snsec-net/dga-detection-drift26dsn

DRIFT Dataset

Longitudinal Benign and DGA Domain Name Dataset (2017–2025)

• Authors: Chaeyoung Lee*, Chaeri Jung*, Seonghoon Jeong (* Equal contribution)
• Affiliation: Division of Artificial Intelligence Engineering, Sookmyung Women's University
• Venue: IEEE/IFIP International Conference on Dependable Systems and Networks (DSN 2026) — accepted, IEEE Xplore link coming soon
• arXiv: arxiv.org/abs/2605.10436
• GitHub: snsec-net/2026-DSN-DRIFT
• Hugging Face: Paper · Dataset · Model
• IEEE DataPort: 10.21227/za2s-9e09

---

Description

This dataset was curated and used in the paper "DRIFT: Drift-Resilient Invariant-Feature Transformer for DGA Detection", accepted at IEEE/IFIP DSN 2026.

This dataset provides a nine-year longitudinal collection (2017–2025) of benign and DGA (Domain Generation Algorithm) domain names, designed to support evaluation of DGA detectors under real-world concept drift. Most existing DGA datasets are constructed from a single time period, which masks how quickly detection models degrade when exposed to evolving domain distributions. This dataset addresses that gap by providing temporally aligned benign and malicious domains, enabling forward-chaining experiments that simulate realistic deployment conditions.

	Count	Source
Benign domains	~49.4M unique	Alexa Top 1M + Tranco Top 1M
DGA domains	~149.4M unique	DGArchive · 148 families (133 char-based, 15 word-based)
Total	~198.8M unique

Cumulative Dataset Statistics (from Table I of the paper)

Period	Benign (Total)	Benign (Unique)	DGA (Total)	DGA (Unique)	Unique DGA Families
2017	1,913,418	1,913,418	14,888,780	14,888,780	58
2017–2018	18,729,906	17,129,997	30,004,459	29,305,992	62
2017–2019	40,508,820	29,359,365	46,423,032	44,940,295	65
2017–2020	58,149,929	36,796,092	65,116,780	62,272,656	71
2017–2021	72,679,722	41,856,865	84,661,350	79,752,910	77
2017–2022	88,262,644	47,358,571	104,424,484	97,107,537	80
2017–2023	92,962,312	48,190,058	124,186,047	114,189,860	81
2017–2024	94,833,375	48,805,692	144,981,221	132,177,150	148
2017–2025	96,849,575	49,433,110	165,824,441	149,405,584	148

Note: counts above are cumulative across years (as reported in the paper). The per-year parquet files in this repository are the per-period snapshots used as model input.

---

Data Collection

Benign domains are sourced from Alexa Top 1M and Tranco Top 1M. Historical Alexa snapshots were retrieved via the Internet Archive Wayback Machine, and historical Tranco lists were obtained through the Tranco API. Yearly snapshots were collected from 2017 to 2025. Domains appearing in both the benign and DGA sets were removed to prevent cross-contamination.

DGA domains are sourced from DGArchive, maintained by Fraunhofer FKIE. DGArchive provides deterministic domain outputs derived directly from reverse-engineered malware algorithms and seeds, along with per-domain timestamps. 148 out of 151 available families were selected based on suitability for longitudinal analysis, covering both character-based (133 families) and word-based (15 families) generation schemes.

Preprocessing:

• All domain names are lowercased
• Effective second-level domains (eSLDs) are extracted by stripping TLD and ccTLD suffixes
• Deduplication is performed on effective eSLDs
• Characters are restricted to alphanumerics, hyphens (-), and dots (.) per IETF RFC 1035

