Abstract
The rise of automated bots has reduced the effectiveness of traditional CAPTCHA systems. Designed as reverse Turing tests to prevent credential stuffing, spam, and unauthorized account creation, CAPTCHAs are now facing growing challenges due to advances in deep learning, particularly convolutional neural networks (CNNs), which enable automated solvers to bypass these systems with high accuracy. This study investigates the vulnerability of CNN-based CAPTCHA solvers to small adversarial perturbations, demonstrating that even minor input modifications can significantly deteriorate model performance. Furthermore, lightweight defense strategies are evaluated to counter such attacks while preserving human readability. These results highlight the urgent need to strengthen CAPTCHA mechanisms against adversarial threats while ensuring that security enhancements do not compromise user accessibility.