnormal connections. Unfortunately, with these 41 derived features as inputs, IDS systems take long time to converge when training and work slowly during on-line detections. We quantize derived features to digital type before feeding them to IDS systems. We reduce the number of inputs while keeping IDS systems high detection rates. After a long time of hard work, we achieved a good architecture, i.e. 18-35-1, of BP neural network for IDS systems. And we choose trainbfg as training function.
其他摘要：There are increasing demands for accessing information over the Internet, more and more networks are designed and deployed. The information and network security becomes a key issue for us to study. Neural network is effective to detect network intrusion. Much effort has been taken in this field. Stolfo et al put forward 41 higher-level derived features to distinguish normal connections from abnormal connections. Unfortunately, with these 41 derived features as inputs, IDS systems take long time to converge when training and work slowly during on-line detections. We quantize derived features to digital type before feeding them to IDS systems. We reduce the number of inputs while keeping IDS systems high detection rates. After a long time of hard work, we achieved a good architecture, i.e. 18-35-1, of BP neural network for IDS systems. And we choose trainbfg as training function.
关键词：Intrusion detection; neural network; derived features
