desant-ai/LoginPagePhishingVsBenign-2026

Desant Phishing Screenshot Dataset — 2026-01-26

A dataset of malicious vs non-malicious login web page screenshots for training phishing detection models.

Built by Desant.ai for the Desant Anti Phishing Chrome extension.

---

Dataset Summary

This dataset contains 2,846 PNG screenshots of web pages, evenly split between malicious (phishing) and non-malicious (legitimate) login pages. The screenshots were captured using a distributed Docker-based pipeline that collects phishing URLs from multiple threat intelligence feeds, navigates to them with a headless browser, and classifies the resulting screenshots using OCR-based login form detection.

The dataset was used to fine-tune a CLIP RN50x64-based binary classifier vision model for binary phishing screenshot classification. The trained model is available at desant-ai/desant-phishing-inference.

---

Dataset Structure

screenshots_20260126_1504/
├── 1.1_url_to_screenshots/
│   ├── screenshots_MALICIOUS/        # 1,423 PNG files
│   ├── screenshots_NOT_MALICIOUS/    # 1,423 PNG files
└── README.md                         # This file

File Counts

Folder	Files	Description
1.1_url_to_screenshots/screenshots_MALICIOUS/	1,423	Phishing login page screenshots
1.1_url_to_screenshots/screenshots_NOT_MALICIOUS/	1,423	Legitimate login page screenshots
Total images	2,846

Image Resolutions

Resolution	Count	Class
1920 x 941	1,413	Malicious
960 x 401	10	Malicious
1920 x 802	1,423	Non-malicious

All images are 8-bit RGB PNG files, non-interlaced.

---

Classes

Label	Class ID	Count	Description
MALICIOUS	1	1,423	Screenshots of phishing login pages — web pages designed to steal user credentials by impersonating legitimate services
NON_MALICIOUS	0	1,423	Screenshots of legitimate login forms web pages

Non-Malicious File Breakdown

Prefix	Count	Description
login_non_malicious_login_*	828	Legitimate login pages (Google, Microsoft, banks, etc.)
non_malicious_login_*	432	Non-malicious pages identified during processing
unknown_non_malicious_login_*	156	Pages where form type was uncertain, manually verified as safe
error_non_malicious_login_*	7	Error/unavailable pages captured during screenshotting

---

Data Collection

Malicious Screenshots

Phishing URLs were collected from four threat intelligence sources using an 8-thread parallel downloader:

• PhishTank — Community-verified phishing URL database
• OpenPhish — Automated phishing detection feed
• URLhaus (Abuse.ch) — Malware and phishing URL tracker
• AlienVault OTX — Open Threat Exchange indicators

URLs were screenshotted using 8 parallel Docker workers running headless Chrome, with automatic cookie consent handling and page load timeouts. Post-processing included:

1. Duplicate removal — Perceptual hashing to remove visually identical screenshots
2. Useless screenshot removal — OCR-based detection and removal of blank/error pages
3. Login form detection — OCR-based classification to identify pages containing login forms

Non-Malicious Screenshots

Legitimate login page URLs were sourced from a curated list of 500,000 known non-malicious login forms. The same screenshotting pipeline was used with identical settings. From the initial processing of 6,016 pages, 2,116 were identified as login forms, and the final set was quality-filtered to remove blank images, low-content pages, and processing errors.

Collection Date

January 26, 2026

---

Intended Use

This dataset is intended for:

• Training phishing detection models based on visual features of web pages
• Evaluating screenshot-based phishing classifiers
• Research in cybersecurity, computer vision, and adversarial machine learning
• Benchmarking vision models for security applications

Associated Model

This dataset was used to train the desant-ai/desant-phishing-inference model, a CLIP RN50x64-based binary classifier with a custom 3-layer MLP classifier head achieving 92–96% accuracy on phishing detection.

---

Limitations

• Language bias: The dataset is predominantly composed of English-language web pages. Phishing pages in non-Latin scripts are underrepresented.
• Login-form focus: The dataset is focused on login-page phishing. Other phishing types (fake downloads, payment scams, etc.) are not well represented.
• Temporal decay: Phishing pages are ephemeral. The URLs in this dataset may no longer be active, and phishing techniques evolve over time.
• Resolution variance: Malicious and non-malicious screenshots were captured at slightly different viewport sizes (941px vs 802px height), which could introduce a resolution-based shortcut for classifiers if not addressed during preprocessing.

---

Ethical Considerations

• This dataset is provided for defensive cybersecurity research only.
• The screenshots may contain brand logos and trademarks of impersonated companies (Google, Microsoft, banks, etc.). These are included solely for research purposes.
• Do not use this dataset to create, improve, or deploy phishing attacks.
• URLs and sensitive content have been removed; only the screenshot images remain.

---

Citation

@dataset{desant2026phishing_screenshots,
  title={Desant Phishing Screenshot Dataset},
  author={Desant.ai},
  year={2026},
  url={https://huggingface.co/datasets/desant-ai/desant-phishing-screenshots},
  license={CC-BY-4.0}
}

---

Contact

• Company: Desant.ai
• Model: desant-ai/desant-phishing-inference
• Demo: Hugging Face Space
• Chrome extension: Desant Anti Phishing Chrome extension.