PyDomains: Classifying the Content of Domains

This repository is no longer actively maintained. Check out: https://github.com/themains/piedomains

The package provides two broad ways of learning about the kind of content hosted on a domain. First, it provides convenient access to curated lists of domain content like the Shallalist, DMOZ, PhishTank, and such. Second, it exposes models built on top of these large labeled datasets; the models estimate the relationship between sequence of characters in the domain name and the kind of content hosted by the domain.

Quick Start

import pandas as pd
from pydomains import *

# Get help
help(dmoz_cat)

# Load data
df = pd.read_csv('./pydomains/examples/input-header.csv')

#  df
#       label                                url
#   0   test1                        topshop.com
#   1   test2                   beyondrelief.com

# Get the Content Category from DMOZ, phishtank
df_dmoz  = dmoz_cat(df, domain_names = 'url')
df_phish = phish_cat(df, domain_names = 'url')

# Predicted category from shallalist, toulouse
df_shalla   = pred_shalla(df, domain_names = 'url')
df_toulouse = pred_toulouse(df, domain_names = 'url')

Installation

Installation is as easy as typing in:

pip install pydomains

API

1. dmoz_cat, shalla_cat, and phish_cat: When the domain is in the DMOZ, Shallalist, and Phishtank data, the functions give the category of the domain according to the respective list. (Phishtank just gives whether or not the domain has been implicated in phishing.) Otherwise, the function returns an empty string.

   • Arguments:

     • df: pandas dataframe. No default.
     • domain_names: column with the domain names/URLs. Default is domain_names
     • year. Specify the year from which you want to use the data. Currently only DMOZ data from 2016, and Shallalist and Phishtank data from 2017 is available.
     • latest. Boolean. Default is False. If True, the function checks if a local file exists and if it exists, if the local file is the latest. If it isn't, it downloads the latest file from the GitHub link and overwrites the local file.

   • Functionality:

     • converts domain name to lower case, strips http://.
     • Looks for dmoz_YYYY.csv, shalla_YYYY.csv, or phish_YYYY.csv respectively in the local folder. If it doesn't find it, it downloads the latest DMOZ, Shallalist, or Phishtank file from pydomains/data/dmoz_YYYY.csv.bz2, pydomains/data/shalla_YYYY.csv.bz2, or pydomains/data/phish_YYYY.csv.bz2respectively.
     • If the latest flag is planted, it checks if the local file is older than the remote file. If it is, it overwrites the local file with the newer remote file.

   • Output:

     • Appends the category to the CSV. By default it creates a column (dmoz_year_cat or shalla_year_cat or phish_year_cat).
     • If no match is found, it returns nothing.
     • DMOZ sometimes has multiple categories per domain. The categories are appended together with a semi-colon.

   • Examples:

     import pandas as pd
     
     df = pd.DataFrame([{'domain_names': 'http://www.google.com'}])
     
     dmoz_cat(df)
     shalla_cat(df)
     phish_cat(df)
     

2. pred_shalla: We use data from Shallalist to train a LSTM model. The function uses the trained model to predict the category of the domain based on the domain name.

   • Arguments:

     • df: pandas dataframe. No default.
     • domain_names: column with the domain names/URLs. Default is domain_names
     • year. Year of the model. Default is 2017. Currently only a model based on data from 2017 is available.
     • latest. Boolean. Default is False. If True, the function checks if a local model file exists and if it exists, is it older than what is on the website. If it isn't, it downloads the latest file from the GitHub link and overwrites the local file.

   • Functionality:

     • converts domain name to lower case, strips http://.
     • Uses the model to predict the probability of content being from various categories.

   • Output

     • Appends a column carrying the label of the category with the highest probability (pred_shalla_year_lab) and a series of columns with probabilities for each category (pred_shalla_year_prob_catname).

   • Examples:

     pred_shalla(df)
     

3. pred_toulouse: We use data from http://dsi.ut-capitole.fr/blacklists/ to train a LSTM model that predicts the category of content hosted by the domain. The function uses the trained model to predict the category of the domain based on the domain name.

