Deep Learning for Web Attack Detection Using Transformers (Prof. Khalid Al-Zamel)

As web applications continue to grow, the attack surface for service providers has expanded significantly due to the constant development of attack technologies. This exposes any Internet-connected node to the risk of being targeted within a very short period of its public release. In our research, we introduce an innovative approach to detect web attacks using HTTPBERT, a transformer-based model that is fine-tuned on HTTP requests using Masked Language Modeling (MLM). Our model surpasses existing research in the field and maintains consistent performance across four different datasets, demonstrating significant potential for real-world applications. Our HTTPBERT model achieves an impressive accuracy rate of 99.2% and an F1-score of 99.5% on average. Furthermore, recognizing the scarcity of recent and realistic publicly available datasets, we have developed a dataset that can serve as a foundational resource for future research in this domain.

Supervisor: Prof. Khalid Al-Zamel
Convener: Prof. Mehmet Karaata
Examination Committee: Prof. Mehmet Karaata , Prof. Mahmoud Ben Naser