CrossPlane-5G: Dual-Plane Telemetry Benchmark Dataset for 5G SA and O-RAN NIDS

Overview

CrossPlane-5G is an open, reproducible, dual-plane benchmark dataset generated from a containerized 3GPP Release-16 5G Standalone (SA) testbed utilizing OpenAirInterface Core and RAN with RFSimulator.

It features simultaneous, hardware TSC-aligned packet captures across two distinct observation planes:

1. Layer-A (UE Plane): Unencapsulated user-plane PDU session tunnel telemetry (oaitun_ue1).
2. Layer-B (N3 Transport Plane): Encapsulated 3GPP GTP-U transport telemetry with 32-bit GTP-U Tunnel Endpoint Identifiers (TEIDs).

---

Dataset Release Profile & Repository Structure

To support diverse computational environments (from high-performance GPUs to resource-constrained laptops), the dataset is made available in three complementary distribution formats:

1. Master Columnar Parquet File

• crossplane5g_master_subflows.parquet (931.9 MB): Single-file binary compressed Apache Parquet format containing all 12,252,610 windowed subflow records and 46 feature columns. Recommended for PyTorch, TensorFlow, and fast Pandas data loading.

2. Chunked Raw CSV Files (Full Dataset Split into 12 Parts)

For researchers preferring plain-text CSV format or operating under RAM limits, the 12.25 million master subflow records are split sequentially into 12 uncompressed CSV part files (~1,021,051 rows each, ~250 MB per file):

• crossplane5g_csv_part_01.csv
• crossplane5g_csv_part_02.csv
• crossplane5g_csv_part_03.csv
• crossplane5g_csv_part_04.csv
• crossplane5g_csv_part_05.csv
• crossplane5g_csv_part_06.csv
• crossplane5g_csv_part_07.csv
• crossplane5g_csv_part_08.csv
• crossplane5g_csv_part_09.csv
• crossplane5g_csv_part_10.csv
• crossplane5g_csv_part_11.csv
• crossplane5g_csv_part_12.csv

3. Quick Inspection Stratified Sample CSV

• crossplane5g_sample_100k.csv (49.5 MB): A 100,000-row stratified sample dataset preserving exact class distribution ratios for rapid exploratory data analysis, plotting, and hyperparameter tuning.

4. Raw Multi-Plane Packet Captures (PCAP Archives)

• campaign_pcap_raw.tar.gz.part_00 to part_22 (~524 MB each): 462.5 GB of raw microsecond-aligned packet captures spanning the full 24-hour experimental execution across 882 PCAP files. To reassemble:

  cat campaign_pcap_raw.tar.gz.part_* > campaign_pcap_raw.tar.gz
  tar -xzf campaign_pcap_raw.tar.gz
  

---

13-Class Canonical Scenario Taxonomy

Class ID	Scenario Class Name	Layer-A Subflows	Layer-B Subflows	Total Combined Rows	Description
00	00_benign_baseline	1,937,079	1,941,925	3,879,004	Multi-thread background profile (HTTP, Video, IoT, Pause)
01	01_dos_icmp_flood	60,751	58,036	118,787	High-rate ICMP echo request flood
02	02_dos_udp_flood	68,794	66,345	135,139	High-rate UDP datagram flood
03	03_dos_syn_flood	1,230,386	1,260,119	2,490,505	Volumetric TCP SYN handshake exhaustion
04	04_dos_http_flood	1,907,529	1,922,731	3,830,260	Application-layer GET request flood
05	05_dos_slowrate	265,013	246,917	511,930	Low-and-slow HTTP headers exhaustion (Slowloris)
06	06_scan_syn	95,815	92,027	187,842	Stealth SYN port scanning
07	07_scan_tcp_connect	96,119	92,529	188,648	Full 3-way TCP connect scanning
08	08_scan_udp	83,303	79,232	162,535	UDP service discovery scanning
09	09_benign_embb	42,595	39,598	82,193	3GPP eMBB high-throughput video streaming
10	10_benign_urllc	58,571	56,041	114,612	3GPP URLLC low-latency industrial control
11	11_benign_mmtc	66,213	63,191	129,404	3GPP mMTC massive IoT telemetry
12	12_benign_agentic_llm	214,687	207,064	421,751	Agentic LLM multi-leg API token streaming
Total		6,126,855	6,125,755	12,252,610	100% Deterministic Ground Truth

---

Data Loading Quickstart Examples

Option 1: Load Full Dataset from Master Parquet File (Fastest)

import pandas as pd

# Load master Parquet file directly into Pandas
df = pd.read_parquet("crossplane5g_master_subflows.parquet")
print(f"Loaded {len(df):,} dual-plane subflow records.")
print(df['scenario_label'].value_counts())

Option 2: Recombine and Load All 12 CSV Part Files

import glob
import pandas as pd

# Find and load all CSV parts in numerical order
csv_files = sorted(glob.glob("crossplane5g_csv_part_*.csv"))
print(f"Found {len(csv_files)} CSV part files. Loading into memory...")

df_list = [pd.read_csv(f) for f in csv_files]
df_full = pd.concat(df_list, ignore_index=True)

print(f"Recombined full dataset: {len(df_full):,} rows across {len(df_full.columns)} columns.")

Option 3: Process CSV Parts Iteratively (Memory-Efficient / Batching)

import glob
import pandas as pd

for csv_part in sorted(glob.glob("crossplane5g_csv_part_*.csv")):
    print(f"Processing chunk: {csv_part}...")
    chunk_df = pd.read_csv(csv_part)
    # Perform feature transformations or batch training here

Option 4: Load Quick Stratified Sample (Lightweight Analysis)

import pandas as pd

# Load 100k sample for fast prototyping
df_sample = pd.read_csv("crossplane5g_sample_100k.csv")
print(f"Loaded sample dataset ({len(df_sample):,} rows).")

---

Hosting, Citation, and Distribution

• Zenodo Repository (DOI): https://doi.org/10.5281/zenodo.23165032
• HuggingFace Datasets: https://huggingface.co/datasets/CrossPlane5G/CrossPlane5G-Benchmark
• GitHub Benchmark Toolkit: https://github.com/TchallaAlbertKwame/CrossPlane5G

Citation

If you use CrossPlane-5G in your research, please cite our benchmark paper:

@article{tchalla2026crossplane,
  title={CrossPlane-5G: A Dual-Plane Telemetry Dataset and Benchmark for Cross-Layer Intrusion Detection in 5G Standalone and O-RAN Architectures},
  author={Tchalla, Albert Kwame and Quansah Junior, Anthony Jacklingo Kwame and Paolini, Emilio and Tchao, Eric Tutu and Valcarenghi, Luca and Agbemenu, Andrew Selasi},
  journal={Computer Networks},
  year={2026},
  doi={10.5281/zenodo.23165032}
}