Detecting Web-Based Attacks: A Comparative Analysis of Machine Learning and BERT Transformer Approaches

As web attacks continue to evolve, web applications are increasingly vulnerable to various security threats and network attacks. Malicious actors can inject harmful code in an HTTP request to launch attacks like SQL injection, XSS, buffer overflow, and others. Detecting and classifying unknown web attacks is essential for enhancing the reliability and security of web applications. In this study, we employ a Transformer called Bidirectional Encoder Representations (BERT) and several machine learning techniques (CNN, SVM, Random Forest, Naive Bayes, etc.) to categorize HTTP requests based on their attack type. We then compare the results obtained from all the techniques and observe that BERT achieves the highest accuracy of 99% compared to all other classification methods used.