Setting the threshold for high throughput detectors: A mathematical approach for ensembles of dynamic, heterogeneous, probabilistic anomaly detectors

TL;DR

This paper introduces a mathematical approach and an algorithm for setting thresholds in ensembles of heterogeneous, adaptive anomaly detectors to control alert rates effectively in high-volume cyber security environments.

Contribution

It provides a rigorous, a priori threshold-setting algorithm for dynamic, heterogeneous detectors and analyzes the impact of data distribution knowledge on its effectiveness.

Findings

Algorithm effectively regulates alert rates in large detector ensembles.

Empirical validation on real network data demonstrates practical utility.

Analysis reveals when model refitting is necessary due to distribution shifts.

Abstract

Anomaly detection (AD) has garnered ample attention in security research, as such algorithms complement existing signature-based methods but promise detection of never-before-seen attacks. Cyber operations manage a high volume of heterogeneous log data; hence, AD in such operations involves multiple (e.g., per IP, per data type) ensembles of detectors modeling heterogeneous characteristics (e.g., rate, size, type) often with adaptive online models producing alerts in near real time. Because of high data volume, setting the threshold for each detector in such a system is an essential yet underdeveloped configuration issue that, if slightly mistuned, can leave the system useless, either producing a myriad of alerts and flooding downstream systems, or giving none. In this work, we build on the foundations of Ferragut et al. to provide a set of rigorous results for understanding the relationship between threshold values and alert quantities, and we propose an algorithm for setting the threshold in practice. Specifically, we give an algorithm for setting the threshold of multiple, heterogeneous, possibly dynamic detectors completely a priori, in principle. Indeed, if the underlying distribution of the incoming data is known (closely estimated), the algorithm provides provably manageable thresholds. If the distribution is unknown (e.g., has changed over time) our analysis reveals how the model distribution differs from the actual distribution, indicating a period of model refitting is necessary. We provide empirical experiments showing the efficacy of the capability by regulating the alert rate of a system with $\approx$2,500 adaptive detectors scoring over 1.5M events in 5 hours. Further, we demonstrate on the real network data and detection framework of Harshaw et al. the alternative case, showing how the inability to regulate alerts indicates the detection model is a bad fit to the data.