Dense-Autoencoder Anomaly Detection for 5G Network Traffic: A Reproducible Study of Detection Performance, Seed Sensitivity, and Reconstruction-Error Attribution

The decentralization of 5G network intelligence to Multi-access Edge Computing (MEC) nodes replaces a small number of centrally defended data centres with thousands of distributed, resource-constrained nodes, each requiring its own anomaly-detection capability rather than relying on centralized inspection. This shift demands detection methods that operate within strict per-node latency and compute budgets while remaining interpretable enough to support autonomous or operator-assisted decisions at the edge. We evaluate a Dense Autoencoder, trained exclusively on benign traffic, as a candidate for such distributed deployment, using the 5G-NIDD dataset and a fixed, pre-specified protocol across five independent random seeds. The detector sustains single-flow inference suitable for commodity edge hardware (mean latency 0.095\,ms for detection alone and 0.178\,ms including attribution on the primary machine, with a second machine showing the same qualitative pattern) and holds benign-flow false-positive rate near its 5% design target (5.05% \((\pm)\) 0.12%) with precision of 0.953 \((\pm)\) 0.097 (on a holdout that is 88.5% attack traffic). However, overall detection performance varies substantially across seeds (F1 = 0.561 \((\pm)\) 0.365), with per-class recall ranging from 0.9998 (ICMP Flood) to 0.220 (UDP Flood). We further compare two interpretability mechanisms suited to distributed, resource-limited nodes: a lightweight reconstruction-error attribution usable in real time, and an offline KernelSHAP explainer; the two agree on the dominant contributing feature in only 15% of flagged flows, exposing a gap between fast and principled explanation methods for networked edge security. We report seed instability and attribution disagreement as central findings relevant to deploying such detectors across heterogeneous, distributed 5G edge infrastructure, and release the complete pipeline for independent verification.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-20
DOI
https://doi.org/10.21203/rs.3.rs-11317159/v1
Primary Topic
Network Security and Intrusion Detection
Type
preprint

Funders

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

Dense-Autoencoder Anomaly Detection for 5G Network Traffic: A Reproducible Study of Detection Performance, Seed Sensitivity, and Reconstruction-Error Attribution

R. Ashok Kumar, Prajwal V.
1 links
Research Square
Network Security and Intrusion Detection
preprint

Dense-Autoencoder Anomaly Detection for 5G Network Traffic: A Reproducible Study of Detection Performance, Seed Sensitivity, and Reconstruction-Error Attribution

R. Ashok Kumar, Prajwal V.
preprint en

Abstract

The decentralization of 5G network intelligence to Multi-access Edge Computing (MEC) nodes replaces a small number of centrally defended data centres with thousands of distributed, resource-constrained nodes, each requiring its own anomaly-detection capability rather than relying on centralized inspection. This shift demands detection methods that operate within strict per-node latency and compute budgets while remaining interpretable enough to support autonomous or operator-assisted decisions at the edge. We evaluate a Dense Autoencoder, trained exclusively on benign traffic, as a candidate for such distributed deployment, using the 5G-NIDD dataset and a fixed, pre-specified protocol across five independent random seeds. The detector sustains single-flow inference suitable for commodity edge hardware (mean latency 0.095\,ms for detection alone and 0.178\,ms including attribution on the primary machine, with a second machine showing the same qualitative pattern) and holds benign-flow false-positive rate near its 5% design target (5.05% \((\pm)\) 0.12%) with precision of 0.953 \((\pm)\) 0.097 (on a holdout that is 88.5% attack traffic). However, overall detection performance varies substantially across seeds (F1 = 0.561 \((\pm)\) 0.365), with per-class recall ranging from 0.9998 (ICMP Flood) to 0.220 (UDP Flood). We further compare two interpretability mechanisms suited to distributed, resource-limited nodes: a lightweight reconstruction-error attribution usable in real time, and an offline KernelSHAP explainer; the two agree on the dominant contributing feature in only 15% of flagged flows, exposing a gap between fast and principled explanation methods for networked edge security. We report seed instability and attribution disagreement as central findings relevant to deploying such detectors across heterogeneous, distributed 5G edge infrastructure, and release the complete pipeline for independent verification.

Research Square
B.M.S. College of Engineering (IN)
BMS College of Engineering
Industry, innovation and infrastructure
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
Dense-Autoencoder Anomaly Detection for 5G Network Traffic: A Reproducible Study of Detection Performance, Seed Sensitivity, and Reconstruction-Error Attribution — R. Ashok Kumar, Prajwal V. · Research Square (2026) | TGRS Research Map | TGRS