   • Arguments:

     • df: pandas dataframe. No default.
     • domain_names: column with the domain names/URLs. Default is domain_names
     • year. Year of the model. Default is 2017. Currently only a model based on data from 2017 is available.
     • latest. Boolean. Default is False. If True, the function checks if a local model file exists and if it exists, is it older than what is on the website. If it isn't, it downloads the latest file from the GitHub link and overwrites the local file.

   • Functionality:

     • converts domain name to lower case, strips http://.
     • Uses the model to predict the probability of it being a domain implicated in distributing malware.

   • Output:

     • Appends a column carrying the label of the category with the highest probability (pred_toulouse_year_lab) and a series of columns with probabilities for each category (pred_toulouse_year_prob_catname).

   • Examples:

     pred_malware(df)
     

4. pred_phish: Given the importance, we devote special care to try to predict domains involved in phishing well. To do that, we use data from PhishTank and combine it with data from http://s3.amazonaws.com/alexa-static/top-1m.csv.zip, and train a LSTM model. The function gives the predicted probability based on the LSTM model.

   • Arguments:

     • df: pandas dataframe. No default.
     • domain_names: column with the domain names/URLs. Default is domain_names
     • year. Year of the model. Default is 2017. Currently only a model based on data from 2017 is available.
     • latest. Boolean. Default is False. If True, the function checks if a local model file exists and if it exists, is it older than what is on the website. If it isn't, it downloads the latest file from the GitHub link and overwrites the local file.

   • Functionality:

     • converts domain name to lower case, strips http://.
     • Uses the model to predict the probability of it being a domain implicated in phishing.

   • Output:

     • Appends column pred_phish_year_lab which contains the most probable label, and a column indicating the probability that the domain is involved in distributing malware (pred_phish_year_prob).

   • Examples:

     pred_phish(df)
     

5. pred_malware: Once again, given the importance of flagging domains that carry malware, we again devote extra care to try to predict domains involved in distributing malware well. We combine data on malware domains http://mirror1.malwaredomains.com/ with data from http://s3.amazonaws.com/alexa-static/top-1m.csv.zip, and train a LSTM model. The function gives the predicted probability based on the LSTM model.

   • Arguments:

     • df: pandas dataframe. No default.
     • domain_names: column with the domain names/URLs. Default is domain_names
     • year. Year of the model. Default is 2017. Currently only a model based on data from 2017 is available.
     • latest. Boolean. Default is False. If True, the function checks if a local model file exists and if it exists, is it older than what is on the website. If it isn't, it downloads the latest file from the GitHub link and overwrites the local file.

   • Functionality:

     • converts domain name to lower case, strips http://.
     • Uses the model to predict the probability of it being a domain implicated in distributing malware.

   • Output:

     • Appends column pred_malware_year_lab and a column indicating the probability that the domain is involved in distributing malware (pred_malware_year_prob).

   • Examples:

     pred_malware(df)
     

Using pydomains

>>> import pandas as pd
>>> from pydomains import *
Using TensorFlow backend.

>>> # Get help of the function
... help(dmoz_cat)
Help on function dmoz_cat in module pydomains.dmoz_cat:

dmoz_cat(df, domain_names='domain_names', year=2016, latest=False)
    Appends DMOZ domain categories to the DataFrame.

    The function extracts the domain name along with the subdomain
    from the specified column and appends the category (dmoz_cat)
    to the DataFrame. If DMOZ file is not available locally or
    latest is set to True, it downloads the file. The function
    looks for category of the domain name in the DMOZ file
    for each domain. When no match is found, it returns an
    empty string.

    Args:
        df (:obj:`DataFrame`): Pandas DataFrame. No default value.
        domain_names (str): Column name of the domain in DataFrame.
            Default in `domain_names`.
        year (int): DMOZ data year. Only 2016 data is available.
            Default is 2016.
        latest (Boolean): Whether or not to download latest
            data available from GitHub. Default is False.

    Returns:
        DataFrame: Pandas DataFrame with two additional columns:
            'dmoz_year_domain' and 'dmoz_year_cat'


>>> # Load an example input with columns header
... df = pd.read_csv('./pydomains/examples/input-header.csv')

>>> df
    label                                url
0   test1                        topshop.com
1   test2                   beyondrelief.com
2   test3                golf-tours.com/test
3   test4                    thegayhotel.com
4   test5  https://zonasequravlabcp.com/bcp/
5   test6                http://privatix.xyz
6   test7              adultfriendfinder.com
7   test8            giftregistrylocator.com
8   test9                 bangbrosonline.com
9  test10                scotland-info.co.uk

>>> # Get the Content Category from DMOZ
... df = dmoz_cat(df, domain_names='url')
Loading DMOZ data file...

