Neural Network-Based Detection of Phishing URLs Using Embeddings

Phishing attacks threaten online security, deceiving users into divulging sensitive information to malicious websites. Currently, as traditional blacklist methods struggle to keep pace with the constantly evolving phishing techniques, machine learning offers an adaptive solution. In this study, we propose creating a neural network-based model to identify phishing URLs using features directly extracted from the URL string. Instead of relying on raw URL tokens or external methods such as website content, our method extracts features related to characteristics of the URL string, including key phrases, and transforms them into dense vector embeddings. These embeddings serve as the input to the neural network, which then teaches complex relationships among feature combinations. This allows the model to generalize beyond external patterns and resist overfitting. Evaluated using the most up-to-date dataset, this method achieved 99 percent accuracy in detecting phishing URLs, outperforming other classical machine learning models.