Two sophisticated techniques to improve HMM-based intrusion detection systems

Hidden Markov model (HMM) has been successfully applied to anomlay detection as a technique to model normal behavior. Despite its good performance, there are some problems in applying it to real intrusion detection systems: it requires large amount of time to model normal behaviors and the false-positive error rate is relatively high. To remedy these problems, we have proposed two techniques: extracting privilege flows to reduce the normal behaviors and combining multiple models to reduce the false-positive error rate. Experimental results with real audit data show that the proposed method requires significantly shorter time to train HMM without loss of detection rate and significantly reduces the false-positive error rate.