>>> df
    label                                url         dmoz_2016_domain  \
0   test1                        topshop.com              topshop.com
1   test2                   beyondrelief.com         beyondrelief.com
2   test3                golf-tours.com/test           golf-tours.com
3   test4                    thegayhotel.com          thegayhotel.com
4   test5  https://zonasequravlabcp.com/bcp/     zonasequravlabcp.com
5   test6                http://privatix.xyz             privatix.xyz
6   test7              adultfriendfinder.com    adultfriendfinder.com
7   test8            giftregistrylocator.com  giftregistrylocator.com
8   test9                 bangbrosonline.com       bangbrosonline.com
9  test10                scotland-info.co.uk      scotland-info.co.uk

                                    dmoz_2016_cat
0  Top/Regional/Europe/United_Kingdom/Business_an...
1                                                NaN
2                                                NaN
3                                                NaN
4                                                NaN
5                                                NaN
6                                                NaN
7                                                NaN
8                                                NaN
9  Top/Regional/Europe/United_Kingdom/Scotland/Tr...
>>> # Predict Content Category Using the Toulouse Model
... df = pred_toulouse(df, domain_names='url')
Loading Toulouse model, vocab and names data file...

>>> df
    label                                url         dmoz_2016_domain  \
0   test1                        topshop.com              topshop.com
1   test2                   beyondrelief.com         beyondrelief.com
2   test3                golf-tours.com/test           golf-tours.com
3   test4                    thegayhotel.com          thegayhotel.com
4   test5  https://zonasequravlabcp.com/bcp/     zonasequravlabcp.com
5   test6                http://privatix.xyz             privatix.xyz
6   test7              adultfriendfinder.com    adultfriendfinder.com
7   test8            giftregistrylocator.com  giftregistrylocator.com
8   test9                 bangbrosonline.com       bangbrosonline.com
9  test10                scotland-info.co.uk      scotland-info.co.uk

                                    dmoz_2016_cat  \
0  Top/Regional/Europe/United_Kingdom/Business_an...
1                                                NaN
2                                                NaN
3                                                NaN
4                                                NaN
5                                                NaN
6                                                NaN
7                                                NaN
8                                                NaN
9  Top/Regional/Europe/United_Kingdom/Scotland/Tr...

pred_toulouse_2017_domain pred_toulouse_2017_lab  \
0               topshop.com               shopping
1          beyondrelief.com                  adult
2            golf-tours.com               shopping
3           thegayhotel.com                  adult
4      zonasequravlabcp.com               phishing
5              privatix.xyz                  adult
6     adultfriendfinder.com                  adult
7   giftregistrylocator.com               shopping
8        bangbrosonline.com                  adult
9       scotland-info.co.uk               shopping

pred_toulouse_2017_prob_adult  pred_toulouse_2017_prob_audio-video  \
0                       0.133953                             0.003793
1                       0.521590                             0.016359
2                       0.186083                             0.008208
3                       0.971451                             0.001080
4                       0.065503                             0.001063
5                       0.986328                             0.002241
6                       0.939441                             0.000211
7                       0.014645                             0.000570
8                       0.945490                             0.004017
9                       0.256270                             0.003745

pred_toulouse_2017_prob_bank  pred_toulouse_2017_prob_gambling  \
0                  1.161209e-04                      2.911613e-04
1                  3.912278e-03                      6.484169e-03
2                  1.783388e-03                      8.022175e-04
3                  8.920387e-05                      6.256429e-05
4                  6.226773e-04                      1.073759e-04
5                  6.823016e-07                      1.969112e-06
6                  1.742063e-07                      6.485808e-08
7                  3.973934e-04                      1.019526e-05
8                  9.122109e-05                      1.142884e-04
9                  3.962536e-04                      4.977396e-04

