Signature Manipulation Authenticator (SigMA)

Inspiration

We started this project with some difficulties in picking a topic. Both Ash and Omar wanted to make something using machine learning, and one of the first ideas we came up with was a lock system based facial expressions, based on things like smart doorbells which are already installed in many homes. We quickly realized that this system would be both impractical and out of scope for the time given, so we switched to focusing on handwritten signatures instead.

What it does

The project is an application where a user types the name of a person and then uploads a picture of them, and then the program certifies whether the signature is valid or not.

How we built it

We used a tensorflow model on a dataset of 2640 signatures, split 50/50 between real and forged signatures.

Challenges we ran into

On the first day, we struggled with coming up with the idea, and we lost many hours of potential work time. This limited the scope of our project. On the second day, we had a particularly time costing run of a model where the model took more than 2 hours to train before crashing. This was after a series of bad training runs where the model achieved a low accuracy of only 0.9%. The dataset itself was quite tricky to work with, and took a lot of manual fiddling with the labeling of the data to make the model train properly. On the final day, we found that the gui library wasn't compatible with mac, and that pushed our ability to demo to the last possible moment.

Accomplishments that we're proud of

we achieved a model that is 79% accurate, which is quite high considering the limitations imposed by a hackathon setting in both

What we learned

We learned about how to use Tensorflow and Tkinter to build applications using neural networks. We also learned how to use github copilot during the workshop.