Phishing Attack Dataset

Citation Author(s):: Ozgur Koray Sahingoz (Computer Engineering Department, Biruni University, 34100, Istanbul, Turkey)

Ebubekir Buber (Computer Engineering Department, Yildiz Technical University, 34469, Istanbul, Turkey)

Emin Kugu (Software Engineering Department, TED University, 06420, Ankara, Turkey)
Submitted by:: Emin Kugu
Last updated:: Sat, 05/03/2025 - 09:25
DOI:: 10.21227/4098-8c60
Data Format:: *.avi; *.csv; *.txt; *.zip

1621 views

Categories:

Keywords:

phishing attack; artificial intelligence; deep learning

ACCESS DATASET CITE

Abstract

The data set has been prepared as 2 different versions. The data set was shared in two versions due to the fact that the researchers could easily reproduce the tests and hardware limitations. The first version (small_dataset) was prepared using a 10% sub-sample of all dataset. The other version (big_dataset) contains the entire data. In this study, the scenarios tested were run on the small_dataset. The most successful configuration that was selected as a result of the analysis on small_dataset was applied to big_dataset.

Both versions were divided into 3 parts: train, validation and test. The proportions used in the split process are as follows; train: 0.7, val: 0.2 and test: 0.1

Phishing URLs are collected from Phishtank, Legal URLs are collected from Commoncrawl.

Datasets are shared under dataset/ folder. The dataset folder has the following structure.

Instructions:

This project has been developed to detect Phishing URLs using Deep Learning. A comprehensive dataset has been collected for this purpose. This data set has been shared so that one can use them in their own studies. The data set consists of two classes: Phishing URLs and Legitimate URLs.

Installation / Clone and install requirements

$ git clone https://github.com/ebubekirbbr/dephides
$ cd dephides/
$ sudo pip install -r requirements.txt

Dataset

Both versions were divided into 3 parts: train, validation and test. The proportions used in the split process are as follows; train: 0.7, val: 0.2 and test: 0.1

The details of the data set are as follows;

Dataset	Train	Validation	Test	Overall
small_dataset	364,199	104,576	51,510	520,285
big_dataset	3,641,986	1,045,774	515,080	5,202,841

Dataset	Phishing	Legitimate	Overall
small_dataset	232,090	288,195	520,285
big_dataset	2,320,893	2,881,948	5,202,841

Phishing URLs are collected from Phishtank, Legal URLs are collected from Commoncrawl. Please read the article for detailed information.

Datasets are shared under dataset/ folder. The dataset folder has the following structure.

dataset/
├── big_dataset
│   ├── meta.txt
│   ├── test.txt
│   ├── train.txt
│   └── val.txt
├── small_dataset
│   ├── meta.txt
│   ├── test.txt
│   ├── train.txt
│   └── val.txt
├── top100k_cc.txt
└── top_100k_errored_cc.txt

meta.txt contains information about the sample counts in the datasets. top100k_cc.txt contain legal domains from the CommonCrawl. These domains were used to collect legitimate URLs.

Train

There are 2 main scripts in this project. dl.py is the script that performs the necessary operations for the training and testing. dl_models.py 'is the script that contains the model architectures tested.

dl.py parameters are as follows.

$ dl.py [-h] [-ep EPOCH] -arch ARCHITECTURE [-bs BATCH_SIZE]

optional arguments:
  -h, --help            show this help message and exit
  -ep EPOCH,          --epoch EPOCH
                            The number of epoch
  -arch ARCHITECTURE, --architecture ARCHITECTURE
                            Architecture function in dl_models.py
  -bs BATCH_SIZE,     --batch_size BATCH_SIZE
                        batch size

A model in dl_models.py must be selected for train process. The function name should be given to the dl.py script as a architecture parameter.

