Threshold Selection and Attack Type Generalization in XGBoost and Denoising Autoencoder Intrusion Detection

Hybrid intrusion detectors must divide a false-positive budget between supervised and anomaly scores. We study allocation stability, changed alerts, and attack-type transfer using extreme gradient boosting (XGBoost) and a benign-trained denoising autoencoder (DAE). Experiments include cross-day evaluation, fixed-model resampling, type and grouped-family holdouts, matched baselines, and corrected CICIDS2017 and UNSW-NB15 replications. At the 1% calibration target, exploratory Friday recall is 67.46% for DAE alone versus 47.47% for selected fusion, with observed false-positive rates of 7.29% and 5.69%. With re-estimated thresholds, original CICIDS2017 allocations change in 25.0–54.0% of development resamples at the 0.1% target and 65.0–78.5% at 1%; mean retrospective Friday prediction disagreement at 1% is only 0.81–1.04%. Selected fusion gains 5.86 percentage points of macro recall over XGBoost in six individual-type holdouts at 1%. With all four denial-of-service (DoS) types excluded together, however, 1% within-family recall is 19.26% for selected fusion, 14.25% for XGBoost, and 54.78% for DAE; both-Patator exclusion gives 5.46% for selected fusion versus 7.66% for XGBoost. Ordinary-autoencoder fusion performs comparably to DAE fusion in the individual-type protocol, and longer denoising training does not consistently improve detection. On corrected CICIDS2017, selected fusion reaches 66.18% macro recall versus 70.71% for XGBoost. UNSW-NB15 holdouts disable the anomaly branch throughout, with a 4.38% mean test false-positive rate at the 1% target; a separate all-known audit finds rare switching. The contribution separates allocation variability, changed decisions, and budget transfer. Neither stable selection nor gains on individual attack types establish reliable unseen-family detection.

Authors

Institutions

Publication Details

Journal
Information
Published
2026-11-16
DOI
https://doi.org/10.3390/info17111155
Primary Topic
Network Security and Intrusion Detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Threshold Selection and Attack Type Generalization in XGBoost and Denoising Autoencoder Intrusion Detection

Young Im Cho, Zhansaya Abildaeva, Galymzhan Kazhymurat, Tulegenova Elmira et al.
1 links
Information
Network Security and Intrusion Detection
article

Threshold Selection and Attack Type Generalization in XGBoost and Denoising Autoencoder Intrusion Detection

Young Im Cho, Zhansaya Abildaeva, Galymzhan Kazhymurat, Tulegenova Elmira, Nazerke Karim, Akmaral Almenova, Aigul Dauitbayeva, Nurbek Konyrbayev
article en

Abstract

Hybrid intrusion detectors must divide a false-positive budget between supervised and anomaly scores. We study allocation stability, changed alerts, and attack-type transfer using extreme gradient boosting (XGBoost) and a benign-trained denoising autoencoder (DAE). Experiments include cross-day evaluation, fixed-model resampling, type and grouped-family holdouts, matched baselines, and corrected CICIDS2017 and UNSW-NB15 replications. At the 1% calibration target, exploratory Friday recall is 67.46% for DAE alone versus 47.47% for selected fusion, with observed false-positive rates of 7.29% and 5.69%. With re-estimated thresholds, original CICIDS2017 allocations change in 25.0–54.0% of development resamples at the 0.1% target and 65.0–78.5% at 1%; mean retrospective Friday prediction disagreement at 1% is only 0.81–1.04%. Selected fusion gains 5.86 percentage points of macro recall over XGBoost in six individual-type holdouts at 1%. With all four denial-of-service (DoS) types excluded together, however, 1% within-family recall is 19.26% for selected fusion, 14.25% for XGBoost, and 54.78% for DAE; both-Patator exclusion gives 5.46% for selected fusion versus 7.66% for XGBoost. Ordinary-autoencoder fusion performs comparably to DAE fusion in the individual-type protocol, and longer denoising training does not consistently improve detection. On corrected CICIDS2017, selected fusion reaches 66.18% macro recall versus 70.71% for XGBoost. UNSW-NB15 holdouts disable the anomaly branch throughout, with a 4.38% mean test false-positive rate at the 1% target; a separate all-known audit finds rare switching. The contribution separates allocation variability, changed decisions, and budget transfer. Neither stable selection nor gains on individual attack types establish reliable unseen-family detection.

InformationVol. 17(11)
Gachon University (KR), Korkyt Ata Kyzylorda State University (KZ)
Network Security and Intrusion 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