Phishing URL detection with neural networks: an empirical study

Area of Science:

• Computer Science
• Cybersecurity
• Artificial Intelligence

Background:

• Phishing websites pose a significant threat by mimicking legitimate sites to steal sensitive data.
• Accurate classification of URLs is crucial for cybersecurity defense.
• Existing methods require enhancement to combat sophisticated phishing tactics.

Purpose of the Study:

• To develop and evaluate advanced artificial intelligence and machine learning models for URL classification.
• To introduce a unique probabilistic neural network approach for improved phishing URL detection.
• To demonstrate the efficacy of machine learning models trained on combined open-source and private datasets in production environments.

Main Methods:

• Application of deterministic and probabilistic neural network models for URL classification.
• Development of a novel approach using probabilistic neural networks.
• Creation of a comprehensive dataset combining open-source (Alexa, PhishTank, OpenPhish) and private production data (EasyDMARC).
• Feature engineering conducted without third-party dependencies.

Main Results:

• Achieved an average of 97% accuracy on a validation dataset through daily model validation.
• Demonstrated the success of a machine learning model trained on a hybrid dataset in a production setting.
• Evaluated both deterministic and probabilistic models, showing high accuracy on URLs of all lengths.

Conclusions:

• Probabilistic neural networks offer an improved approach to URL classification accuracy.
• Machine learning models trained on diverse datasets are effective for real-world phishing detection.
• The developed models provide a robust solution for identifying both short and long phishing URLs.