An example usage;

python dl.py -arch cnn_base

The epoch number and batch size can be given as parameters for train. For example;

python dl.py -arch cnn_base -ep 30 -bs 1000

Training log

.
.
.
Total params: 449,044
Trainable params: 449,044
Non-trainable params: 0
_________________________________________________________________
Instructions for updating:
Use tf.cast instead.
Train on 10 samples, validate on 10 samples
Epoch 1/10
14000/3641986 [==============================] - 118s 32us/step - loss: 0.1481 - acc: 0.9352 - val_loss: 0.0850 - val_acc: 0.9667
Epoch 2/10
21000/3641986 [==============================] - 118s 32us/step - loss: 0.1481 - acc: 0.9352 - val_loss: 0.0850 - val_acc: 0.9667
.
.
.

Tested Deep Learning Models

The available models tested in the article are as follows;

rnn_base
brnn_base
cnn_base
ann_base
att_base
rnn_complex
brnn_complex
ann_complex
att_complex
cnn_complex
cnn_complex2
cnn_complex3

Define Custom DL Model

If you want to train your own model, all you need to do is define the model into dl_models.py. Then you should run your own model with as follows;

python dl.py -arch custom_model

It is worth repeating that you should give the function name of the model you defined in dl_models.py as a parameter. For the example above, the model you create in dl_models.py must be defined in the function named custom_model.

Test Results

Many information about the test results described in the article are shared in the test_results folder. Shared assets for each test performed are as follows.

.
├── accuracy.png                    : Accuracy Graphic
├── char_embeddings.json            : Character Embeddings (dim. 50), if you want to use transfer learning, you can use this embeddings
├── classification_report.txt       : Precision, Recall, F-Measures
├── confusion_matrix.png            : Confusion Matrix
├── floyd_command.txt               : The command used to run the test in Floydhub
├── gpu_utilization.png             : Resources consumed during running
├── loss.png                        : Loss Graphics
├── model.json                      : Model (only json)
├── model_all.h5                    : Model with Weights
├── model_summary.txt               : Model Summary
├── normalized_confusion_matrix.png : Normalized Confusion Matrix
├── raw_test_results.json           : All debug info as json
└── weights.h5                      : Model Weights

The most successful model in the article was cnn_complex1 on big dataset. Test results for cnn_complex model as follows;

path of assets: 'test_results/big_dataset/'

Model Summary

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding_1 (Embedding)      (None, 200, 50)           4900      
_________________________________________________________________
conv1d_1 (Conv1D)            (None, 198, 128)          19328     
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 66, 128)           0         
_________________________________________________________________
dropout_1 (Dropout)          (None, 66, 128)           0         
_________________________________________________________________
conv1d_2 (Conv1D)            (None, 66, 128)           114816    
_________________________________________________________________
dropout_2 (Dropout)          (None, 66, 128)           0         
_________________________________________________________________
conv1d_3 (Conv1D)            (None, 66, 128)           82048     
_________________________________________________________________
dropout_3 (Dropout)          (None, 66, 128)           0         
_________________________________________________________________
conv1d_4 (Conv1D)            (None, 66, 128)           49280     
_________________________________________________________________
max_pooling1d_2 (MaxPooling1 (None, 22, 128)           0         
_________________________________________________________________
dropout_4 (Dropout)          (None, 22, 128)           0         
_________________________________________________________________
conv1d_5 (Conv1D)            (None, 22, 128)           82048     
_________________________________________________________________
dropout_5 (Dropout)          (None, 22, 128)           0         
_________________________________________________________________
conv1d_6 (Conv1D)            (None, 22, 128)           49280     
_________________________________________________________________
max_pooling1d_3 (MaxPooling1 (None, 7, 128)            0         
_________________________________________________________________
dropout_6 (Dropout)          (None, 7, 128)            0         
_________________________________________________________________
conv1d_7 (Conv1D)            (None, 7, 128)            49280     
_________________________________________________________________
max_pooling1d_4 (MaxPooling1 (None, 2, 128)            0         
_________________________________________________________________
dropout_7 (Dropout)          (None, 2, 128)            0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 256)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 2)                 514       
=================================================================
Total params: 451,494
Trainable params: 451,494
Non-trainable params: 0
_________________________________________________________________