Adversarial Attacks and Defense Evaluation in Deep Learning-Based CAPTCHA Recognition Systems

Deep learning-based CAPTCHA recognition systems are increasingly used in modern authentication frameworks; however, their robustness against adversarial attacks remains insufficiently explored, particularly in sequence-based recognition architectures. This study evaluates targeted adversarial vulnerabilities in a Convolutional Recurrent Neural Network (CRNN) optimized using Connectionist Temporal Classification (CTC) loss on a publicly available real-world dataset of 1,070 CAPTCHA images. Unlike prior studies that primarily focus on adversarial robustness in image classification tasks, this work investigates the impact of targeted adversarial attacks on CAPTCHA sequence recognition systems and assesses the effectiveness of a simple preprocessing-based defense. To assess robustness, white-box Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks are applied. Under clean conditions, the CRNN model achieved a sequence-level recognition accuracy of 98.73%. FGSM achieved an attack success rate of 75.66%, reducing recognition accuracy to 23.07%, while PGD achieved an attack success rate of 34.00%, with post-attack accuracy of 64.73%. A Gaussian blue preprocessing defense was evaluated at inference time to mitigate adversarial perturbations. The defense reduces FGSM attack success rate from 75.66% to 51.45% and partially restored recognition performance, but provided no measurable improvement against PGD-generated adversarial examples. The results demonstrate that CRNN-based CAPTCHA recognition systems remain highly vulnerable to targeted adversarial manipulation and that simple preprocessing defenses offer limited protection against adaptive gradient-based attacks. These findings highlight the need for more robust defense mechanisms to improve the security and reliability of AI-driven authentication systems.