Webinar Q&A: Gen AI fraud & OnlyFake

Yes. This entirely depends on the sophistication of the liveness detection: 

• Static ones—where a person takes a selfie while holding the ID card or a paper with the onboarding date or a provided code—can easily be fooled by readily available Gen AI image generating models. 
• Video verifications can also be fooled, though with slightly more complicated deepfaking technology that overlays and replaces facial features, either on a prerecorded video—if the liveness check follows a static request pattern (look left, up, right, etc)—or by in the video stream itself.
• Depth scanning verifications—think of the iphone’s FaceID, which creates a 3D map of the face—are likely the most difficult to fool. However, unlike the first two options, this isn’t a universally available technology, which dramatically limits.

However, one clever way to bypass all of them is for a fraudster to use their own portrait (or that of a paid mule) on the ID, since the liveness will then perfectly match their face.

We've come across multiple different tactics, including: 

• Replacing the portrait on the ID to match the person who will be passing the liveness check
• Modifying the face of the person passing the liveness check to match the photo on the ID, either in a live video stream or on a static selfie.
• Modifying the face on both the ID and the person undergoing the liveness check.

No. Large Language Models can operate in any language. The quality might vary depending on the training of the model, but they are pretty proficient in any language. 

We are document agnostic, meaning we don’t really care what the document is. So long as it’s a document in digital format (PDF or Image), we can analyze it for fraud. That means we can analyze tax forms, business registration documents, bank statements, utility bills, income statements, paystubs, diplomas, etc… We have analyzed over 100M documents from over 130 countries.

Potential regulatory requirements might be to: 

• Ensure there is no discrimination against those who are unbanked, not Plaid-enabled, meaning forcing institutions to accept documents even when they would rather not.
• Ensure that any decisions made about a document, whether it was approved or denied, is clearly recorded for future audit. 
  
  

Resistant AI is document agnostic. That means it works with any document loan underwriters use, from anywhere—even those we've never seen before. That includes tax forms and other tax documents, bank statements, pay stubs, account statements, IDs, passports, invoices, and more.

Resistant AI works with all PDF and image formats. Different formats have different risk profiles, and we can help you set up a policy that matches your risk appetite. 

Yes, we do this kind of retroactive analysis often with our customers—often finding far more fraud than they knew about. This kind of past check is best run through an offline batch upload to our API. Book a meeting to get started.

Not necessarily Generative AI, as its purpose is to generate content, and perhaps bring some contextual understanding of content through its ability to explain things. With the exception of reading spam/scam messages and classifying them as such, the actual predictive and preventative capabilities to stop other types of fraud would be handled by other forms of AI. Generative AI might play a role in explaining the output of those other AIs. 

Yes. We work with several institutions that can approve student visa applications and also background check companies, where transcripts and certifications come up. 

While crypto asset platforms are vulnerable, we find that any financial service that allows people to open accounts (banks, payment firms, etc) are particularly at risk to false documentation, as it enables organized criminals to commit many more crimes—from receiving the proceeds of scams and fraud perpetrated elsewhere, to money laundering without the need for or risk from real money mules, building up fake credit scores for bust out frauds, or just reselling those accounts as valuable commodities to other criminals.

Only if poorly made. Today, most regular document fraud done digitally cannot be detected with the human eye. Bad AI is only making the problem far worse. Frankly, the arms race has begun, and there is no going back: if humans even had a slim chance against the quality of fraud produced by AI, they certainly don’t against the scale and speed at which it can be created. And things are only going to get worse as the bad AI improve. They need good AI to back them up.

We offer several different models. We offer several different deployment models, including deployment in partner/customer clouds under specific configurations. If you’d like to discuss in more detail, schedule a quick call with our team.

It largely depends on the individual document and fraud execution. Generally speaking, we have detectors that work on images in all formats, including scans, screenshots, and photos amongst others. We're also able to rank different formats and document types by risk, and help you build an appropriate intake policy that will match your risk appetite. To learn more, book a call.

We don't believe in risk scores: What do you do with a 68% vs an 82% document risk score?

Instead, Resistant AI gives you clear actionable verdicts:

You can either use Resistant AI manually by dropping documents in our web interface, or use our API integration to automate fraud detection at scale.

• In dealing with AI Generated document fraud, you need AI-powered document fraud detection. You need good AI to fight bad AI.
• Taking a layered approach to detecting fraud. The document is just one element of any fraudulent attack. Having other data-driven anomaly detections that look at behaviors, transactional patterns, device biometrics and more will reveal other elements.
• Having a defense in depth approach, where detections on one layer get used to reinforce the other layers. 

Resistant AI analyzes documents in any language. In fact, we don't read the content of the documents to catch fraud—we analyze the whole context of how the document is made instead. Simply put, we look at how the words are written, not what they say. 

By analyzing millions of documents every month, we've found that 1 in 5 documents used in digital loan applications show signs of tampering, and up to 2% of documents are based on reused or generated documents indicative of crowd-sourced first party fraud or organized crime rings.

Yes, you can analyze any and all documents necessary for the lending underwriting process: From tax forms and other tax documents, to bank statements, pay stubs, account statements, IDs, passports, invoices, and more.

It takes less than 20 seconds on average for you to get your verdict when analyzing a document for fraud with Resistant AI.