Detecting the Abnormal: Machine Learning in Computer Security

Feb 14, 1997

Abstract

Two problems of importance in computer security are to:

detect the presence of an intruder masquerading as the valid user
detect the perpetration of abusive actions on the part of an otherwise innocuous user.
In this talk I present a machine learning approach to anomaly detection, designed to handle these two problems. Our system learns a user profile for each user account and subsequently employs it to detect anomalous behavior in that account. Based on sequences of actions (UNIX commands) of the current user's input stream, the system compares each fixed-length input sequence with a historical library of the account's command sequences using a similarity measure. The system must learn to classify current behavior as consistent or anomalous with past behavior using only positive examples of the account's valid user. Our empirical results demonstrate that in most cases it is possible to distinguish the legitimate user from an intruder and, furthermore, that an instance selection technique based on a memory page-replacement algorithm is capable of drastically reducing library size without hindering detection accuracy.