Security event anomaly detection
What it does
Anomaly detection watches how many security events of each kind arrive for each customer, hour by hour, and alerts you when an hour is far outside that customer's own recent pattern. It catches a sudden burst in one kind of collected security event, such as syslog or SNMP traps, and a sudden drop in one kind, which can mean a source stopped sending. It catches changes that a fixed threshold would never notice.
Anomaly detection is part of the native SIEM preview. On the Alert Rules page, the Native SIEM tab reads "Off unless the native SIEM preview is switched on." Ask support whether the preview is on for your workspace before you rely on these alerts.
How it works
- Each customer and each kind of collected security event (from the sources your workspace collects, such as network device logs, SNMP traps, device control events and email security feeds) has its own baseline. The baseline is the customer's recent hourly counts for that kind: up to the last 48 closed hours that were not themselves judged a spike or drop.
- A new hour is compared with the middle of that history and its normal spread. The comparison is robust: one unusual hour in the past does not distort what counts as normal.
- An hour is reported as a spike when it is far above normal, and as a drop when it is far below normal. Small changes are ignored: an hour must also differ from normal by at least 10 events.
- Only closed hours are judged, so the current, partly filled hour never looks like a drop.
- Hours that were reported as a spike or drop are kept out of the baseline, so a long storm cannot teach the detector that the storm is normal.
- One customer's counts never affect another customer's baseline.
States
- Learning: fewer than 12 hours of history for that customer and event kind. Nothing is reported while learning.
- Normal: the hour is within the customer's usual range.
- Spike or drop: the hour is far outside the usual range. An alert opens.
Step by step: check or tune the alert
- Go to Administration > General > Alert Rules. This page needs an Owner or Admin role.
- On the Native SIEM tab, find the alert kind Security event volume anomaly. When the preview is on, it is on by default with Warning severity, and it does not open a case or send a notification until you choose to.
- Set the severity, and choose whether it opens a case, sends a notification or creates a ticket. Switch it off if you do not want these alerts.
- Make sure the customers you care about are sending security events.
Day to day
- An anomaly alert names the event kind, the count in that hour, the hour itself and the customer's usual level, so you can judge the size of the change at a glance.
- Repeated anomalies for the same customer and event kind fold into one open alert instead of creating a new alert each hour. Each hour is delivered at most once, even after a restart.
- Treat a drop as seriously as a spike: a source that goes quiet can mean a device stopped logging or logging was turned off.
Limits
- Detection works on hourly counts of collected security events per customer and event kind, not per source. It does not judge the content of single events; correlation rules do that. One quiet source shows only when the whole kind's hourly count drops far enough.
- Learning counts hours, including hours with no events. A customer is replayed over the last 48 hours when it is first seen, so its first busy hour, or the first busy hour of a new event kind, can be reported as a spike.
- A customer with no security events in the last 48 hours is not checked.
