Peer-Group-Aware Anomaly Detection for Credential Misuse in Enterprise Authentication Logs
Online - Zoom
Dr. Ashok Kumar Patel
apatel38@umassd.edu
https://umassd.zoom.us/j/99756832558?pwd=5Jj2KWIfqIAz7H1P7QiC8E9LgRRWMm.1
College of Engineering
Computer & Information Science
Location: https://umassd.zoom.us/j/99756832558?pwd=5Jj2KWIfqIAz7H1P7QiC8E9LgRRWMm.1
Thesis Advisor: Dr. Ashokkumar Patel, CIS/DSC - University of Massachusetts, Dartmouth
Committee Members:
- Dr. Gokhan Kul, CIS/DSC - University of Massachusetts, Dartmouth
- Dr. Yuchou Chang, CIS/DSC - University of Massachusetts, Dartmouth
Abstract
Credential misuse is hard to detect because every login in an attack can succeed. This thesis asks whether judging each login against accounts with similar roles, its peer group, reduces false alarms at the same detection rate. A peer-aware detector is proposed: accounts are clustered on behavior and host usage, each group is modeled by its own isolation forest, and scores become peer-relative tail probabilities. It is compared with a first-time-host rule and with per-user, network-aware and peer-column models on the Los Alamos National Laboratory authentication data, using three chronological partitions, causal features, a test set locked until all choices are frozen, and automated leakage checks. On a test period of 44.2 million logins with 333 red-team logins, peer features used as extra inputs raised average precision 1.7-fold over the network-aware model on every seed, and the peer-aware detector needed 13% fewer alerts at exactly 80% recall, a difference not significant across seeds. Neither advantage held at 90% recall, and calibration did not improve ranking. Thresholds set to catch 80% of validation attacks caught only 59–69% of test attacks, and grouping alerts per account, destination and day cut workload about fivefold. Peer context helps in specific places, but threshold transfer between attack periods matters more for deployment than the differences between detectors.
For further information please contact Dr. Ashokkumar Patel. All CIS MS students are encouraged to attend.