Скачать или смотреть Human-Computer Decision Systems for Cybersecurity

Security decision systems aim to distinguish malicious activity from benign and often use a combination of human expert and
automated analysis, including machine learning (ML). Systems using only human experts scale badly; pure ML systems are
susceptible to structured attack by adversaries and, in most cases, have unsatisfactory performance on their own. [Jones 2014]
In order for machine learning to reach its potential as a tool for cyber-defense, we must improve collaboration between experts
and automation. The well-established subfield of Active Learning studies the ability of an ML algorithm to learn from ongoing
human analyst feedback, but it provides only an incomplete view of the system. We must also account for the influence that the
ML algorithm has on the human (e.g., through experience gained, boredom induced, or other factors). Furthermore, the Active
Learning literature usually assumes that an analyst is a single, flawless, oracle rather than a heterogeneous and fallible team of
individuals.