Phishing Detection with Adaptive Loss Function

A research project proposing a novel loss function for phishing detection: Phishing-Aware Adaptive Focal Loss (PAFL).

Overview

This project proposes a new loss function that addresses class imbalance and domain adaptation in phishing URL detection, with a lightweight implementation that runs on Google Colab’s free tier.

Research Contributions

1. Phishing-specific loss function: Incorporates URL structural features (domain trust, brand similarity, structure) into the loss.
2. Dynamic class weighting: Adjusts weights during training based on the distribution of hard vs. easy samples.
3. Domain adaptation: Explicitly models the distribution gap between source and target domains in the loss.

Project Structure

phishing/
├── notebooks/          # Jupyter notebooks for Google Colab
├── src/                # Source code
│   ├── data/          # Data loaders and feature extraction
│   ├── models/        # Model definitions
│   ├── losses/        # Loss function implementations
│   └── utils/         # Utilities (metrics, visualization)
├── data/              # Datasets (after download)
└── requirements.txt   # Dependencies

Setup

pip install -r requirements.txt

Datasets

• PhiUSIIL Phishing URL Dataset (235,795 URLs, 54 features)
• Feature-Engineered URL Dataset (111,660 URLs, 22 features)

Usage

See the notebooks for step-by-step usage:

• notebooks/01_data_preparation.ipynb — Load and prepare data
• notebooks/02_baseline_models.ipynb — Train and evaluate baselines
• notebooks/03_proposed_loss_function.ipynb — Train with PAFL and compare losses

To run experiments from the command line:

python run_experiments.py

License

MIT License