Abstract:
The intrusion detection problem could be modeled as data stream classification problem.For traditional data stream classification algorithms,a lot of labeled training samples should be provided to train the classifier,with a high expense,and decrease the utility of the algorithms.For positive unlabeled learning (PU Learning) based algorithms,only positive and unlabeled samples are required to train classifier.In this paper,we propose to use dynamic classifier ensemble of PU learning algorithm to cope with intrusion detection,which only requires label some positive samples for building data stream classifier. The experiment result on both synthetic and real-life dataset shows that the proposed algorithm has excellent classification performance,and it outperforms 3 PU learning algorithms on data streams with shewed distribution,with high intrusion detection ratio.