Communication-aware dynamic graph learning for V2X misbehavior detection using GAT–BiLSTM and bagged trees

Vehicle-to-Everything (V2X) communication supports cooperative driving, which also makes the malicious vehicle-state reports influence nearby vehicles. Due to the isolated use of physical state and cross-time communication context, existing methods struggle to detect malicious reports that are physically self-consistent and only reveal inconsistencies at the temporal communication level. Thus, this paper proposes a sequence-level misbehavior detection framework based on graph attention networks (GAT) and bidirectional long short-term memory (BiLSTM). Recorded sender-receiver interactions define a directed communication graph for each one-second window. The GAT can learn the neighboring information contributes to detection. The BiLSTM captures changes across ten consecutive graph snapshots. The resulting representation is fused with physical statistics. Then, a bagged-tree ensemble is introduced to classify it. Experiments on curated VeReMi NextGen data distinguish normal behavior from three attack types. These are position offset, speed offset and data replay. Across ten random seeds the method achieves 96.40 ± 0.49% accuracy and 96.42 ± 0.48% Macro-F1. Compared with Physical features + Trees, the paired mean improvements are approximately 0.61 percentage points in accuracy and 0.62 percentage points in Macro-F1. The corresponding 95% confidence intervals exclude zero. Ablation results identify physical statistics as the main source of detection information. Combined with these statistics, the learned representation can provide a modest additional gain. Further tests show higher detection scores under feature-space Gaussian noise and changes in attack-related feature magnitude. Additional tests examine sensitivity to cross-class window substitution. Non-overlapping evaluation produces only a small decline in detection performance.

Authors

Institutions

Publication Details

Journal
PLoS ONE
Published
2026-11-16
DOI
https://doi.org/10.1371/journal.pone.0362960
Primary Topic
Vehicular Ad Hoc Networks (VANETs)
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Communication-aware dynamic graph learning for V2X misbehavior detection using GAT–BiLSTM and bagged trees

Xinyu Wang, Ping Wang, Yuee Zhang, Xiao Zheng
1 links
PLoS ONE
Vehicular Ad Hoc Networks (VANETs)
article

Communication-aware dynamic graph learning for V2X misbehavior detection using GAT–BiLSTM and bagged trees

Xinyu Wang, Ping Wang, Yuee Zhang, Xiao Zheng
article en

Abstract

Vehicle-to-Everything (V2X) communication supports cooperative driving, which also makes the malicious vehicle-state reports influence nearby vehicles. Due to the isolated use of physical state and cross-time communication context, existing methods struggle to detect malicious reports that are physically self-consistent and only reveal inconsistencies at the temporal communication level. Thus, this paper proposes a sequence-level misbehavior detection framework based on graph attention networks (GAT) and bidirectional long short-term memory (BiLSTM). Recorded sender-receiver interactions define a directed communication graph for each one-second window. The GAT can learn the neighboring information contributes to detection. The BiLSTM captures changes across ten consecutive graph snapshots. The resulting representation is fused with physical statistics. Then, a bagged-tree ensemble is introduced to classify it. Experiments on curated VeReMi NextGen data distinguish normal behavior from three attack types. These are position offset, speed offset and data replay. Across ten random seeds the method achieves 96.40 ± 0.49% accuracy and 96.42 ± 0.48% Macro-F1. Compared with Physical features + Trees, the paired mean improvements are approximately 0.61 percentage points in accuracy and 0.62 percentage points in Macro-F1. The corresponding 95% confidence intervals exclude zero. Ablation results identify physical statistics as the main source of detection information. Combined with these statistics, the learned representation can provide a modest additional gain. Further tests show higher detection scores under feature-space Gaussian noise and changes in attack-related feature magnitude. Additional tests examine sensitivity to cross-class window substitution. Non-overlapping evaluation produces only a small decline in detection performance.

PLoS ONEVol. 21(11)
Yanshan University (CN), Shandong Huayu University of Technology (CN)
Vehicular Ad Hoc Networks (VANETs)
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