Adversarial Attack on Deep Learning Model Detecting XSS

Cross-Site Scripting (XSS) remains one of the most persistent vulnerabilities in web applications, necessitating robust detection mechanisms to counter increasingly sophisticated attack patterns. This study proposes a novel approach to attack XSS detection model by using a recurrent neural network (RNN) to generate adversarially perturbed payloads. These perturbed samples mimic real-world malicious inputs by selectively encoding characters with their encoded counterparts. Using a character-level tokenization strategy, the RNN model is trained to predict and perturb sequences that challenge the detection model’s robustness. After generating adversarial samples through probabilistic character-level encoding, we evaluate our XSS detection model on both perturbed and unperturbed samples to assess the outcome. Our attacking technique has successfully managed to reduce the metrics such as accuracy, precision, recall and F1-Score against the detection model of the malicious class.