mail2sia/VANET-IDS26

Vehicular Ad Hoc Network (VANET)-IDS26 Dataset

VANET-IDS26 is a large-scale Vehicular Ad Hoc Network intrusion detection dataset containing benign VANET messages and 26 cyberattack classes.

The dataset is designed for research on VANET cybersecurity, intrusion detection systems, machine learning, deep learning, transformer-based models, graph learning, and federated learning.

This repository provides documentation, scripts, manifests, and a small balanced sample. The full dataset is hosted on Hugging Face Datasets.

---

Full Dataset Download

The full dataset is available on Hugging Face:

https://huggingface.co/datasets/mail2sia/VANET-IDS26/tree/main/data

Hugging Face repository:

https://huggingface.co/datasets/mail2sia/VANET-IDS26

Full master file:

data/vanet_ids26_master.csv.gz

The full master file is approximately 84.0 GB compressed (about 78.2 GiB).

---

Operational 99:1 Test Split

For deployed NIDS-style evaluation, this release also includes a pre-built operational test split with benign traffic dominating the stream:

data/vanet_ids26_test_99_1_operational_max.csv.gz

This split is intended for realistic rare-anomaly evaluation where benign safety messages are approximately 99% of traffic and attacks are 1%.

Total rows:   37,957,500
Benign rows:  37,577,925
Attack rows:     379,575
Ratio:        99:1 exactly
Attack labels: multiclass_label 1-26 all represented
Attack sampling: proportional stratification across the 26 attack classes
Random seed: 42

Supporting metadata:

manifests/vanet_ids26_test_99_1_operational_max_manifest.json
manifests/vanet_ids26_test_99_1_operational_max_attack_distribution.csv
scripts/create_realistic_vanet_ids26_test_split.py

The operational split uses the maximum strict 99:1 size supported by the available benign rows. It uses 37,577,925 of the 37,577,963 benign rows and leaves 38 benign rows unused because the benign count is not exactly divisible by 99.

---

Citation Requirement

If you use VANET-IDS26 in any publication, thesis, dissertation, report, benchmark, software release, model release, or derivative dataset, please cite this dataset and the associated research work.

A formal citation and DOI will be added after the dataset paper or archive DOI is finalized.

Temporary citation:

@ARTICLE{Ahsan2024-mo,
  title  = "Privacy-preserving intrusion detection in software-defined {VANET} using Federated Learning with {BERT}",
  author = "Ahsan, Shakil Ibne and Legg, Phil and Iftekharul Alam, S M",
  year   =  2024,
  eprint = "2401.07343"
}

---

Dataset Summary

Item	Value
Dataset name	VANET-IDS26
Full master file	data/vanet_ids26_master.csv.gz
Full master file size	Approximately 84.0 GB compressed
Operational 99:1 split	data/vanet_ids26_test_99_1_operational_max.csv.gz
Operational split size	Approximately 5.16 GB compressed
Operational split rows	37,957,500
Total rows	786,798,378
Benign rows	37,577,963
Attack rows	749,220,415
Columns	46
Binary classes	2
Multiclass labels	0-26
Attack classes	26
Final simulation runs	20
Density levels	4
Attacker ratios	5%, 10%, 20%, 30%
Sample rows	27,000
Sample size	Approximately 2.26 MB

---

Attack Class Coverage by File

Both released data files cover all 26 attack families. The full master file is the complete attack-heavy research dataset. The operational 99:1 split is the deployment-style rare-anomaly test set with proportionally stratified attacks.

