Phishing email classification

The script is a toolkit for evaluating and comparing the performance of different machine learning models on the task of email classification. These models include:

• Logistic Regression
• Decision Trees
• Random Forests
• Gradient Boosting
• Naive Bayes
• Multi-Layer Perceptrons

By analysing the performance of these models, you can identify the most effective approach for accurately classifying emails and protecting users from phishing scams.

• Data preparation. Loads and preprocesses email data from a CSV file, including label encoding and splitting the dataset for training and evaluation.
• Custom email dataset class. A custom EmailDataset class handles tokenisation so the data meets the BERT model’s input requirements.
• Model comparison. Trains and evaluates Logistic Regression, Decision Trees, Random Forests, Gradient Boosting, Naive Bayes and Multi-Layer Perceptrons on the same task.
• BERT in action. Uses PyTorch and the Hugging Face Transformers library to train and optimise a BERT model specifically for email classification.
• Making predictions. A predict function classifies new emails as phishing or safe with the trained models.
• Saving trained models. Model states are saved, so they can be reloaded for classification without retraining.
• Understanding predictions. The LIME (Local Interpretable Model-Agnostic Explanations) library explains the model’s predictions on specific emails.

• Python
• PyTorch
• Hugging Face Transformers
• BERT
• LIME
• Logistic Regression
• Decision Trees
• Random Forests
• Gradient Boosting
• Naive Bayes
• Multi-Layer Perceptrons

To assess each model during training, the script calculates training and validation loss, accuracy, precision, recall and F1-score.

Authors: Bohdan Boiprav & Catalin Bondari. Dublin City University coursework. Licensed under the MIT License.