FedURL: A BERT-based Federated Malicious URL Detection Framework

From the standpoint of cybersecurity, malicious websites have always been a significant concern. Traditional URL safety measures relied on blacklists, but the surge in malicious URLs has necessitated advanced machine learning and deep learning techniques. The ever-evolving nature of malicious URLs poses a significant challenge, requiring continuous model training from diverse data sources to maintain optimal performance. Given the inherent security risks in malicious URLs and their dispersed distribution, direct data transfer is impractical. Moreover, data custodians often face challenges in data sharing due to commercial and privacy concerns. To address these challenges, we introduced FedURL, a BERT-based federated framework that enabled effective malicious URL detection without compromising data security or privacy. Upon leveraging a pre-trained BERT model, we fine-tuned it for malicious URL detection. Both open-source and real-world datasets were used to evaluate the performance of FedURL. Additionally, we evaluated factors such as communication cost, training strategies and client numbers during fine-tuning to ensure its efficient real-world deployment. Our results reveal that FedURL not only matches the performance of state-of-the-art methods but also offers enhanced data security and diversity through the application of federated learning.