Full Master File

File:

data/vanet_ids26_master.csv.gz

Label	Attack name	Rows present
1	constant_position	23,337,078
2	position_offset	23,493,062
3	random_position	23,512,220
4	speed_manipulation	23,443,579
5	acceleration_manipulation	23,362,114
6	heading_manipulation	23,355,519
7	lane_spoofing	23,331,397
8	impossible_kinematics	23,492,276
9	eventual_stop	23,270,948
10	false_brake_event	23,361,712
11	false_emergency_vehicle	23,328,526
12	false_hazard_event	23,412,792
13	replay	23,349,590
14	delayed_message	23,544,098
15	timestamp_shift	23,286,623
16	stale_message_replay	23,409,666
17	sybil	23,447,400
18	impersonation	23,399,568
19	pseudonym_abuse	23,380,414
20	flooding_ddos	116,960,630
21	beacon_rate_abuse	46,968,602
22	gnss_spoofing	23,570,391
23	map_location_spoofing	23,450,720
24	ghost_vehicle	46,923,926
25	false_object_injection	23,395,318
26	object_position_shift	23,432,246

Master file totals:

Benign rows: 37,577,963
Attack rows: 749,220,415
Total rows:  786,798,378

Operational 99:1 Split

File:

data/vanet_ids26_test_99_1_operational_max.csv.gz

Label	Attack name	Rows present
1	constant_position	11,823
2	position_offset	11,902
3	random_position	11,912
4	speed_manipulation	11,877
5	acceleration_manipulation	11,836
6	heading_manipulation	11,832
7	lane_spoofing	11,820
8	impossible_kinematics	11,902
9	eventual_stop	11,790
10	false_brake_event	11,836
11	false_emergency_vehicle	11,819
12	false_hazard_event	11,862
13	replay	11,829
14	delayed_message	11,928
15	timestamp_shift	11,798
16	stale_message_replay	11,860
17	sybil	11,879
18	impersonation	11,855
19	pseudonym_abuse	11,845
20	flooding_ddos	59,255
21	beacon_rate_abuse	23,796
22	gnss_spoofing	11,941
23	map_location_spoofing	11,881
24	ghost_vehicle	23,773
25	false_object_injection	11,853
26	object_position_shift	11,871

Operational split totals:

Benign rows: 37,577,925
Attack rows:    379,575
Total rows:  37,957,500
Ratio:        99:1 exactly

---

Repository Contents

This Hugging Face dataset repository contains the compressed master dataset, the operational 99:1 test split, manifests, scripts, and a small sample for quick loading tests.

VANET-IDS26/
|-- data/
|   |-- vanet_ids26_master.csv.gz
|   `-- vanet_ids26_test_99_1_operational_max.csv.gz
|-- docs/
|-- manifests/
|-- samples/
|   `-- vanet_ids26_sample.csv.gz
|-- scripts/
`-- README.md

---

Sample File

A small balanced sample is included for quick testing:

samples/vanet_ids26_sample.csv.gz

The sample contains:

1,000 benign rows
1,000 rows from each attack class 1-26
Total rows: 27,000
Columns: 46
Compressed size: approximately 2.26 MB

This sample has the same schema as the full master file and balanced labels from 0 through 26.

---

Dataset Purpose

VANET-IDS26 was created to support reproducible research on intrusion detection in vehicular networks.

It can be used for:

binary intrusion detection
multiclass attack classification
transformer-based IDS using text_input
classical machine learning IDS
deep learning IDS
graph-based IDS
federated learning IDS
non-IID client partitioning
density-based robustness evaluation
attacker-ratio robustness evaluation
attack-family generalization

---

Dataset Design

The dataset contains two main types of records:

benign_base      original benign VANET messages
attack_overlay   attacked or synthetic attack overlay messages

The final master file combines:

benign VANET messages + attack overlay messages

into one single compressed CSV file:

vanet_ids26_master.csv.gz

This design allows researchers to load one file directly and create their own train, validation, and test splits.

---

Why One Master File?

The dataset is released as one model-ready file so users do not need to reconstruct the dataset from raw simulation folders.

Researchers can directly load:

data/vanet_ids26_master.csv.gz

and choose their own splitting method, such as:

random stratified split
run-level split
density-level split
attacker-ratio split
attack-family split
federated client split

The dataset intentionally does not force one fixed train/validation/test split.

