PaperlayerView exact revision on arXiv

Source-linked AI summary

SPADE: SPaT Attack Detection from the Connected Vehicle's Perspective

James Di Novo, Hany Ragab, Sylvain P. Leblanc

arXiv:2609.02741v1cs.CRcs.LG

TL;DR

SPADE addresses the absence of a purpose-built dataset for onboard CV-side detection of SPaT attacks, which existing work treats mainly from infrastructure or V2V perspectives. It constructs a labelled, multi-modal simulation dataset through SAE J2735 application-layer attack injection and provides balanced scenario-based data for IDS research. Its scope is bounded by a closed-world label set and simulation limitations, including fixed SPaT timing and vehicles that do not act on falsified messages.

  • Problem

    Existing IDS research lacks a purpose-built dataset combining onboard CV-side SPaT integrity with vehicle sensors and V2V communications.

  • Method

    SPADE uses Eclipse MOSAIC to inject six SPaT attack classes and one benign class at the SAE J2735 application layer while fusing SPaT, camera, and cooperative V2V data.

  • Results

    ∼1,890,000 balanced labelled records across 180 scenario runs provide a multi-modal dataset for deep-learning IDS research.

  • Takeaways & Limitations

    SPADE supports reproducible and comparative IDS research on C-V2X SPaT attacks from the connected vehicle’s perspective.

  • Takeaways & Limitations

    SPADE assumes seven exhaustive label classes, uses fixed timing, and does not let vehicles operate on falsified malicious messages.

Abstract

from arXiv · show

Signal Phase and Timing (SPaT) messages are a cornerstone of connected vehicle (CV) safety, enabling CVs to perceive and respond to intersection state through Vehicle-to-Infrastructure (V2I) and Vehicle-to-Vehicle (V2V) communication. The integrity of these messages is threatened by a range of application-layer attacks that can bypass conventional authentication when a roadside unit or peer vehicle is compromised. Existing intrusion detection research either defends the infrastructure side or targets V2V Basic Safety Message (BSM) / Cooperative Awareness Message (CAM) misbehavior, leaving the onboard CV perspective on SPaT integrity unaddressed.To close this gap, we introduce SPADE --- the SPaT Attack Detection and Evaluation dataset --- a labelled, multi-modal, simulation-based dataset designed specifically for deep learning IDS research in this space. SPADE is generated through Eclipse MOSAIC using runtime attack injection at the SAE J2735 application layer across six attack classes and one benign class. By combining four intersection geometries, six operating conditions, and five independent random-seed repetitions, SPADE comprises 180 unique base scenario runs, yielding $\sim$1,890,000 labelled timestep records (270,000 per class). Each record fuses SPaT message fields, onboard camera confidence scores, and cooperative V2V peer data across 40 features, reflecting the multi-modal signal space required to distinguish deliberate attacks from environmental degradation. The dataset, generation code, and scenario configurations are released publicly to support reproducible and comparative IDS research in C-V2X security. The developed toolbox, instructions, and dataset link are publicly available on GitHub: https://github.com/jdinovo/SPADE.

I. INTRODUCTION

SPADE addresses the lack of a purpose-built onboard dataset for detecting SPaT attacks by combining application-layer attack injection with synchronized CV-side modalities. It provides labelled data spanning multiple attack classes and scenario configurations for IDS research.

  • Threat and research gap: A compromised RSU or peer CV can inject false, stale, conflicting, suppressed, or impersonated SPaT messages despite valid credentials.Application-layer detection must therefore complement cryptographic authentication by examining message content and cross-modal consistency.
  • Threat and research gap: Existing IDS datasets and studies focus on infrastructure-side signal security or V2V BSM misbehavior rather than onboard CV-side SPaT integrity.The paper identifies no prior dataset combining onboard SPaT data with vehicle sensors and V2V communications.
  • Dataset contribution: SPADE injects attacks at the SAE J2735 application layer in Eclipse MOSAIC while maintaining consistency among SPaT/MAP, camera, and CAM/CPM streams.This design represents the multi-modal observations available to a receiving connected vehicle.
  • Dataset contribution: ∼1,890,000 labelled records cover six SPaT attack classes and one benign class across 180 unique scenario configurations.The dataset contains 270,000 records per class and targets deep-learning IDS research.

II. RELATED WORK AND DATASET GAP

Prior work addresses infrastructure-side traffic-signal security, general V2X or V2V misbehavior, and real-world vehicular attacks, but not the combined onboard CV-side SPaT and cross-modal setting targeted by SPADE.

