capptcha.github.io
CAPTCHA
We have build many models to solve some of the difficult open sourced CAPTCHAs that are available on the internet. We have obtained about more than 99.5% accuracy on most of the models, which converges at about 5 epochs. The generators folder have some of the modified codes that we have used to generate the data to feed into the model. The pyfiles folder section have all of the models and their corresponding python codes.
Results
| CAPTCHA name | CAPTCHA img | Algorithm used | Accuracy Obtained | Try out in Google Colab |
|---|---|---|---|---|
| JAM CAPTCHA |  |
kNN | 99.53% |  |
| CNN_c4l_16x16_550 |  |
CNN - modified CIFAR 10 | 99.91% |  |
| captcha-1L |  |
Own CNN model - multilabel classification | 99.67% |  |
| captcha_4_letter |  |
LSTM model - multilabel classification | 99.87% |  |
| captcha_v2 |  |
Own CNN - multilabel classification | 90.102% |  |
| circle_captcha |  |
Alex Net with multilabel classification | 99.99% |  |
| faded |  |
Alex Net with multilabel classification | 99.44% |  |
| fish_eye |  |
Alex Net with multilabel classification | 99.46% |  |
| mini_captcha |  |
Alex Net with multilabel classification | 97.25% |  |
| multicolor |  |
Alex Net with multilabel classification | 95.69% |  |
| railway_captcha |  |
Own CNN model | 99.94% |  |
| sphinx |  |
Alex Net with multilabel classification | 99.62% |  |
Documentation
| [Thesis - Deceiving computers in Reverse Turing Test through Deep Learning (Research paper)] | [Slides] |
Advisor
Acknowledgements
Frequently Asked Questions
- Are these the only notebooks? - No, https://colab.research.google.com/github/Jimut123/CAPTCHA/blob/master/pyfiles/sphinx/sphinx_33_10e_9873.ipynb is the path for testing the notebooks in Colab, please use this format for testing other notebooks, there are some awesome visualizations too...
- Do we need to download the data? - No, it is automatically downloaded, you just need to plug and play for getting the job done in Google Collaboratory.
- Training time is taking too long? - Yes, some of the CAPTCHAs really take long time to train, (over 10 hrs for just 10 epochs even in GPUs). It is good to have multiple GPUs when you are using this on your own machine.
- Found a bug? or version issue? - PRs welcome, fork it, and send a pull request!
Contribution
Please feel free to raise issues and fix any existing ones. Further details can be found in our code of conduct.
While making a PR, please make sure you:
- Always start your PR description with “Fixes #issue_number”, if you’re fixing an issue.
- Briefly mention the purpose of the PR, along with the tools/libraries you have used. It would be great if you could be version specific.
- Briefly mention what logic you used to implement the changes/upgrades.
- Provide in-code review comments on GitHub to highlight specific LOC if deemed necessary.
- Please provide snapshots if deemed necessary.
- Update readme if required.
BibTeX and citations
@article{DBLP:journals/corr/abs-2006-11373,
author = {Jimut Bahan Pal},
title = {Deceiving computers in Reverse Turing Test through Deep Learning},
journal = {CoRR},
volume = {abs/2006.11373},
year = {2020},
url = {https://arxiv.org/abs/2006.11373},
archivePrefix = {arXiv},
eprint = {2006.11373},
timestamp = {Tue, 23 Jun 2020 17:57:22 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2006-11373.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}