Combining heterogeneous anomaly detectors for improved software security

Abstract

Abstract Host-based Anomaly Detection Systems (ADSs) monitor for significant deviations from normal software behavior. Several techniques have been investigated for detecting anomalies in system call sequences. Among these, Sequence Time-Delay Embedding (STIDE), Hidden Markov Model (HMM), and One-Class Support Vector Machine (OCSVM) have shown a high level of anomaly detection accuracy. Although ADSs can detect novel attacks, they generate a large number of false alarms due to the difficulty in obtaining complete descriptions of normal software behavior. This paper presents a multiple-detector ADS that efficiently combines the decisions from heterogeneous detectors (e.g., STIDE, HMM, and OCSVM), using Boolean combination in the Receiver Operating Characteristics (ROC) space, to reduce the false alarms. Results on two modern and large system call datasets generated from Linux and Windows operating systems show that the proposed ADS consistently outperforms an ADS based on a single best detector and on an ensemble of homogeneous detectors. At an operating point of zero percent alarm rate, the proposed multiple-detector ADS increased the true positive rate by 500% on the Linux dataset and by 25% on the Window dataset. Furthermore, the combinations of decisions from multiple heterogeneous detectors make the ADS more reliable and resilient against evasion and adversarial attacks.