Robust Support Vector Machines for Anomaly Detection in Computer Security.

Using the 1998 DARPA BSM data set collected at MIT’s Lincoln Labs to study intrusion detection systems, the performance of robust support vector machines (RVSMs) was compared with that of conventional support vector machines and nearest neighbor classifiers in separating normal usage profiles from intrusive profiles of computer programs. The results indicate the superiority of RSVMs not only in terms of high intrusion detection accuracy and low false positives but also in terms of their generalization ability in the presence of noise and running time. Keywords—Intrusion detection, computer security, robust support vector machines, noisy data.