Document fraud: adapting to evolving fraud tactics

How Veriff detects document fraud

Let’s explore the methods for detecting document fraud. Subtle indicators, such as mismatched fonts or incorrect borders, can signal a fraudulent document. Paco Romero Ferrero, a data scientist in Verification Automation at Veriff, emphasizes the crucial role of collaboration between the Fraud Operations Engineering Teams in staying ahead of evolving fraud tactics. By monitoring trends, annotating fraud cases, and regularly updating machine learning models, these teams enhance detection capabilities. 

Useful links:

Onboard more genuine customers with Veriff's Identity and Document Verification solution. It's proven to deliver speed, convenience, and low friction for your users resulting in high conversion rates, fraud mitigation, and operational efficiency for your business. Learn more.

Do you want to be part of the push to bring accountability to the online world, creating a safe space of innovation and inspiration, like Paco? Find your role with Veriff.

Document fraud detection in detail

Veriff's AI model detects fraud by aggregating multiple signals from documents, such as the New York State ID shown here. Our well-trained model, built on a large dataset, enhances accuracy. Unlike "black box" models, this approach offers transparency, allowing us to pinpoint which specific signals triggered a fraud alert. This transparency not only improves model performance but also aids fraud operations teams in focusing on critical issues, fostering a productive collaboration between AI and humans.

Useful links:
Veriff’s SVP of Operations Mike O’Callaghan talks more about ‘Widgets’, ‘black boxes’, and ‘bionics’ in operations at Veriff, in this recent Veriff Voices episode.

In this blog, we take a closer look at why it’s important to unlock the potential of both AI and human intelligence to provide the best possible identity verification solutions to customers.

The signals Veriff uses to detect fraudulent documents

We use a range of signals to detect fraudulent documents, including analyzing the document’s background, shape, and borders to ensure they align with known specimens. Our specialized models are trained to spot anomalies in portraits, fonts, and characters, such as unintended text like "specimen" that signal fraudulent documents. Our system also cross-references information like serial numbers, issue dates, and barcodes to ensure consistency. These methods help identify potential fraud by catching subtle irregularities that might otherwise go unnoticed.

Useful links:

We believe a key pillar to developing a sophisticated anti-fraud strategy is crosslinking. Learn more in this blog about how we enable our fraud prevention teams to look for patterns across multiple sessions.

Detecting fraudulent signatures

Our AI system evaluates approximately 2,000 different signals, including signatures, fonts, and portraits, to assess a document’s authenticity. For instance, the model detects fraudulent signatures using a technique called embedding, which compares a signature to a database of known frauds. If the similarity is high, the document is flagged as fraudulent. The transparency of our model fosters trust by revealing which signals were pivotal in the decision-making process, ensuring both accuracy and reliability in fraud detection.

While significant progress has been made, there are still areas for improvement. We will continue to explore the use of embeddings to analyze additional features of documents beyond signatures. We aim to incorporate validation from videos and registries to enhance accuracy with the ultimate goal being to ensure that the person being verified is truly who they claim to be.