pred_toulouse_2017_prob_games  pred_toulouse_2017_prob_malware  \
0                       0.002073                         0.003976
1                       0.022408                         0.018371
2                       0.013352                         0.006392
3                       0.000713                         0.000934
4                       0.012431                         0.077391
5                       0.001021                         0.004949
6                       0.000044                         0.000059
7                       0.004112                         0.016339
8                       0.002216                         0.000422
9                       0.014452                         0.006615

pred_toulouse_2017_prob_others  pred_toulouse_2017_prob_phishing  \
0                        0.014862                          0.112132
1                        0.046011                          0.172208
2                        0.021287                          0.060633
3                        0.005018                          0.017201
4                        0.031691                          0.416989
5                        0.003069                          0.002094
6                        0.001674                          0.058497
7                        0.015631                          0.131174
8                        0.017964                          0.012574
9                        0.057622                          0.111698

pred_toulouse_2017_prob_press  pred_toulouse_2017_prob_publicite  \
0                   8.404775e-04                           0.000761
1                   2.525988e-02                           0.002821
2                   1.853482e-02                           0.000990
3                   2.208834e-04                           0.000135
4                   2.796387e-03                           0.000284
5                   4.559151e-06                           0.000252
6                   1.133891e-07                           0.000007
7                   1.115335e-02                           0.000436
8                   5.098383e-04                           0.000785
9                   7.331154e-04                           0.000168

pred_toulouse_2017_prob_shopping
0                          0.727203
1                          0.164577
2                          0.681934
3                          0.003094
4                          0.391121
5                          0.000038
6                          0.000066
7                          0.805531
8                          0.015817
9                          0.547802

Models

For more information about the models, including the decisions we made around curtailing the number of categories, see here

For model performance and comparison to Random Forest and SVC models, see the relevant notebooks and this folder with eps images of the ROC. We also checked if the probabilities were calibrated. We find LSTM to be pretty well calibrated. The notebooks are posted here

Underlying Data

We use data from DMOZ, Shallalist, PhishTank, and a prominent Blacklist aggregator. For more details about how the underlying data, see here

Validation

We compare content categories according to the TrustedSource API with content category from Shallalist and the Shallalist model for all the unique domains in the comScore 2004 data:

1. comScore 2004 Trusted API results
2. comScore 2004 categories from pydomains
3. comparison between TrustedSource and Shallalist and shallalist model

Application

We use the package to answer two questions:

• Do poor people, minorities, and the less-well-educated visit sites that distribute malware or engage in phishing more frequently than their respective complementary groups---the better-off, the racial majority, the better educated?
• How does consumption of pornography vary by education and age?

See the repository for the application.

Paper

For more details about the performance and for citation, see the paper.

comScore Domain Data Categories

To make it easier to learn browsing behavior of people, we obtained the type of content hosted by a domain using all the functions in pydomains for all the unique domains in all the comScore data from 2002 to 2016 (there are some missing years). We have posted the data here

Notes and Caveats

• The DMOZ categorization system at tier 1 is bad. The category names are vague. They have a lot of subcategories that could easily belong to other tier 1 categories. That means a) it would likely be hard to classify well at tier 1 and b) not very valuable. So we choose not to predict tier 1 DMOZ categories.
• The association between patterns in domain names and the kind of content they host may change over time. It may change as new domains come online and as older domains are repurposed. All this likely happens slowly. But, to be careful, we add a year variable in our functions. Each list and each model is for a particular year.
• Imputing the kind of content hosted by a domain may suggest to some that domains carry only one kind of content. Many domains don't. And even when they do, the quality varies immensely. (See more here.) There is much less heterogeneity at the URL level. And we plan to look into predicting at URL level. See TODO for our plans.
• There are a lot of categories where we do not expect domain names to have any systematic patterns. Rather than make noisy predictions using just the domain names (the data that our current set of classifiers use), we plan to tackle this prediction task with some additional data. See TODO for our plans.

Documentation

For more information, please see project documentation.

Authors

Suriyan Laohaprapanon and Gaurav Sood

Contributor Code of Conduct

The project welcomes contributions from everyone! In fact, it depends on it. To maintain this welcoming atmosphere, and to collaborate in a fun and productive way, we expect contributors to the project to abide by the Contributor Code of Conduct

License

The package is released under the MIT License.