VerSign: Signature Identification and Verification using One Class Support Vector Machines

Signatures are used daily, whether to sign for credit cards or for contracts. However the validity of a signature is often left up to a human, and isn’t checked for accuracy.

31.8 million Credit cards were breached in 2014 alone and therefore it’s imperative that a method of verification for signatures be developed.

This experiment builds a working one class support vector machine to better determine the authenticity of signatures.

The one class support vector machine uses features based upon the histogram of the image. By doing so the algorithm is able to identify which values for which features qualify an image as a genuine, and which are forgeries.

The datasets contain different amounts of genuine signatures, and forgeries as well as a varying training set. The training set size ranged from 9 – 25 to determine the effect of number of training signatures on accuracy.

In order to determine the optimal configuration parameters of the OCSVM were changed to help increase the efficacy of the algorithm.

After completing the hyper parameter selection, the OCSVM reached accuracies of nearly 90% with perfect settings. Its efficacy and accuracy demonstrated the possibility of a writer independent system in signature verification.

The optimum hyperparameter selection consists of a nu value of .2 - .25, and a gamma value of 0.01 – 0.02. This provides the least false positives, while maximizing the amount of true negatives.

The OCSVM model can be used in Point-Of-Sale systems, in signature databases to validate contracts, or as a biometric password.