---

Source Data Used to Build the Master File

The final master file was built from the final 20 large simulation runs:

dataset/processed/run_001001/messages.csv
dataset/processed/run_001002/messages.csv
...
dataset/processed/run_001020/messages.csv

and attack overlays:

dataset/overlays/run_001001/**/attack_overlay.csv
dataset/overlays/run_001002/**/attack_overlay.csv
...
dataset/overlays/run_001020/**/attack_overlay.csv

The following old pilot folders are excluded from the final master file:

dataset/processed/run_000001
dataset/processed/run_000002
dataset/attacks/

---

Simulation Runs and Density Levels

The final dataset contains 20 large simulation runs.

Runs	Density
run_001001-run_001005	low
run_001006-run_001010	medium
run_001011-run_001015	high
run_001016-run_001020	very_high

The density is included in the density column.

---

Attacker Ratios

Attack overlays were generated with four attacker ratios:

Value in dataset	Meaning
5	5% attacker ratio
10	10% attacker ratio
20	20% attacker ratio
30	30% attacker ratio

For benign rows:

attack_ratio = 0

---

Labels

Binary Label

Use binary_label for binary intrusion detection.

0 = benign
1 = attack

Multiclass Label

Use multiclass_label for multiclass intrusion detection.

0  = benign
1  = constant_position
2  = position_offset
3  = random_position
4  = speed_manipulation
5  = acceleration_manipulation
6  = heading_manipulation
7  = lane_spoofing
8  = impossible_kinematics
9  = eventual_stop
10 = false_brake_event
11 = false_emergency_vehicle
12 = false_hazard_event
13 = replay
14 = delayed_message
15 = timestamp_shift
16 = stale_message_replay
17 = sybil
18 = impersonation
19 = pseudonym_abuse
20 = flooding_ddos
21 = beacon_rate_abuse
22 = gnss_spoofing
23 = map_location_spoofing
24 = ghost_vehicle
25 = false_object_injection
26 = object_position_shift

---

Attack Taxonomy

Label	Attack type	Family	Brief description
0	benign	benign	Normal VANET message
1	constant_position	position_falsification	Reports a fixed false position
2	position_offset	position_falsification	Reports a shifted false position
3	random_position	position_falsification	Reports randomly falsified positions
4	speed_manipulation	kinematic_falsification	Manipulates claimed speed
5	acceleration_manipulation	kinematic_falsification	Manipulates claimed acceleration
6	heading_manipulation	kinematic_falsification	Manipulates claimed heading
7	lane_spoofing	lane_map_falsification	Reports a false lane
8	impossible_kinematics	kinematic_falsification	Creates physically implausible motion
9	eventual_stop	kinematic_falsification	Falsifies behavior toward stopping
10	false_brake_event	event_falsification	Injects or reports false braking events
11	false_emergency_vehicle	event_falsification	Claims false emergency-vehicle behavior
12	false_hazard_event	event_falsification	Injects or reports false hazard events
13	replay	timing_replay	Replays earlier VANET messages
14	delayed_message	timing_replay	Delays message timing
15	timestamp_shift	timing_replay	Shifts message timestamps
16	stale_message_replay	timing_replay	Replays stale messages as current
17	sybil	identity_spoofing	Creates multiple claimed identities
18	impersonation	identity_spoofing	Impersonates another vehicle
19	pseudonym_abuse	identity_spoofing	Abuses pseudonym or identity-change behavior
20	flooding_ddos	availability_flooding	Floods the network with excessive messages
21	beacon_rate_abuse	availability_flooding	Abuses beacon transmission rate
22	gnss_spoofing	position_falsification	Spoofs GNSS-derived position
23	map_location_spoofing	lane_map_falsification	Spoofs map-matched location
24	ghost_vehicle	identity_spoofing	Injects a non-existent vehicle identity
25	false_object_injection	event_falsification	Injects a false perceived object
26	object_position_shift	event_falsification	Shifts the reported position of an object

---

Column Schema

The full master file and the sample file both contain 46 columns.

