Signatures are still very important in the banking world. Although check use has declined, over half of Americans wrote a check last year, and they’re especially popular for business clients. Annually, consumers write well over 10 billion checks worth over $20 trillion, and additionally, the majority of loan documents are still completed with wet signatures.
In many financial institutions, signature verification is still a human-handled back office process. This is time-consuming and error-prone. In contrast, automating signature verification with artificial intelligence (AI) can improve efficiency, reduce errors, and save time. But how does it work? Here’s an overview of the basics.
At the highest level, signature verification software uses AI in the following ways:
- Special descriptive language — The software presents both the reference and the presented signatures as a sequence of hieroglyphic elements and linear transformations to identify correlations and discrepancies.
- Signature segmentation — After taking a holistic approach, the software breaks down the signature into nodes. Then, it creates triangles with apexes in the nodes to compare the similarity of different segments.
- Neural networks — When analyzing the signatures, the software takes one or more input variables and runs them through a network of equations to result in one or more output variables. This allows the system to assess multiple variables in different ways.
There are several different ways to explain how the tools assess signatures. If you break down the process into its core steps, they are detection, preprocessing, and analysis.
Automated signature detection
With checks, the location of the signature is constant. The verification tools simply need to assess the signature and endorsement lines. With loan and account opening documents, however, the software needs to be able to find the signature in different places.
Using AI-powered detection tools signature verification software scans the documents, detects the signature, and crops it for comparison with a reference signature. AI makes this possible with documents in a range of formats in multiple languages.
Signature preprocessing
Once the tool has detected the signature, it preprocesses the signature and a reference signature. Preprocessing cleans unstructured data so that it can be analyzed. For instance, in this situation, preprocessing removes all of the background text that surrounds the signature in a contract. It also cuts the data into smaller segments for analysis.
Preprocessing is a critical part of dealing with unstructured data, and in fact, signature verification is one of the most popular uses of unstructured data analysis. Unstructured data includes images and videos, but it’s not just visual. It also includes audio files, text files, geospatial data, and surveillance data.
The majority of data is unstructured, and it’s much more challenging to gather and assess unstructured than structured data.
Analysis
Once preprocessing is complete, the tool compares the extracted signature to the reference signature and generates a confidence score. For example, a different signature from the same signer may get a 92%, while a skilled forgery may get a 42%, and a blind forgery a 10%. Note that these are just sample numbers. Then, if the score is over a certain threshold, it flags the signature for manual review.
These tools don’t just assess visual similarity. They also analyze dynamic and temporal elements such as speed, fluidity, pen stroke, and pressure. To do this, they rely on complex algorithms informed by pattern recognition, image processing, and geometrical analysis.
When reviewing different types of tools, ask about the risk of false positives. This is a risk with any tool, but to ensure you select the best tool, you should look for the lowest number possible. In particular, you should ask how the tool deals with differences in legitimate signatures.
Differences in legitimate signatures
Both human analysts and automated signature verification tools must be aware of the differences in legitimate signatures. People sign their names in slightly different ways based on their mood, the time available, the size of the signature line, and multiple other factors.
AI tools calculate a dynamic threshold to allow for the variations that occur between multiple signatures from the same person. You can feed the tool multiple signatures from the same person to enhance its ability to detect forgeries. To do this, these tools look at authorized signature images. Then, they find the bottlenecks and compute cosine similarities between all of the different signatures.
In contrast, human analysts struggle when they have multiple reference signatures. This actually reduces their accuracy levels.
AI Tools With Human Exception Handling
The most effective way to spot forgeries is to combine artificial and human intelligence. You need AI-driven signature verification tools and humans that can review exceptions.
Ready to learn more? Then, contact us today. At SQN Banking Systems, we offer a robust suite of tools and services to help our clients detect forgeries and minimize their risk of all other types of bank fraud. For a no-cost fraud process review, contact us today.