An efficient hidden Markov model training scheme for anomaly intrusion detection of server applications based on system calls

Recently hidden Markov model (HMM) has been proved to be a good tool to model normal behaviours of privileged processes for anomaly intrusion detection based on system calls. However, one major problem with this approach is that it demands excessive computing resources in the HMM training process, which makes it inefficient for practical intrusion detection systems. In this paper a simple and efficient HMM training scheme is proposed by the innovative integration of multiple-observations training and incremental HMM training. The proposed scheme first divides the long observation sequence into multiple subsets of sequences. Next each subset of data is used to infer one sub-model, and then this sub-model is incrementally merged into the final HMM model. Our experimental results show that our HMM training scheme can reduce the training time by about 60% compared to that of the conventional batch training. The results also show that our HMM-based detection model is able to detect all denial-of-service attacks embedded in testing traces.