Karakter Düzeyinde Doğal Dil İşleme ve 1D-CNN Tabanlı Karar Destek Sistemi ile Web Uygulama Güvenlik Duvarları

Özet (abstract)

Rule- and signature-based Web Application Firewalls (WAFs) may struggle with evolving malicious requests. This study presents a WAF decision-support framework that combines character-level HTTP request representation with a one-dimensional convolutional neural network (1D-CNN). The sigmoid output is treated as an uncalibrated attack score, which allows threshold adjustment without retraining. Across five seeds, each run evaluates 999 validation thresholds ranging from 0.001 to 0.999. The threshold that maximizes validation F1-score is locked before the corresponding held-out test evaluation. Cross-seed one-standard-error analysis is used only to characterize threshold stability. It is not interpreted as a confidence interval for the optimal threshold. Evaluation on CSIC 2010 and SR-BH 2020 includes baseline comparisons, request-length sensitivity analysis, model-only CPU runtime characterization, and normalized-prefix group-disjoint testing. Group-disjoint F1-scores were 0.8590 ± 0.1079 and 0.9912 ± 0.0005, respectively. CSIC 2010 showed greater variability and performance degradation relative to random splitting, indicating sensitivity to group composition. Under matched random splits, the 1D-CNN achieved higher mean F1-scores than Character-MLP and Char TF-IDF + LR. However, the improvement over the competitive TF-IDF baseline was modest. The baseline models were not evaluated under group-disjoint partitions. The principal contribution is therefore a reproducible evaluation framework and operational threshold analysis rather than substantial gains in detection accuracy.

Yazarların özeti; kaynağından alınmıştır. Electronics, 2026 · DOI ↗