FMURL-H: A Federated Multimodal Approach with Hyperparameter Optimization for Malicious URL Detection

Abstract

One important threat vector for malware, phishing, and defacement attacks is malicious Uniform Resource Locators (URLs). A number of approaches, including those based on machine learning, have been developed to deal with this issue. Because they frequently rely on centralized learning and lexical features, traditional detection models continue to be inadequate and constrained in terms of scalability and confidentiality. 

This study introduces FMURL-H (Federated Multimodal URL Detection with Hyperoptimization), a novel framework that combines transformer-based encoders (DeBERTa/RoBERTa), multimodal features, federated learning, and hyperparameter optimization with Optuna. When tested on a massive dataset of 651,191 URLs, FMURL-H outperformed other approaches with an Accuracy of 99.30%, Precision of 99.42%, Recall of 99.48%, and F1 score of 99.45%. This model establishes a new standard for malicious URLs that are both scalable and privacy-conscious.