salamiMan/Email_Phising_V2.4

Phishing Dataset

Phishing datasets compiled from various resources for classification and phishing detection tasks.

Dataset Details

All datasets have been preprocessed in terms of eliminating null, empty and duplicate data. Class balancing has also been performed to avoid possible biases. Datasets have the same structure of two columns: text and label. Text field can contain samples of:

• URL
• SMS messages
• Email messages
• HTML code

Depending on the dataset it belongs to; if it is the combined dataset it will have all data types. In addition, all records are labeled as 1 (Phishing) or 0 (Benign).

Source Data

Datasets correspond to a compilation of 4 sources, which are described below:

• Mail dataset that specifies the body text of various emails that can be used to detect phishing emails, through extensive text analysis and classification with machine learning. Contains over 18,000 emails generated by Enron Corporation employees.

• SMS message dataset of more than 5,971 text messages. It includes 489 Spam messages, 638 Smishing messages and 4,844 Ham messages. The dataset contains attributes extracted from malicious messages that can be used to classify messages as malicious or legitimate. The data was collected by converting images obtained from the Internet into text using Python code.

• URL dataset with more than 800,000 URLs where 52% of the domains are legitimate and the remaining 47% are phishing domains. It is a collection of data samples from various sources, the URLs were collected from the JPCERT website, existing Kaggle datasets, Github repositories where the URLs are updated once a year and some open source databases, including Excel files.

• Website dataset of 80,000 instances of legitimate websites (50,000) and phishing websites (30,000). Each instance contains the URL and the HTML page. Legitimate data were collected from two sources: 1) A simple keyword search on the Google search engine was used and the first 5 URLs of each search were collected. Domain restrictions were used and a maximum of 10 collections from one domain was limited to have a diverse collection at the end. 2) Almost 25,874 active URLs were collected from the Ebbu2017 Phishing Dataset repository. Three sources were used for the phishing data: PhishTank, OpenPhish and PhishRepo.

It is worth mentioning that, in the case of the website dataset, it was unfeasible to bring the total 80,000 samples due to the heavy processing required. It was limited to search the first 30,000 samples, of which only those with a weight of less than 100KB were used. This will make it easier to use the website dataset if you do not have powerful resources.

Combined dataset

The combined dataset is the one used to train BERT in phishing detection. But, in this repository you can notice that there are two datasets named as combined:

• combined full
• combined reduced

Combined datasets owe their name to the fact that they combine all the data sources mentioned in the previous section. However, there is a notable difference between them:

• The full combined dataset contains the 800,000+ URLs of the URL dataset.
• The reduced combined dataset reduces the URL samples by 95% in order to keep a more balanced combination of data.

Why was that elimination made in the reduced combined dataset? Completely unifying all URL samples would make URLs 97% of the total, and emails, SMS and websites just 3%. Missing data types from specific populations could bias the model and not reflect the realities of the environment in which it is run. There would be no representativeness for the other data types and the model could ignore them. In fact, a test performed on the combined full dataset showed deplorable results in phishing classification with BERT. Therefore it is recommended to use the reduced combined dataset. The combined full dataset was added for experimentation only.

Processing combined reduced dataset

Primarily, this dataset is intended to be used in conjunction with the BERT language model. Therefore, it has not been subjected to traditional preprocessing that is usually done for NLP tasks, such as Text Classification.

You may be wondering, is stemming, lemmatization, stop word removal, etc., necessary to improve the performance of BERT?

In general, NO. Preprocessing will not change the output predictions. In fact, removing empty words (which are considered noise in conventional text representation, such as bag-of-words or tf-idf) can and probably will worsen the predictions of your BERT model. Since BERT uses the self-attenuation mechanism, these "stop words" are valuable information for BERT. The same goes for punctuation: a question mark can certainly change the overall meaning of a sentence. Therefore, eliminating stop words and punctuation marks would only mean eliminating the context that BERT could have used to get better results.

However, if this dataset plans to be used for another type of model, perhaps preprocessing for NLP tasks should be considered. That is at the discretion of whoever wishes to employ this dataset.

For more information check these links:

• https://stackoverflow.com/a/70700145
• https://datascience.stackexchange.com/a/113366

How to use them

You can easily use any of these datasets by specifying its name in the following code configuration:

from datasets import load_dataset

dataset = load_dataset("salamiMan/Email_Phising_V2.4")

For example, if you want to load combined reduced dataset, you can use:

dataset = load_dataset("salamiMan/Email_Phising_V2.4")

Due to the implementation of the datasets library, when executing these codes you will see that only a training split is generated. The entire downloaded dataset will be inside that split. But if you want to separate it into test and training sets, you could run this code:

from datasets import Dataset
from sklearn.model_selection import train_test_split

df = dataset['train'].to_pandas()
train, test = train_test_split(df, test_size=0.2, random_state=42)
train, test = Dataset.from_pandas(train, preserve_index=False), Dataset.from_pandas(test, preserve_index=False)