Existing approaches that leverage system log data for anomaly detection can be broadly classified into three groups: PCA based approaches over log message counters [39], invariant mining based methods to capture co-occurrence patterns between different log keys [21], and workflow based methods to identify execution anomalies in program logic flows [42]. Even though they are successful in certain scenarios, none of them is effective as a universal anomaly detection method that is able to guard against different attacks in an online fashion.