---
license: mit
task_categories:
- text-classification
- image-classification
language:
- en
size_categories:
- 1M<n<10M
configs:
- config_name: images
  data_files:
  - split: train
    path: images/train-*
  - split: validation
    path: images/validation-*
  - split: test
    path: images/test-*
- config_name: pairs
  data_files:
  - split: train
    path: pairs/train-*
  - split: validation
    path: pairs/validation-*
  - split: test
    path: pairs/test-*
dataset_info:
- config_name: images
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': real
          '1': phishx
  - name: text
    dtype: string
  - name: pair_id
    dtype: int32
  splits:
  - name: train
    num_bytes: 2660477680
    num_examples: 1799810
  - name: validation
    num_bytes: 760218759
    num_examples: 514232
  - name: test
    num_bytes: 379976401
    num_examples: 257116
  download_size: 3776742555
  dataset_size: 3800672840
- config_name: pairs
  features:
  - name: domain
    dtype: large_string
  - name: homoglyphs
    dtype: large_string
  - name: pair_id
    dtype: int32
  splits:
  - name: train
    num_bytes: 49094585
    num_examples: 899905
  - name: validation
    num_bytes: 14028004
    num_examples: 257116
  - name: test
    num_bytes: 7010618
    num_examples: 128558
  download_size: 40293542
  dataset_size: 70133207
tags:
- cybersecurity
- domain-spoofing
- homoglyph
- idn-homograph
- phishing
- security
pretty_name: GlyphNet Homoglyph Domains
---
# GlyphNet: Homoglyph Domains Dataset

Data for detecting [homoglyph](https://en.wikipedia.org/wiki/IDN_homograph_attack)
phishing domains (e.g. `facebook.com` spoofed with visually-similar Unicode
characters). Every genuine domain is paired with a synthetically generated
homoglyph variant, and each domain is also rendered to a 256x256 grayscale image
so the task can be tackled as text **or** image classification.

Paper: [arXiv:2306.10392](https://arxiv.org/abs/2306.10392) ·
Code: [github.com/Akshat4112/Glyphnet](https://github.com/Akshat4112/Glyphnet)

## Configs & splits

Both configs share the same `train` / `validation` / `test` split (70/20/10).
Splits are assigned **per pair** (seeded), so a genuine domain and its homoglyph
always fall in the same split - preventing leakage between near-identical pairs.

| Config | Rows | Fields |
|--------|------|--------|
| `pairs`  | 1,285,579 pairs | `domain` (genuine), `homoglyphs` (spoofed), `pair_id` |
| `images` | 2,571,158 images | `image` (256x256 grayscale PNG), `label` (`real`/`phish`), `text`, `pair_id` |

`pair_id` links the two `images` rows (one `real`, one `phish`) that came from
the same source pair, and matches the `pair_id` in the `pairs` config.

```python
from datasets import load_dataset

pairs  = load_dataset("Akshat4112/Glyphnet", "pairs")
images = load_dataset("Akshat4112/Glyphnet", "images")
train_img = images["train"]        # or "validation" / "test"
```

## How it was generated

Homoglyphs are produced by substituting one or two characters with Unicode
confusables (`code/dataGeneration.py`). Images are rendered with the **DejaVu
Sans** font - Arial (used in the paper) is proprietary and not redistributed
here, so glyph shapes may differ slightly from the original figures.

## Intended uses & limitations

- Intended for research on homoglyph / IDN-homograph phishing detection.
- Homoglyphs are **synthetic**, not harvested from real attacks, so the
  distribution may not match live phishing in the wild.
- The genuine and spoofed strings within a pair are near-identical; always
  respect the provided splits (or group by `pair_id`) to avoid leakage.
- Rendering font differs from the paper (DejaVu Sans vs Arial).

## License

MIT (same as the source repository).

## Citation

If you use this dataset in your research, please cite the GlyphNet paper:

```bibtex
@article{gupta2023glyphnet,
  title   = {GlyphNet: Homoglyph domains dataset and detection using attention-based Convolutional Neural Networks},
  author  = {Gupta, Akshat and Tomar, Laxman Singh and Garg, Ridhima},
  journal = {arXiv preprint arXiv:2306.10392},
  year    = {2023}
}
```

> Gupta, A., Tomar, L. S., & Garg, R. (2023). *GlyphNet: Homoglyph domains
> dataset and detection using attention-based Convolutional Neural Networks.*
> arXiv:2306.10392.