Detection and discrimination of injected network faults

Six hundred faults were induced by injection into five live campus networks at Carnegie Mellon University in order to determine whether or not particular network faults have unique signatures as determined by out-of-band monitoring instrumentation. If unique signatures span networks, then the monitoring instrumentation can be used to diagnose network faults, or distinguish among fault classes, without human intervention, using machine-generated diagnostic decision rules. This would be especially useful in large, unmanned systems in which the occurrence of novel or unanticipated faults can be catastrophic. Results indicate that significant accuracy in automated detection and discrimination among fault types can be obtained using anomaly signatures as described.