  • Infrastructure-side work: Infrastructure-oriented studies detect attacks against traffic-signal controllers using infrastructure sensors, traffic invariants, or adaptive-controller resilience methods.These approaches do not represent detection from the receiving vehicle’s perspective.
  • CV-side and ML work: The closest cited CV-oriented patent describes V2I misbehavior identification but provides no labelled dataset, machine learning, or cross-modal data.A survey likewise identifies onboard ML detection of SPaT-specific attacks as an open area.
  • Existing datasets: VeReMi and VeReMi NextGen target V2V BSM spoofing, while ROAD contains real-world vehicular attacks from the infrastructure perspective rather than SPaT-specific CV-side attacks.Neither VeReMi dataset incorporates camera or cross-modal fusion data.
  • Dataset scale: 270,000 samples per class produced competitive F1 scores for seven-class classification in related VeReMi-extension work, motivating SPADE’s per-class target.Other cited datasets address different attack surfaces, including CAN-bus or general VANET communications.

C. Simulation of Urban Mobility (SUMO) Scenarios

The paper defines a CV-perspective threat model for SAE J2735 application-layer SPaT attacks, with scope bounded by compromised authenticated nodes, excluded attack surfaces, and controlled simulation assumptions.

  • Scenario and gap summary: SPADE uses a SUMO scenario based on Ottawa to represent North American road types and complex signalized intersections.The authors position Ottawa as suitable for a SPaT-centric dataset.
  • Scenario and gap summary: The six selected attack types apply specifically to SPaT messages from the CV perspective, although related attacks have been studied against CAM, BSM, or traffic-signal controllers.This distinguishes the threat focus from prior perspectives.
  • Assumptions: The attacker controls a compromised RSU or peer CV that retains valid PKI credentials, so authentication does not flag its transmissions.Detection therefore relies on message content and consistency across modalities.
  • Assumptions: The attacker seeks to make the CV misinterpret intersection state by altering phase perception, stopping behavior, cross-source agreement, or SPaT availability.The threat model targets effects observable by the receiving vehicle.
  • Assumptions: Physical-layer jamming, GPS spoofing, CAN-bus attacks, and adversarial attacks on trained models are excluded from the application-layer scope.All six attacks modify SPaT messages without requiring network-layer simulation.

B. Attack Classes

SPADE’s label schema contains six SPaT-relevant attack classes and one benign class, with attacks represented through application-layer effects observable by a receiving CV.

  • Label schema: Table II defines six attack classes and one benign class using targeted SAE J2735 fields and observable multi-modal signatures.The schema links each label to both manipulated message content and cross-modal evidence.
  • Denial of Service: Denial of Service is modeled as application-layer message suppression rather than channel flooding.This keeps the attack representation within the receiving CV’s observable message layer.
  • Denial of Service: Suppression is detectable through the inter-message interval because SPaT broadcasts are absent within the expected reception window.The representation does not require ns-3 or OMNeT++ integration.

IV. SPADE DATASET GENERATION

SPADE uses Eclipse MOSAIC to simulate autonomous connected vehicles and inject attacks at the SAE J2735 application layer. It logs synchronized SPaT, V2V, and camera-perception data while modeling environmental visibility and camera confidence.

  • Simulation Environment: Eclipse MOSAIC integrates traffic mobility, vehicle applications, and V2X simulation for SPADE generation.Custom V2X applications run on individual simulation objects, enabling SAE J2735-level attack injection.
  • Simulation Environment: Autonomous vehicles receive SPaT, transmit CAMs, and log SPaT content, V2V peer data, and camera perception in per-vehicle CSV files.The vehicle setup removes human intervention, producing predictable sensor- and communication-driven decisions.
  • Perception Modeling: Environmental degradation is represented by visibility scores Vs from 0.95 for clear conditions to 0.20 for severe impairment.The model treats dense fog as an example of severe impairment, while weather and lens-distortion occlusion are not directly simulated.
  • Perception Modeling: Camera confidence Ccam is modeled from distance, angular offset, maximum range, and field of view, then bounded with noise and broadcast through V2V CPM.Distance and off-axis angle reduce detection quality quadratically, while a 0.4 floor represents residual detectability at the field-of-view boundary.

B. Scenario Configurations

SPADE varies geometry, traffic density, and camera visibility to create diverse simulation conditions. Combining four geometries with six operating-condition combinations yields 36 base configurations and 180 seeded scenario runs.

  • Geometry and Operating Conditions: SPADE includes four geometries: two four-way crossings, a T-intersection, and a pedestrian-actuated mid-block crossing.The geometries are derived from representative intersection and crossing layouts.
  • Geometry and Operating Conditions: Traffic density and camera visibility each have low, medium, and high levels across the scenario configurations.The operating dimensions include low, medium, and high traffic density, plus high, medium, and low camera visibility.
  • Configuration Matrix: 36 base scenario configurations result from combining four geometry types with six operating-condition combinations.The six combinations arise from three traffic-density levels and three visibility levels.
  • Configuration Matrix: 180 unique scenario runs result when every base configuration is repeated with five independent random seeds.Seeds vary vehicle route assignments and inter-arrival times.
  • Configuration Matrix: Each scenario run is executed seven times, once for every label class, producing 1260 simulation executions.The resulting record count varies by scenario according to traffic density and vehicle proximity.

