- Solves all types of puzzle piece captchas
- Created as a proof of concept to show how front-end captchas are vulnerable to AI solving
- Also wanted to build an add-on to my previously published project which bypasses DataDome's detection almost every time
- 1st Place Online Presentation at JCC Honors Experience
- Manually labeling hundreds of thousands of images would be infeasible which requires us to create our own versions of their captchas
- GeeTest uses a pool of cutouts and base images to create captchas, places them within a frame randomly and fades the cutout
- DataDome does not have a demo site, only pops up on a protected site or have a bad cookie or send an invalid POST request, uses random images with same cutout
- Originally started with TensorFlow but switched to YOLOv11
- Used YOLOv11n.pt with 7 epochs of training since it instantly plateaued, all args are here
Training (Optional as I Added the Weights Here)
- Generate captchas with generatecaptchas.py using images in the assets folder
- Create the dataset using this script
- If you've never used YOLO before here's their setup page which goes over training
- I created a template/POC you can use for solving in browser with the GeeTest demo page.
- You will need to change some of the captcha identifiers depending on the site