VeReMiVNDN

A labeled dataset and simulation framework for misbehavior detection in Vehicular Named Data Networking (VNDN), built on the JubST (Al-Jubail) SUMO scenario.

This repository accompanies the IEEE Access article:

B. G. Aldahlan and M. Uzair, "VeReMiVNDN: A Dataset for Misbehavior Detection in Vehicular Named Data Networking Using the JubST SUMO Scenario," IEEE Access, 2026. doi:10.1109/ACCESS.2026.3681896

---

Download the dataset

The data is published as GitHub Releases, not as files in this repository. Git is a poor fit for a few GB of compressed CSV, so the release assets carry the data and this repository carries the code that produced it.

The main archive is uploaded in 48 MB parts. A single large asset has no resume, and this release was produced over a slow uplink, so a stalled transfer would otherwise mean starting the whole file again. Fetch the parts and join them with cat; the order is guaranteed by the shell glob.

# the labeled machine-learning dataset, one gzipped CSV per run
for i in 00 01 02 03 04 05 06 07; do
  curl -LO https://github.com/Muhammaduazir69/VeReMiVNDN-dataset/releases/download/v1.0.0/veremivndn_ml.tar.gz.part$i
done
cat veremivndn_ml.tar.gz.part* > veremivndn_ml.tar.gz
tar xzf veremivndn_ml.tar.gz

The per-neighbor view is small enough to come as one file:

curl -LO https://github.com/Muhammaduazir69/VeReMiVNDN-dataset/releases/download/v1.0.0/veremivndn_plane.tar.gz

Verify what you downloaded. CHECKSUMS.sha256 covers both the individual parts and the reassembled archives, so it will catch a truncated part and a bad join.

curl -LO https://github.com/Muhammaduazir69/VeReMiVNDN-dataset/releases/download/v1.0.0/CHECKSUMS.sha256
sha256sum -c CHECKSUMS.sha256

Load it

Every file is a gzipped CSV with a header, so pandas reads it directly.

import glob
import pandas as pd

df = pd.concat(
    [pd.read_csv(f) for f in glob.glob("data/ml/*.csv.gz")],
    ignore_index=True,
)

print(len(df), "rows")
print(df.attack.value_counts())

Each row already carries the columns that identify which run it came from, so the experimental grid can be rebuilt from the released files alone without consulting any side-car index.

---

The label schema

This is the part most people ask about, so it is stated explicitly.

Column	Meaning
attack	The attack configured for the run: None, InterestFlooding, NamePrefixHijacking, InterestAggregation or RoutingInfoFlood. Constant within a run.
attackType	The ground truth for the row. The attack the recording node was executing at that instant, or Benign.
isAttack	1 when attackType is not Benign, 0 otherwise. This is the binary training target.
attackIntensity	Configured intensity of the active attack, 0.0 when benign.
severity, attackLayer	Severity class and protocol layer of the active attack.
nodeId, timestamp	Recording node and simulation time.

Two distinctions matter and are easy to get wrong:

attack is not attackType. attack describes the run; attackType describes the row. In an attack run, a malicious vehicle is labeled Benign before its attack window opens and again after it closes, and every benign vehicle in that run is labeled Benign throughout. Training on attack rather than attackType labels the majority of an attack run's rows as malicious and will produce misleadingly high scores.

Ground truth is not an input feature. The 69 feature columns are traffic and table statistics observed at the recording node. The label columns are resolved from the simulator's own configuration and are supervision only. Drop them before fitting:

LABELS = ["timestamp", "nodeId", "attackType", "isAttack",
          "attackIntensity", "severity", "attackLayer"]
IDS = ["run", "config", "attack", "density", "vehicles",
       "attacker_ratio_pct", "seed"]

X = df.drop(columns=LABELS + IDS)
y = df.isAttack

---

What is in a release

Asset	Contents
veremivndn_ml.tar.gz.part00 .. part07	The main dataset, in 48 MB parts. One gzipped CSV per run, 69 features plus labels. Join with cat.
veremivndn_plane.tar.gz	A second view: 23 per-neighbor features, one row per monitor and observed neighbor.
veremivndn_scalars.tar.gz.part00 .. part03	The OMNeT++ scalar file for every run, holding the network-level measurements. Join with cat.
CHECKSUMS.sha256	SHA-256 of every part and of the reassembled archives.

Documentation lives in docs/DATA_DICTIONARY.md, which defines every column, and docs/INSTRUCTIONS.md, which covers loading, splitting and the limitations to read first.

The four attacks

The paper evaluates four forwarding-layer attacks, each with its own run configuration.

Configuration	Attack	What the attacker does
Attack01_InterestFlooding	Interest Flooding	Emits Interests for names that cannot be satisfied, so PIT entries accumulate and expire.
Attack05_NamePrefixHijacking	Name Prefix Hijacking	Advertises prefixes it does not serve, drawing Interests away from legitimate producers.
Attack12_InterestAggregation	Interest Aggregation Exploitation	Abuses PIT aggregation so that one Interest suppresses the forwarding of others.
Attack18_RoutingInfoFlood	Routing Information Flood	Floods routing updates to destabilize the FIB.
BenignTraffic	None	Attacker-free control.

The simulator in src/attacks/ implements a wider set of attacks across the caching, privacy, trust and cross-layer planes. Those are not part of this dataset and are not evaluated in the paper. Only the five configurations above were run.

