Improvement of Text CAPTCHA Codes by Comparing Adversarial Techniques Against Deep Learning Model Attacks

CAPTCHAs are essential tools in computer security to distinguish between humans and automated programs. Although widely used in web applications to prevent unauthorized access and spam, advances in artificial intelligence have increased attacks against these systems. This study focuses on improving the security of CAPTCHAs using adversarial techniques such as FGSM and PGD, exploring their effectiveness against a deep learning model. Furthermore, a generative adversarial network is employed to strengthen resistance to these attacks. The research also includes human validation to evaluate the robustness of different types of CAPTCHAs against simulated attacks. Our findings demonstrate that while adversarial modifications enhance security, they require careful calibration to avoid excessive usability degradation.