BustedURL: Collaborative Multi-agent System for Real-Time Malicious URL Detection

Phishing is a major security issue in cyberspace, and time-stamping the detection of malicious URLs is crucial to safeguarding users and systems. Current approaches like URLNet and URLTran are reasonable, but the methods generally cannot adapt to quick changes and cannot be easily scaled as a standalone agent. In this paper, we present BustedURL, a sustainable framework designed to address these limitations. Leveraging a distributed, collaborative multi-agent architecture, BustedURL incorporates advanced methodologies, including transformers, ensemble learning, stacking, big-data aggregation, and sophisticated learning pipelines. These innovations enhance the framework's adaptability across diverse environments. We empirically evaluate the proposed architecture using real-time datasets from OpenPhish and various other sources, demonstrating that BustedURL outperforms current state-of-the-art solutions across various performance metrics in dynamic phishing contexts, owing to its distributed, scalable, and adaptive characteristics. Specifically, the scalability experiments demonstrated an approximately 50-fold improvement compared to the competitive baseline models.