Solving Pattern Identification CAPTCHAs

Wuyang Du will present his senior thesis talk at this public event.

Abstract: CAPTCHAs, designed to differentiate between human users and automated bots, are increasingly vulnerable to sophisticated automated attacks, especially those utilizing deep neural networks. Our research addresses this challenge by developing a solver for Click-Based Pattern Identification CAPTCHAs that uses a combination of generative inpainting (to reconstruct the underlying patterns) and segmentation (to dissect the CAPTCHA elements). Unlike existing CAPTCHA solvers that require fine-tuning on unseen datasets, our method generalizes better to unseen CAPTCHAs with 50%~70% accuracy without fine-tuning, outperforming other models with less than 10% of overall accuracy. Our work underscores the need for more robust CAPTCHA designs in the face of advancing AI technologies.