tushaargvs/phishing

Feature Selection for Spam and Phishing Detection

The ease of communicating through email saw a huge increase in Unsolicited Bulk Email(UBE). Unsolicited emails are broadly divided into 2 categories: Spam(mass mailing approach to marketing) and Phishing(impersonatisation for the purpose of stealing data). This project gives a machine learning to approach to classify them into spam and phishing categories. It considers a total of 40 features which are broadly categorised into URL based, body based, sender based, subject based and script based features.

An addition to this extraction involved feature selection using:

• Low Variance filter
• High correlation filter
• Feature importances
• mRMR

Getting Started

The project has been divided into 3 modules for convenience. The first module deals with extraction of intricate features from the emails and preparing the dataset for application fo the module. The entire procedure is illustrated as below!

Requirements

• Python2
• Jupyter Notebook
• Necessary Python libraries (check first cells of Feature Extraction and Feature Selection)

Running

• Run Feature Extraction sequentially to obtain 3 datasets in CSV format
• Once CSVs with 40 features have been generated, now run the Feature Selection sequentially
• In Feature Selection, the reference file chosen was dataset_HSP.csv which can be changed in the calling of mRMR_CSV('dataset_HSP', 'label')

Testing and Results

Voting Ensemble classifier with SVM, Naive Bayes, LDA, Adaboost, Random Forest and CART was used in the implementation of mRMR feature selection The selected feature accuracy was tested against the origibal (40 features) using the following algorithms:

• Voting Ensemble classifier (same as above)
• SVM
• Stochastic Gradient Boosting
• Extra Trees classifier
• Adaboost
• Random Forest
• Bagged Decision Tree classifier (CART)
• Naive Bayes