Problem
About 60 ML models flag anomalous events in a SIEM that receives about 20 thousand events per second, almost 2 billion per day. Even at 0.999 quality, atomic flags give almost 2 million false positives a day, which no team can review. A rule-based layer over model triggers scores a process only with its two nearest ancestors, so a low-scoring process inside an attack gets lost.
Example from the article: in a real attack tree most processes scored up to 140 points, but conhost.exe, part of the same attack, scored about 40 and would be ignored by an analyst.
Solution in brief
Cut the process launch tree at a system process and give each node the vector of model triggers. A graph convolution (mean over all nodes within a chosen depth) adds the neighborhood context to every node, and an SVM decides whether the node belongs to an attack tree. In the production check the false-positive rate fell from about 4% to about 0.001%.
Questions this case answers
- How to turn many atomic detector triggers into one risk score for a tree of related events, cutting false alerts without missing low-scoring parts of a rare pattern
- How to score a node in a hierarchy of events from its neighbours within a fixed depth, instead of from its immediate ancestors only
- How to cut false positives of an anomaly detection system that scores events one by one
- How to turn trigger flags of many detection models into a risk score for a tree of related events
Tools and methods
- Recommender "user - process" — collaborative filtering; anomaly when launched processes differ from recommended ones.
- Suspicious process chains — transition matrix between each process and the final one of a chain.
- Rule-based profiler — risk score from rules over model triggers plus parent scores.
- Graph convolution — mean over neighbors up to depth gives context features.
- SVM, CatBoost, Random Forest, Decision Tree — classifiers over the aggregated vectors.
Google or a one-time email code. No password.