C. Runtime Attack Injection

SPADE injects attacks at runtime through malicious V2X applications rather than modifying exported data afterward. The attack classes target SPaT state, timing, availability, identity, replay behavior, or peer-consensus signals.

  • Injection Mechanism: Runtime injection uses custom V2X applications on malicious RSUs or compromised peer vehicles, preserving synchronization across multimodal data sources.Attacks are simulated within MOSAIC rather than applied as posthoc CSV modifications.
  • SPaT Manipulation: FALSE_STATE modifies MovementPhaseState values, while REPLAY re-injects captured SPaT messages after a configurable delay.These attacks represent falsified-information and replay behaviors, respectively.
  • SPaT Manipulation: TIMING_MANIP preserves MovementPhaseState but alters MinEndTime and MaxEndTime in TimeChangeDetails.The attack targets SPaT timing information without changing the reported movement phase.
  • V2V and Availability Attacks: SYBIL uses phantom CV nodes to transmit conflicting perceived signal states over V2V, creating peer-consensus failure.The attack is implemented through multiple transmitting identities.
  • V2V and Availability Attacks: DOS silences the legitimate RSU for a defined interval, whereas IMPERSONATION spoofs the IntersectionID or source address.DOS produces SPaT absence detectable through the inter-message interval.

D. Ground-Truth Labelling

SPADE derives ground-truth labels from the simulation schedule and aligns them with exported vehicle records by timestamp. Runtime injection keeps the labelled multimodal streams consistent with each scenario.

  • Ground-Truth Labelling: The active attack class is logged at every simulation timestep for corroboration with the generated data.This provides a direct record of the label active during each timestep.
  • Generation Pipeline: Runtime attack injection ensures separate data streams reflect one consistent scenario during SPADE generation and use.This consistency applies across the multimodal streams produced by the pipeline.
  • Ground-Truth Labelling: Ground-truth labels are produced by joining the scenario schedule to exported CSV records on timestamp, without manual annotation.The labels are an inherent simulation byproduct rather than a separate annotation process.

V. DATASET DESCRIPTION

SPADE provides a balanced, scenario-level partitioned dataset with 40 multimodal features for training and evaluating SPaT intrusion-detection models. Its scale is motivated by prior vehicular IDS findings, while scenario diversity remains central to generalization.

  • 270,000 records per class yield an approximately balanced seven-class dataset designed for deep learning IDS training.The full dataset contains approximately 1,890,000 records distributed equally across seven classes.
  • 180 unique base runs are split by scenario into 126 training, 27 validation, and 27 test runs.This produces 189,000 training, 40,500 validation, and 40,500 test records per class.
  • Scenario-level partitioning prevents records from the same simulation run appearing across multiple splits, avoiding leakage from random record-level splitting.Random record-level splitting could place windows from one run in both training and testing and inflate detection rates.
  • 40 features combine SPaT/MAP messages, camera perception, V2V cooperative messages, and vehicle proprioception for cross-stream legitimacy checks.The multimodal design allows an IDS to identify discrepancies rather than relying solely on SPaT changes.
  • SPADE emphasizes 180 unique base runs because generalization risk is dominated by scenario diversity rather than raw sample count.The input space contains 40 features and is described as sufficiently compact for scenario diversity to be the primary design lever.

VI. FUTURE WORK

Future work extends SPADE from a released dataset toward comparative IDS evaluation. Proposed directions include benchmarking several architecture families on the dataset.

  • SPADE is released as a dataset contribution, while deep-learning model training and comparative benchmarking remain ongoing work.
  • Natural baseline families include 1D-CNN, stacked LSTM, and transformer encoders with self-attention over the input window.These architectures are identified from prior V2X IDS research and related connected-vehicle IDS applications.

B. Limitations and Future Directions

SPADE’s scope is bounded by a closed-world label set, fixed SPaT timing, and simulation constraints that limit vehicle reactions to falsified messages. The authors identify extensions for more realistic evaluation.

  • SPADE assumes its seven label classes are exhaustive, so models may not detect hybrid or parametrically novel attacks outside the training distribution.The authors identify this as a common limitation of simulation-based IDS datasets.
  • SPaT messages use fixed timing schedules rather than timings dynamically adjusted for traffic volume or other factors.The authors state that incorporating dynamic SPaT would create more difficult malicious-message detection scenarios.
  • SUMO prevents vehicles from operating on falsified attack messages, so logged decisions replace collision-enabled behavior and proprioceptive data remains aligned with ground truth.The authors suggest CARLA or similar tooling for collision avoidance and note that proprioception may still help identify Sybil attacks with CPMs and claimed positions.
  • SPADE is publicly released with generation code and scenario configurations as a labelled, multimodal onboard dataset for SPaT attack detection.
Loading 2609.02741v1…