---
license: cc-by-4.0
task_categories:
- text-classification
- feature-extraction
- tabular-classification
language:
- sq
- th
- it
- tl
- de
- hr
- fi
- da
- lv
- pl
- ca
- ro
- ja
- lt
- af
- ru
- so
- en
- id
- cs
- sw
- es
- sl
- hu
- ko
- nl
- pt
- tr
- sv
- sk
- cy
- bg
- fr
- et
- no
- vi
- multilingual
size_categories:
- 1M<n<10M
configs:
- config_name: default
  data_files:
  - split: train
    pattern:
    - "phishing_features_train.csv"
    - "phishing_url_train0.csv"
  - split: test
    pattern:
    - "phishing_features_val.csv"
    - "phishing_url_val0.csv"
---
**I have decided to release the auto-moderation models all at once sometime in July, 2023. The datasets for training these models will be avaliable first.**

*Upload not complete...*

The *features* dataset is original, and my feature extraction method is covered in [feature_extraction.py](./feature_extraction.py)

In the *features* dataset, there're 770,098 websites online at the time of data collection. The plots below show the regression line and correlation coefficients of 20+ features extracted and whether the URL is malicious.
If we could plot the lifespan of URLs, we could see that the oldest website has been online since Nov 7th, 2008, while the most recent phishing websites appeared as late as July 10th, 2023.

As we can see, there's no correlation between `is_malicious` and the columns `meta_percentage`, `mouseover_changes`, `not_indexed_by_google`, `right_click_disabled`, and `popup_window_has_text_field` as of July, 2023, contrary to some [analyses of researchers in 2013 on phishing detection](./Phishing_Websites_Features.docx).
The majority of features have very weak correlations with `is_malicious`, while a minority has a weak correlation. Is this problematic for training? Not really.

I split the classification task into two stages in anticipation of the limited availability of online phishing websites due to their short lifespan, as well as the possibility that research done on phishing is not up-to-date:
1. a small multilingual BERT model to output the confidence level of a URL being malicious to model #2, by finetuning on 2,154,175 legitimate and malicious URLs
2. (probably) LightGBM to analyze the confidence level, along with roughly 19 extracted features

This way, I can make the most out of the limited phishing websites avaliable.

![Phish Eater Data Analysis](https://i.imgur.com/ADh6luR.png)


## Source of the URLs
- https://phishtank.org/
- https://moz.com/top500
- https://www.kaggle.com/datasets/siddharthkumar25/malicious-and-benign-urls
- https://www.kaggle.com/datasets/sid321axn/malicious-urls-dataset
- https://github.com/ESDAUNG/PhishDataset
- https://github.com/JPCERTCC/phishurl-list
- https://github.com/Dogino/Discord-Phishing-URLs

## Reference
- https://www.kaggle.com/datasets/akashkr/phishing-website-dataset
- https://www.kaggle.com/datasets/shashwatwork/web-page-phishing-detection-dataset
- https://www.kaggle.com/datasets/aman9d/phishing-data

## Side notes
- Cloudflare offers an [API for URL scanning](https://developers.cloudflare.com/api/operations/phishing-url-information-get-results-for-a-url-scan), with a generous global rate limit of 1200 requests every 5 minutes.