Column	Description
dataset_name	Dataset identifier, always VANET-IDS26
record_id	Unique row identifier in the master file
source_type	benign_base or attack_overlay
source_file	Original file path used to create the row
base_run_id	Simulation run ID
density	Traffic density: low, medium, high, very_high
attack_ratio	Attacker ratio: 0, 5, 10, 20, or 30
message_id	Message ID in the master file
base_message_id	Original base message ID
overlay_message_id	Overlay message ID for attack rows
original_message_id	Original message ID for replay or delay attacks when applicable
attack_dataset_id	Attack overlay dataset identifier
time	Simulation time
claimed_time	Claimed message time
physical_sender_id	Actual sender vehicle ID
claimed_sender_id	Claimed sender identity
sequence_number	Message sequence number
true_x	True x-coordinate
true_y	True y-coordinate
true_speed	True vehicle speed
true_acceleration	True vehicle acceleration
true_heading	True vehicle heading
true_lane	True lane identifier
claimed_x	Claimed x-coordinate
claimed_y	Claimed y-coordinate
claimed_speed	Claimed vehicle speed
claimed_acceleration	Claimed vehicle acceleration
claimed_heading	Claimed vehicle heading
claimed_lane	Claimed lane identifier
malicious_delay_ms	Malicious delay in milliseconds, where applicable
sybil_group_id	Sybil group identifier, where applicable
event_type	Event type for event-based attacks
false_object_id	False object identifier, where applicable
object_type	Type of injected or shifted object
object_x	Object x-coordinate
object_y	Object y-coordinate
attack_start_time	Attack start time
attack_notes	Additional attack notes
is_attacker	Whether the physical sender is an attacker
is_synthetic_message	Whether the row is a synthetic injected message
attack_type	Human-readable attack type
attack_label	Numeric attack label
message_label	Original message label from source file
binary_label	0 for benign, 1 for attack
multiclass_label	0 for benign, 1-26 for attack class
text_input	BERT-ready text representation of the VANET message

---

Text Input for Transformer Models

The dataset includes a text_input column for transformer-based models such as BERT, DistilBERT, RoBERTa, and similar architectures.

Example format:

time=0.10 ; claimed_time=0.10 ; physical_sender_id=0 ; claimed_sender_id=0 ; sequence_number=1 ; true_x=4.80 ; true_y=1549.97 ; true_speed=12.68 ; ...

The text_input column intentionally excludes direct label columns such as:

attack_type
attack_label
message_label
binary_label
multiclass_label

This reduces direct label leakage.

---

Loading the Sample

import pandas as pd

df = pd.read_csv("samples/vanet_ids26_sample.csv.gz")

print(df.shape)
print(df["binary_label"].value_counts())
print(df["multiclass_label"].value_counts().sort_index())

Expected:

Rows: 27,000
Columns: 46
Labels: 0 through 26

---

Loading the Full Dataset

The full master file is large, so chunked loading is recommended.

import pandas as pd

path = "vanet_ids26_master.csv.gz"

for chunk in pd.read_csv(path, chunksize=100000):
    print(chunk.shape)
    print(chunk.head())
    break

Load only selected columns:

import pandas as pd

cols = [
    "text_input",
    "binary_label",
    "multiclass_label",
    "attack_type",
    "base_run_id",
    "density",
    "attack_ratio"
]

for chunk in pd.read_csv(
    "vanet_ids26_master.csv.gz",
    usecols=cols,
    chunksize=100000
):
    print(chunk.shape)
    break

---

Binary IDS Example

import pandas as pd
from sklearn.model_selection import train_test_split

df = pd.read_csv("samples/vanet_ids26_sample.csv.gz")

train_df, test_df = train_test_split(
    df,
    test_size=0.2,
    stratify=df["binary_label"],
    random_state=42
)

X_train = train_df["text_input"]
y_train = train_df["binary_label"]

X_test = test_df["text_input"]
y_test = test_df["binary_label"]

---

Multiclass IDS Example

import pandas as pd
from sklearn.model_selection import train_test_split

df = pd.read_csv("samples/vanet_ids26_sample.csv.gz")

train_df, test_df = train_test_split(
    df,
    test_size=0.2,
    stratify=df["multiclass_label"],
    random_state=42
)

X_train = train_df["text_input"]
y_train = train_df["multiclass_label"]

X_test = test_df["text_input"]
y_test = test_df["multiclass_label"]

---

Suggested Evaluation Protocols

Protocol A: Random Stratified Split

Use this for quick baseline experiments.

