Web Attack Classification Method Based on Pre-Trained Language Model ELECTRA

In response to the current challenges of insufficient dataset coverage and suboptimal detection performance in certain categories within Web attack classification tasks, we extends the SR-BH 2020 dataset to develop the new Web-2025 dataset, which includes 15 types of Web attacks. Also, we introduce a classification method based on the pre-trained language model ELECTRA, which substantially improves the detection of various Web attacks through enhanced semantic understanding of complex attack requests. Our method enhances detection performance across various attack types, with the most notable case showing improvements of over 20 percentage points in Precision, Recall, and F1-score. Thus, the ELECTRA-based Web attack classification method we proposed in this study offers an effective solution to enhance the accuracy of Web security defenses.