The experimental grid

225 runs of 300 simulated seconds each:

5 configurations x 3 densities x 3 attacker ratios x 5 seeds = 225

Axis	Levels
Density	low ~77 vehicles, med ~151, high ~356
Attacker ratio	5 %, 10 %, 20 % of the fleet
Seed	1 to 5, varying both mobility and application traffic

Density is set through SUMO demand scaling, calibrated against the JubST route file over a 300 s horizon. Attackers occupy the low vehicle indices. The attack window opens at t = 60 s and runs for 180 s, so every run contains benign traffic before, during and after the attack.

BenignTraffic has no attackers, so its three ratio levels are replicates rather than distinct conditions. The axis is kept so the design stays balanced and the run count matches the published figure; the released rows record an attacker ratio of 0 for all of them.

Dataset at a glance

Property	Value
Runs	225
Rows, machine-learning view	5,989,800
Rows, per-neighbor view	924,567
Feature columns	69
Attack rows (isAttack = 1)	653,140 (10.9 %)
Simulated time	67,500 s
Compressed size	549M

Which attacks are observable, and through what

An attack that runs in the simulator is not automatically visible in the features. Each attack below executes and is labeled, and this table is measured from the released files rather than asserted: Cohen's d is the standardized mean difference between attacker and benign rows for the single most separating feature, pooled over all forty-five runs of that attack.

Attack	Attacker rows	Best Cohen's d	Feature	Verdict
InterestAggregation	163,285	3.09	pitSize	clearly separable
InterestFlooding	163,285	12.57	burstiness	clearly separable
NamePrefixHijacking	163,285	1.72	longTermInterestRate	clearly separable
RoutingInfoFlood	163,285	1.73	longTermInterestRate	clearly separable

All four are separable, but not in the same way, and the difference matters for how you model them.

Interest Flooding and Interest Aggregation change the short-term picture: the attacker's burstiness and PIT occupancy move sharply, so a detector reading a single observation window can find them.

The two routing-plane attacks do not. Name Prefix Hijacking and Routing Information Flood leave every short-window statistic untouched, to three decimal places: interestRate, burstiness and windowInterestCount are the same for attacker and benign rows. They show up only in longTermInterestRate, which accumulates over a longer horizon, where the hijacker sits at about 0.46 against 0.22 for a benign node. The forged advertisements and routing updates are real, and the scalar files record them in the tens of thousands, but they register as a slow rise in accumulated volume rather than as a burst.

The practical consequence is that a model given only single-window features will find the first two attacks and miss the last two. If you care about the routing attacks, keep the long-horizon features, or build sequences over consecutive windows.

There is no purpose-built routing feature: the set has no advertisement or FIB churn column, and hijackingScore is an unimplemented stub. Adding one is the most useful extension available here, and the ground truth for that work is already in data/scalars/ as forgedAdvertisements, hijackedPrefixCount and interceptedInterests.

---

Build and run it yourself

See INSTALL.md for the full toolchain. In short:

make                                     # build the simulator
scripts/veremivndn_paper1_campaign.sh    # regenerate all 225 runs
scripts/consolidate_paper1.py --root datasets/paper1_veremivndn --out release/data

A single run:

scripts/miids/run_one.sh Attack01_InterestFlooding 300s "" 0 \
  --seed-set=1 \
  --'*.manager.launchConfig=xmldoc("../scenarios/sumo/jubst_low_s1.launchd.xml")'

Notes on the data

Read these before drawing conclusions.

• The dataset was regenerated after publication. Five defects in the export path were repaired first, so the released files are not a byte-for-byte reproduction of the tables in the article. Two of them were severe enough that the original pipeline could not have produced a usable dataset at all: the routine that fills the feature window had no caller, so every one of the 69 features held its default value, and no attack module ever reported itself, so every row was labeled benign whatever the configuration. The others were that all recording nodes wrote to one filename and truncated it, that the collector read the window after the extractor had cleared it, and that the mobility features were hard-coded placeholders. RELEASE_NOTES.md lists each one. Numbers computed from this release will differ from the published tables, and the release is the version to trust.
• Some columns are still constant, either because the quantity is not instrumented in this build or because the feature is an unimplemented stub. The data dictionary lists exactly which, measured from the released files rather than assumed.
• One map, one vehicle class. All runs use the JubST road network. Results should not be assumed to transfer to another topology without re-measuring.
• Attackers are static. Each attacker holds a fixed rate and strategy for its whole window and does not react to detection. A detector scored only on this data will look better than it would against an adaptive adversary.
• Class imbalance is real and intended. Attack rows are a minority. Use stratified splits and report precision and recall, not accuracy.

Repository layout

src/           simulator sources (NDN forwarding, attacks, IDS, logging)
simulations/   OMNeT++ configurations and the JubST SUMO scenario
scripts/       campaign, consolidation and release tooling
docs/          data dictionary, instructions and the project page

License

The dataset and documentation are released under CC BY 4.0. The simulator source is released under the MIT license.

Citation

See CITATION.cff, or cite the article directly:

@article{aldahlan2026veremivndn,
  author  = {Aldahlan, Bassma G. and Uzair, Muhammad},
  title   = {{VeReMiVNDN}: A Dataset for Misbehavior Detection in Vehicular
             Named Data Networking Using the {JubST} {SUMO} Scenario},
  journal = {IEEE Access},
  year    = {2026},
  doi     = {10.1109/ACCESS.2026.3681896}
}

Contact

Questions about the data are best raised as a GitHub issue.

Muhammad Uzair, COMSATS University Islamabad, muhammaduzairr69@gmail.com