Split randomly while preserving class distribution.

Protocol B: Run-Level Split

Use this to reduce leakage across simulation runs.

Example:

Train:      run_001001 to run_001014
Validation: run_001015 to run_001017
Test:       run_001018 to run_001020

Protocol C: Density Holdout

Use this to evaluate generalization to unseen traffic density.

Example:

Train: low, medium, high
Test:  very_high

Protocol D: Attacker-Ratio Holdout

Use this to evaluate generalization to unseen attacker intensity.

Example:

Train: 5%, 10%, 20%
Test:  30%

Protocol E: Attack-Family Holdout

Use this to evaluate generalization to unseen attack families.

Example:

Train: position_falsification, kinematic_falsification, timing_replay
Test:  identity_spoofing or availability_flooding

Protocol F: Federated Non-IID Split

Use this for federated learning.

Possible client definitions:

client = physical_sender_id
client = base_run_id
client = density
client = attack_family
client = lane/geographic partition

---

Recommended Model Inputs

For transformer-based models:

text_input

For structured machine learning models:

time
claimed_time
physical_sender_id
claimed_sender_id
sequence_number
true_x
true_y
true_speed
true_acceleration
true_heading
true_lane
claimed_x
claimed_y
claimed_speed
claimed_acceleration
claimed_heading
claimed_lane
malicious_delay_ms
sybil_group_id
event_type
object_x
object_y
is_attacker
is_synthetic_message

Depending on the deployment scenario, users may exclude is_attacker and is_synthetic_message because these fields may be too informative for some real-world IDS assumptions.

---

Columns to Avoid as Features

Do not use direct label or identifier columns as model inputs:

attack_type
attack_label
message_label
binary_label
multiclass_label
record_id
source_file
attack_dataset_id

Using these as model features may cause label leakage.

---

Class Imbalance

The full master file is not balanced:

Benign rows: 37,577,963
Attack rows: 749,220,415

Researchers may choose to:

use all rows
downsample attack rows
upsample benign rows
create balanced subsets
use class weights
use stratified sampling
evaluate per-attack-class metrics

The sample file is balanced across labels 0-26 for easy testing.

---

Validation Status

The dataset was validated before release.

Overlay run folders: 20
Attack overlay files: 2080
Overlay manifest lines: 2081
Processed big-run folders: 20
Master rows: 786,798,378
Master columns: 46
Python quick-read validation: passed
Full gzip validation: MASTER GZIP OK
Sample gzip validation: SAMPLE GZIP OK
Sample labels: 0-26, 1,000 rows each

---

Rebuilding the Master File

The master file was generated using:

scripts/build_vanet_ids26_master.py

Example:

python scripts/build_vanet_ids26_master.py \
  --output research_ready/vanet_ids26_master.csv.gz \
  --compresslevel 1

---

Creating the Sample File

The sample file was generated using:

scripts/create_vanet_ids26_github_sample.py

Example:

python scripts/create_vanet_ids26_github_sample.py

---

Intended Use

VANET-IDS26 is intended for:

defensive cybersecurity research
VANET intrusion detection benchmarking
machine learning experiments
deep learning experiments
transformer-based IDS experiments
federated learning research
cybersecurity education
reproducible benchmarking

---

Ethical Use

This dataset is intended for defensive research only. It should be used to build, evaluate, and improve intrusion detection and resilience methods for vehicular networks.

---

License

This repository is released under the MIT License. See License for the full license text.

---

Maintainer

mail2sia

---

Citation Reminder

Please cite VANET-IDS26 and the associated research work if you use this dataset.