A Multi-Dimension EML Feature Framework and Multi-Model-Based Email Anomaly Detection System

Detecting business email compromise (BEC), account takeover (ATO), and insider threats that use legitimate accounts and business language requires analyzing the sender’s role and prior conversations alongside a message’s technical attributes. However, because data combining corporate email with human resources information is subject to restrictions regarding access to personal and proprietary business information, it is difficult to study organizational roles and communication relationships simultaneously. We propose a multi-dimensional UBA feature framework that integrates raw EML files, MTA logs, and role and authority information from ERP, together with a synthetic email corpus of a fictional company, NexaSoft. Based on an organization of 126 employees and 117 normal business scenarios, we define 23 feature-based and 11 contextual anomaly types and build a dataset of 124,355 emails (76,850 normal, 47,505 anomalous), organized into feature-based and context-based datasets. Linking each message to organizational roles and reply relationships allows normal work and anomalous requests to be represented in the same organizational and conversational environment. The detection method comprises a rule layer for explicit security conditions, an XGBoost layer for structural and statistical anomalies, and a multi-view layer for context-dependent anomalies, in which models M1–M4 analyze message semantics, behavior and time, organizational role and policy, and past conversation. On their assigned classes, M1, M2, and M3 achieve Macro-F1 of 0.9863, 0.9752, and 0.9110 under scenario-separated evaluation, and M4 achieves a mean Macro-F1 of 0.9213 over three seeds. Removing prior conversation lowers M4’s Macro-F1 by 4.34 pp, confirming its contribution. The dataset and detection framework thus enable the study and evaluation of both technical and context-dependent email anomalies, grounded in organizational roles and conversation flow, in one business environment.

Authors

Institutions

Publication Details

Journal
Applied Sciences
Published
2026-11-12
DOI
https://doi.org/10.3390/app162211224
Primary Topic
Spam and Phishing Detection
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Multi-Dimension EML Feature Framework and Multi-Model-Based Email Anomaly Detection System

Kyeongseok Lee, Sangryu Park, Jinhyuk Son, Gyuhyeon Choi et al.
Applied Sciences
Spam and Phishing Detection
article

A Multi-Dimension EML Feature Framework and Multi-Model-Based Email Anomaly Detection System

Kyeongseok Lee, Sangryu Park, Jinhyuk Son, Gyuhyeon Choi, Jin Jeong, Jaewan Hong, Homook Cho
article en

Abstract

Detecting business email compromise (BEC), account takeover (ATO), and insider threats that use legitimate accounts and business language requires analyzing the sender’s role and prior conversations alongside a message’s technical attributes. However, because data combining corporate email with human resources information is subject to restrictions regarding access to personal and proprietary business information, it is difficult to study organizational roles and communication relationships simultaneously. We propose a multi-dimensional UBA feature framework that integrates raw EML files, MTA logs, and role and authority information from ERP, together with a synthetic email corpus of a fictional company, NexaSoft. Based on an organization of 126 employees and 117 normal business scenarios, we define 23 feature-based and 11 contextual anomaly types and build a dataset of 124,355 emails (76,850 normal, 47,505 anomalous), organized into feature-based and context-based datasets. Linking each message to organizational roles and reply relationships allows normal work and anomalous requests to be represented in the same organizational and conversational environment. The detection method comprises a rule layer for explicit security conditions, an XGBoost layer for structural and statistical anomalies, and a multi-view layer for context-dependent anomalies, in which models M1–M4 analyze message semantics, behavior and time, organizational role and policy, and past conversation. On their assigned classes, M1, M2, and M3 achieve Macro-F1 of 0.9863, 0.9752, and 0.9110 under scenario-separated evaluation, and M4 achieves a mean Macro-F1 of 0.9213 over three seeds. Removing prior conversation lowers M4’s Macro-F1 by 4.34 pp, confirming its contribution. The dataset and detection framework thus enable the study and evaluation of both technical and context-dependent email anomalies, grounded in organizational roles and conversation flow, in one business environment.

Applied SciencesVol. 16(22)
Gachon University (KR), Korea Advanced Institute of Science and Technology (KR)
Ministry of Science and ICT, South Korea
Spam and Phishing Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon