Split-parameter federated learning for threat detection across heterogeneous security infrastructures

Federated learning enables collaborative intrusion detection without centralizing sensitive network data, but existing approaches require all participants to share a homogeneous feature space—a fundamental limitation for organizations operating heterogeneous IT, OT, and IoT infrastructures that produce structurally different telemetry. We propose a split-parameter federated architecture that maintains domain-specific encoders (trained locally, never shared) and aggregates only a shared classification head across domains, enabling cross-domain threat detection without feature alignment or projection. We evaluate this architecture on three domains spanning the IT-OT-IoT spectrum: enterprise network traffic (UNSW-NB15, 42 features), industrial control systems (HAI, 61 features), and IoT botnet detection (N-BaIoT, 115 features). Over 20 federation rounds, the architecture achieves a global weighted F1 of 0.9738 (mean across 5 seeds, 95% CI [0.9736, 0.9739]) with 93.5% detection rate and 3.1% false alarm rate, exceeding all single-domain binary configurations in global F1 (best single-domain: 0.954). Per-domain macro F1 is 0.846 (with HAI achieving 80.0% DR), reflecting the performance gap that sample-weighted global metrics conceal. The approach reduces per-round communication by 89.1% relative to a hypothetical full-model reference (4,418 shared parameters vs 40,646 total), and by 49.3% relative to the applicable feature-projection baseline (4,418 vs 8,706 shared parameters at an identical total parameter budget). All primary results are reported under the five-seed protocol rather than single-seed exploratory runs, prioritizing reproducibility and confidence interval reporting. We analyze a persistent oscillation pattern in the ICS domain (HAI stable in 9/20 rounds, collapsed in 11/20 rounds) and correlated volatility in the IT domain, attributable to gradient conflicts in the shared head from semantically distant feature spaces, and evaluate domain-specific gradient scaling and lower learning rate as mitigation strategies.

Authors

Institutions

Publication Details

Journal
Discover Internet of Things
Published
2026-11-15
DOI
https://doi.org/10.1007/s43926-026-00520-6
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

Split-parameter federated learning for threat detection across heterogeneous security infrastructures

Vivek Bhardwaj, Sunil Kumar Yadav, Vivek Kumar Verma
Discover Internet of Things
Network Security and Intrusion Detection
article

Split-parameter federated learning for threat detection across heterogeneous security infrastructures

Vivek Bhardwaj, Sunil Kumar Yadav, Vivek Kumar Verma
article en

Abstract

Federated learning enables collaborative intrusion detection without centralizing sensitive network data, but existing approaches require all participants to share a homogeneous feature space—a fundamental limitation for organizations operating heterogeneous IT, OT, and IoT infrastructures that produce structurally different telemetry. We propose a split-parameter federated architecture that maintains domain-specific encoders (trained locally, never shared) and aggregates only a shared classification head across domains, enabling cross-domain threat detection without feature alignment or projection. We evaluate this architecture on three domains spanning the IT-OT-IoT spectrum: enterprise network traffic (UNSW-NB15, 42 features), industrial control systems (HAI, 61 features), and IoT botnet detection (N-BaIoT, 115 features). Over 20 federation rounds, the architecture achieves a global weighted F1 of 0.9738 (mean across 5 seeds, 95% CI [0.9736, 0.9739]) with 93.5% detection rate and 3.1% false alarm rate, exceeding all single-domain binary configurations in global F1 (best single-domain: 0.954). Per-domain macro F1 is 0.846 (with HAI achieving 80.0% DR), reflecting the performance gap that sample-weighted global metrics conceal. The approach reduces per-round communication by 89.1% relative to a hypothetical full-model reference (4,418 shared parameters vs 40,646 total), and by 49.3% relative to the applicable feature-projection baseline (4,418 vs 8,706 shared parameters at an identical total parameter budget). All primary results are reported under the five-seed protocol rather than single-seed exploratory runs, prioritizing reproducibility and confidence interval reporting. We analyze a persistent oscillation pattern in the ICS domain (HAI stable in 9/20 rounds, collapsed in 11/20 rounds) and correlated volatility in the IT domain, attributable to gradient conflicts in the shared head from semantically distant feature spaces, and evaluate domain-specific gradient scaling and lower learning rate as mitigation strategies.

Discover Internet of ThingsVol. 6(1)
Amity University (IN), Manipal University Jaipur (IN)
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