Skip to content
<- all work
AI / MLSpring 2026·CSDP 710, Machine Learning. Phase 1 of 3

Insider Threat Detection from Windows File-Share Events

Unsupervised ML on 4656/5145 telemetry in a live Zero Trust lab

Windows domains emit file-access telemetry constantly. Events 4656 and 5145 record who touched what, from where, and whether access was granted, and almost nobody looks at it. This work builds an unsupervised detection pipeline on that stream inside a controlled Active Directory lab, folding raw events into behavioral sessions and scoring them with Isolation Forest, LOF, and One-Class SVM.

Results

18,572
Events collected
over 23 days, no simulated attacks
58
Sessions
across 6 domain users
12
Behavioral features
engineered per session
3
Models compared
IF, LOF, One-Class SVM

01The gap

Existing work detects insider threats from logon events, cloud telemetry, or synthetic datasets. None of it runs on Windows file-access events. None compares Isolation Forest, LOF, and One-Class SVM on one shared feature set. None applies SHAP at the file layer.

Zero Trust asks for continuous verification, but that only means something if the behavioral signals are actually being modeled. Right now, at the file layer, they are not.

02Pipeline

A collector on the domain controller subscribes to Events 4656 and 5145, writes each one to JSON, and appends to a rolling log. Events are grouped by subject_logon_id, the per-login session identifier, then aggregated into twelve behavioral features. One of them, co_event_flag, records whether 4656 and 5145 appear together inside a session. A logon-only study has no way to derive it.

All three models train on the full baseline at a 3% contamination rate, treating every session as roughly normal since no attacks were staged. Isolation Forest carries the primary research question. LOF targets gradual behavioral drift. One-Class SVM handles hard permission violations.

03Status

Phase 1 built and validated the pipeline on baseline data: 23 days, six users, two source hosts, four shared drives, 12,954 successful and 5,618 failed access attempts. Phase 2 adds simulated attack sessions for labeled per-category evaluation. Phase 3 puts the model on a live sliding window and measures false-positive rate and latency.