Pelusa (Predictive Engine for Legitimate & Unverified Site Assessment) is a machine learning based application that predicts the legitimacy of a website based on the URL provided. It is built using FastAPI and PostgreSQL, deployed with Docker Compose.
To get started, clone the repository and run with Docker Compose.
git clone https://github.com/javi-aranda/pelusa-server
cd pelusa-server
docker-compose up # add flag -d to run detached
docker-compose exec -T backend alembic upgrade head # run SQLAlchemy migrationsThat should run the application on http://localhost:8000.
You can get a more detailed reference of the API by visiting http://localhost:8000/docs.
But mainly it consists of an endpoint api/v1/analysis that accepts JSON body with {"input": "<URL_TO_CHECK>"}
and returns the legitimacy of the website (1 means potentially bad, 0 means potentially safe).
Those results are stored in a PostgreSQL database, which could be useful to train the model in a future or as persistence mechanism in case an URL is submitted multiple times in a short period of time.
There is a pgAdmin instance running on http://localhost:5050 with credentials
defined in .env file. After connecting to the PostgreSQL server, you can explore the database and run any query you want.
The dataset used for training the model is handmade, it consists on 30000 URLs, 50% legitimate and 50% malicious.
Malicious websites were randomly sampled from PhishTank active threats and legitimate URLs were sampled from multiple Kaggle datasets. After extracting features for both types, the resulting dataset is phishing_dataset.csv
The model is trained using a Random Forest Classifier with an accuracy of 94% over the training dataset and the code is available as a Jupyter Notebook in train.ipynb
This project was made keeping in mind FastAPI Starter as a reference, but bundling the frontend in a different repository, which is available in Pelusa React.