Detection of xss attacks and sql injections using convolutional neural networks

Abstract

Cross-site scripting (XSS) and SQL injection (SQLi) remain critical threats to web application security, often bypassing traditional rule-based defenses through obfuscation techniques. This study proposes a lightweight convolutional neural network (CNN) that detects malicious inputs directly from raw HTML or script fragments using an ASCII-based grayscale representation. Evaluated on a public dataset of approximately 27,000 samples, the model achieves 98.39% accuracy with ~1 ms inference latency. The results demonstrate that compact CNN architectures can provide efficient, real-time detection of injection-based attacks in practical deployment environments.