We tested our trained network on hundreds of signatures to determine a ‘cut off’ value. This value determined which signatures were thought to be real and which were thought to be forged. Our favorite results were asking our advisor to try to forge our signatures. Dr. Morley attempted to forge our signatures in three ways: blind, copy, and seen.

Blind – No previous knowledge of the signature

Copy – Direct copy of the signature

Seen – Look at signature and sign a couple minutes later

The results can be seen below. Verified is asking did the neural network believe the signature was real…this means it was properly forged.

Further Improvements

  • This project was solely done in Matlab. While that is not a problem as that was a clear design decision there is room for growth replication this application in another program, like Python
  • Given more time and less memory constraints it would be great to add more people to the network. The network returns the population normalized value. As the population grows the network will change with it
  • Learned a valuable lesson that 10 hours of experimenting with code can save you 1 hour of research