Alert Analysis and Contextualization in Provenance-Based Intrusion Detection
Abstract
Provenance-based intrusion detection systems (PIDS) often rely on anomaly detection to identify Advanced Persistent Threat (APT) activity. However, their alerts are commonly evaluated only by overlap with documented attack ground truth. As a result, all non-ground-truth alerts are treated as false positives, although some may represent unusual but unlabeled behavior. We distinguish two complementary interpretations of PIDS alerts: attack-sense, which measures overlap with documented attacks, and anomaly-sense, which assesses whether alerts are supported by deviations from learned benign behavior. To instantiate this idea, we introduce a post-processing method for executable-behavior alerts that compares alerted process neighborhoods with benign reference neighborhoods using a behavior-distance metric. Evaluated on DARPA TC, OpTC, ATLASv2, and CARBANAKv2, the method separates alerts into untrained behavior, context-supported deviations, and alerts consistent with benign references. Our results show that contextualization enables an evaluation perspective that is better aligned with the anomaly-detection task than true-positive and false-positive counts alone. The method does not prove maliciousness, but provides behavioral evidence to support analyst triage.