---

Repository Layout

dga-detection-drift26dsn/
├── DRIFT_input_eSLD/         # eSLD, 72 files: T{YY}_{benign,dga}[_{train,val,test}].parquet
│   ├── T17_benign.parquet                    # Full 2017 benign (eSLD)
│   ├── T17_benign_train.parquet              # 1,500,000 samples
│   ├── T17_benign_val.parquet                # 150,000 samples
│   ├── T17_benign_test.parquet               # Remaining samples
│   ├── T17_dga.parquet                       # Full 2017 DGA (eSLD)
│   ├── T17_dga_train.parquet
│   ├── T17_dga_val.parquet
│   ├── T17_dga_test.parquet
│   ├── T18_benign.parquet
│   ├── ... (same pattern for T18–T25)
│   └── T25_dga_val.parquet
└── raw_including_TLD/        # raw + family, 18 files: T{YY}_{benign,dga}.parquet
    ├── T17_benign.parquet                    # Full 2017 benign (with TLD)
    ├── T17_dga.parquet                       # Full 2017 DGA (with TLD + family)
    ├── T18_benign.parquet
    ├── ... (same pattern for T18–T25)
    └── T25_dga.parquet

File naming convention: T{YY}_{class}[_{split}].parquet

• YY: two-digit year (17 = 2017 … 25 = 2025)
• class: benign or dga
• split (optional): train, val, or test — omitted for the full yearly file

Split sizes (DRIFT_input_eSLD only):

• train: 1,500,000 samples per class per year
• val: 150,000 samples per class per year (10% of train)
• test: remaining samples (varies by year and class)

The train size was determined by the smallest yearly class file across all years, ensuring no class/year combination runs out of samples. The full T{YY}_{class}.parquet files are the union of train + val + test for that year and class.

raw_including_TLD/ provides only full yearly files (no pre-split), as it is intended for reference and custom preprocessing rather than direct model input.

---

Format

*_eSLD configs (DRIFT_input_eSLD/)

All files are *.parquet with the following columns:

Column	Type	Description
domain	string	Effective second-level domain (eSLD), TLD and ccTLD stripped (e.g.google)
label	int	0 = benign, 1 = DGA

Example rows:

domain,label
google,0
vutotoid,1
runtime-incorrect,1

*_raw configs (raw_including_TLD/)

All files are *.parquet with the following columns:

Column	Type	Description
domain	string	Raw domain name including TLD (e.g.google.com)
label	int	0 = benign, 1 = DGA
family	string	DGA family name (e.g.virut, qsnatch); "benign" for benign domains

Example rows:

domain,label,family
google.com,0,benign
sbhya.com,1,virut
w0gv.gl,1,qsnatch

---

Quick Start

Usage

This repository exposes two representations, each split into per-year configs (T17–T25):

• T{YY}_eSLD — effective second-level domain only (TLD/ccTLD stripped). Benign + DGA are merged into ready-to-train train / validation / test splits; the label column distinguishes the class. This is the direct DRIFT model input.
• T{YY}_raw — raw domain (including TLD) plus the DGA family label. Provided as benign / dga splits for reference and custom preprocessing.

from datasets import load_dataset

# eSLD model input — load one within-year partition of 2020 (benign + DGA, with label).
# NOTE: `split` here selects a packaged file partition, NOT a forward-chaining role —
# see the Recommended Evaluation Protocol below for how to assemble train/test by year.
drift_input = load_dataset("snsec-net/dga-detection-drift26dsn", "T20_eSLD", split="test")

# raw, with TLD + DGA family — e.g. 2017 DGA domains
raw = load_dataset("snsec-net/dga-detection-drift26dsn", "T17_raw", split="dga")

Recommended Evaluation Protocol (Forward-Chaining)

DRIFT measures temporal robustness, so the train/test distinction is defined by time, not by the split names.

Important — what the split names mean. Each yearly config ships train / validation / test splits, but these are just within-year partitions of that one year's data (an artifact of how the files were packaged); they are not conventional ML train/test roles. The forward-chaining role of any data is decided by which year it belongs to relative to your cutoff — so a training year's test partition is still training data, and a test year's train partition is still test data.

The protocol, stated temporally:

1. Stand at a point in time (in the paper, the end of 2019).
2. Train on all data strictly before that point (2017–2019), holding out a small slice per year for validation / early stopping.
3. Freeze the model, then evaluate it year by year on each later year (2020, 2021, …, 2025) — each year tested independently, so accuracy degradation reveals drift over time.

