We’ve created a framework to identify Handwritten Signature fraud in checks and contracts in which, after being scanned, they can automatically be standardized and inserted into an algorithm that would verify if the analyzed document is authentic or a fraud.
Even though there has been a great number of publications in the area of Handwritten Signature Verification in recent years, the challenge to produce highly-effective tools is still open to new technology.
In a traditional process, an institution employee, usually a bank clerk, receives checks from customers and compares the signatures to a signature database. Employee have only so many working hours per day to complete this non-stop process, which is subject to unpredictability and human error.
Fraud possibilities are classified into three classes. In the first, the scammer doesn’t know the victim’s real signature. The probability of forging a signature similar to the original is low because signatures vary from cursive to abstract forms. The second class is equivalent to an amateur trying to copy a signature they have already seen. The signatures will probably have great similarities, but a well-trained algorithm would be able to spot the differences. This is our potential class. The third and last category is composed of professional forgeries. If a human being trained for this task finds it an immense challenge, it will also be difficult for an algorithm.
Moreover, there are practical and inherent challenges surrounding the subject:
· Standardization of different-sized signatures
· Different check layouts from the same bank
· Microfilm residues present during document scanning
· Signatures that go out of the document’s textual fields
· Documents scanned at a slant or moderately rotated
Until the 2000s, the traditional way to verify signatures used geometric attributes extraction, graph metrics, directionals, mathematical transformations, and texture, among others.
With the application of Neural Network and Deep Learning methods, new opportunities have emerged. Instead of manual attribute extractors, the raw pixels were used for several Deep Learning architectural structures.
Our team’s approach was to use Artificial Intelligence’s most advanced techniques to automate the identification of a document’s signature fields, then extract them to go through an image treatment process. These standardized images are the input sources for Convolutional Neural Network algorithms, which create numerical representations of each signature’s attributes.
The database contains a set of signatures for each customer. We compare them with authentic and forged signatures on checks/contracts through diverse similarity metrics. We feed classification models to obtain the final result. In our tests, we were able to correctly identify the origin of 52 out of 53 checks available for verification.
For more complex cases, in which the algorithm cannot reach a decision about verification, they will be sent to a trained team who will make the final decision.
The use of algorithms provides cost reduction, process optimization and scale gain. However, it is important to highlight that this tool should be used to help the operation focus on more valuable tasks or for cases that require experience and human ability.
More details can be found in the forthcoming paper.
By: João Victor Dias