UNSW-NB15 Network Intrusion Detection Dataset

The UNSW-NB15 dataset for network intrusion detection, provided with two evaluation protocols to enable fair comparison across the literature.

Why This Dataset Exists

Published results on UNSW-NB15 range from 85% to 99% accuracy — but the gap is almost entirely due to evaluation protocol differences, not model quality:

Evaluation Protocol	Typical Accuracy	Example Papers
Standard split (temporal, 175K/82K)	85–93%	Most published results
Random split (from full 2.28M records)	97–99.6%	FWIW, some deep learning papers

The random split achieves higher accuracy because:

1. 30% of records are duplicates — random splitting leaks near-identical flows into both train and test
2. No temporal shift — the standard split has temporal separation between training and testing periods

Both protocols are valid for different purposes:

• Standard split: Realistic deployment scenario (train on past, test on future)
• Random split: Maximum model comparison (controls for temporal shift)

Configurations

standard — Original Temporal Split

The official train/test CSV files from UNSW-NB15, containing 44 features (all features except id).

from datasets import load_dataset

ds = load_dataset("lacg030175/UNSW-NB15", "standard")
# ds["train"]: 175,341 rows
# ds["test"]:   82,332 rows

• Source: UNSW_NB15_training-set.csv and UNSW_NB15_testing-set.csv
• Used by most published papers for benchmarking
• Temporal separation between train and test periods

random — Deduplicated Random Split

Full dataset (2.28M records) with duplicates removed, split 90/10 with random_state=0 and stratified by label. Contains 49 features including IP addresses and ports.

from datasets import load_dataset

ds = load_dataset("lacg030175/UNSW-NB15", "random")
# ds["train"]: 1,425,833 rows
# ds["test"]:    158,426 rows

• Source: All four raw UNSW-NB15 CSV files via Mouwiya/UNSW-NB15
• Preprocessing: drop_duplicates() reduces 2,280,090 → 1,584,259 records
• Split: train_test_split(test_size=0.1, random_state=0, stratify=label)
• Comparable to FWIW evaluation protocol (Susskind et al., 2023)

Baseline Results

Model	Standard Split	Random Split
Random Forest	87.2%	99.6%
XGBoost	87.3%	99.4%
FWIW WNN (paper)	—	98.5%

Labels

• Binary (label): 0 = Normal, 1 = Attack
• Multi-class (attack_cat): Normal, Analysis, Backdoor, DoS, Exploits, Fuzzers, Generic, Reconnaissance, Shellcode, Worms

Class Distribution

Standard split (train):

• Normal: 56,000 (32%) | Attack: 119,341 (68%)

Random split (after dedup):

• Normal: 1,523,904 (96%) | Attack: 60,355 (4%)

Features

Both configs include flow-level network features:

Category	Features	Examples
Flow	6	dur, proto, state, service, sbytes, dbytes
Content	8	sttl, dttl, sloss, dloss, sload, dload, spkts, dpkts
Time	6	sjit, djit, sinpkt, dinpkt, tcprtt, synack
Additional	13	ct_srv_src, ct_dst_ltm, is_sm_ips_ports, ...
Generated	5+	trans_depth, res_bdy_len, swin, dwin, ...

The random config additionally includes: srcip, dstip, sport, dsport, Stime, Ltime.

Citation

If you use this dataset, please cite the original UNSW-NB15 paper:

@inproceedings{moustafa2015unswnb15,
  title={UNSW-NB15: A Comprehensive Data Set for Network Intrusion Detection Systems},
  author={Moustafa, Nour and Slay, Jill},
  booktitle={Military Communications and Information Systems Conference (MilCIS)},
  year={2015},
  organization={IEEE}
}

License

The original UNSW-NB15 dataset is provided under CC BY 4.0 by the University of New South Wales. This reformatted version preserves the original license.