Mapping that onto the packaged splits (full year = train + validation + test):

Forward-chaining role	Years	What to load
Training data	2017–2019	full yearminus the val holdout → train + test partitions
Validation (early stopping / model selection — the one real ML val set)	2017–2019	the validation partition (150k benign + 150k DGA per year)
Test (frozen model, scored per year)	each of 2020–2025	thewhole year → train + validation + test partitions

So do not read split="train" as "the training set" or split="test" as "the test set": for a training year you want train+test (the full year minus the val holdout), and for a later year you want the entire year. Using split="train"/split="test" literally would mix the temporal roles and silently drop most of the data — e.g. split="train" is capped at 1.5M/class, discarding the test partition that holds the bulk of each year (for 2017 DGA, 13.2M of 14.9M domains live in test).

from datasets import load_dataset, concatenate_datasets

REPO = "snsec-net/dga-detection-drift26dsn"
TRAIN_YEARS = ["17", "18", "19"]

# Training set: FULL 2017–2019 data minus the 150k/class held-out validation.
# full = train + validation + test  =>  full - validation = train + test
train = concatenate_datasets([
    load_dataset(REPO, f"T{y}_eSLD", split="train+test") for y in TRAIN_YEARS
])

# Held-out validation: 150k benign + 150k DGA per training year
val = concatenate_datasets([
    load_dataset(REPO, f"T{y}_eSLD", split="validation") for y in TRAIN_YEARS
])

# Test on a strictly-newer year — the FULL year (2020–2025 are never used for training)
test_2020 = load_dataset(REPO, "T20_eSLD", split="train+validation+test")

Not running forward-chaining? If you are not running a longitudinal (forward-chaining) evaluation — e.g. you are training and testing within a single year, or pooling several years together as one ordinary dataset — you can use the train / validation / test partitions at face value. The temporal caveat above applies only when a year's role is decided by its position relative to a cutoff.

---

Citation

If you use this dataset, please cite:

Dataset:

C. Lee, C. Jung, and S. Jeong, "Longitudinal Benign and DGA Domain Name Dataset,"
IEEE DataPort, 2026. doi: 10.21227/za2s-9e09.
Available: https://dx.doi.org/10.21227/za2s-9e09

@misc{lee2026driftdata,
  author       = {Lee, Chaeyoung and Jung, Chaeri and Jeong, Seonghoon},
  title        = {Longitudinal Benign and {DGA} Domain Name Dataset},
  howpublished = {IEEE Dataport},
  year         = {2026},
  doi          = {10.21227/za2s-9e09},
  url          = {https://dx.doi.org/10.21227/za2s-9e09}
}

Paper:

C. Lee, C. Jung, and S. Jeong, "DRIFT: Drift-Resilient Invariant-Feature Transformer
for DGA Detection," in Proc. IEEE/IFIP DSN, 2026. [to be updated with DOI upon publication]

@inproceedings{lee2026drift,
  title     = {{DRIFT}: Drift-Resilient Invariant-Feature Transformer for {DGA} Detection},
  author    = {Lee, Chaeyoung and Jung, Chaeri and Jeong, Seonghoon},
  booktitle = {Proc. IEEE/IFIP International Conference on Dependable Systems and Networks (DSN)},
  year      = {2026}
}

---

Intended Use & Ethical Considerations

This dataset is released for cybersecurity research — building and benchmarking DGA detectors, studying concept drift, and improving network defenses. It must not be used to operate, register, or distribute malicious domains. Use is restricted to non-commercial research and personal use per the license below.

---

License

DGArchive is governed by the CC BY-NC-SA 3.0 license. Therefore, this dataset may not be used for commercial purposes, but is available for research or personal use. If this license does not meet your needs, we are open to individual licensing agreements.

https://creativecommons.org/licenses/by-nc-sa/3.0/

---

References

[1] D. Plohmann, K. Yakdan, M. Klatt, J. Bader, and E. Gerhards-Padilla, "A comprehensive measurement study of domain generating malware," in Proc. USENIX Security, 2016.

[2] V. Le Pochat, T. Van Goethem, S. Tajalizadehkhoob, M. Korczynski, and W. Joosen, "Tranco: A research-oriented top sites ranking hardened against manipulation," in Proc. NDSS, 2019.