Malicious URL Detection Using Character-Level Tokenization with CNN-LSTM Deep Learning Model

As cyberattacks have become increasingly sophisticated, malicious URL detection is now essential in enhancing cybersecurity against phishing, defacement, and malware attacks. This work introduces a practical method for detecting malicious URLs by blending engineered features with deep learning meth- ods. One notable contribution of our method is character-level tokenization, which allows the models to learn refined structure patterns from raw URLs directly, a feature especially useful when dealing with obfuscated or novel URLs. We built a large and varied dataset of $\mathbf{6 5 1, 1 9 1}$ URLs, drawn from publicly available cybersecurity sources, with balanced coverage across benign, defacement, phishing, and malware classes. Following prepro- cessing and feature fusion, our method was tested with various learning architectures. The system had high performance, with accuracy between 85 % and 97 %, demonstrating the strength to integrate sequential and handcrafted features. This framework demonstrates a new and solid direction for URL-based threat intelligence in real-world cyber security applications.