The European Union Agency for Cybersecurity, ENISA, has published its good practice guide for buyers of remote identity verification (IDV) services, dated 12 March 2024. The guide does cover some of the important factors to consider, but it fails to take into account the next generational technology that leading remote identity providers are now selling in this space.
The result? Sadly, from the very date of its publication, ENISA’s guide is out of date.
For this brief response, we will focus on just two conclusions of ENISA’s guide:
- That liveness tests need to involve a movement by the user so that an “active” assessment of their liveness can be made; and
- That humans are required to review documents for fraud attempts as well as using automated fraud detection systems.
How active does an active liveness test need to be?
The guide is correct to conclude that a good liveness test needs to include both passive and active elements. But in the most modern liveness solutions, the lines between active and passive are being blurred from a UX perspective.
The guide gives examples of what active tests look like: movement of the head, pressing in the skin etc. However, modern liveness tests are able to perform both active and passive tests with little to no movement being required from the user.
We would not want a buyer to read the guide and think that they can only buy a solution with large movements in the liveness section. And we certainly would not want a buyer to think that just because the liveness test has these movements it is safe against fraudsters.
Micro movements are sufficient
At IDVerse, our liveness detection needs only micro movements for the active element of its checks. Currently, we simply ask the user to smile into the camera—and in a future version we will not even need the user to smile.
To the user it looks like they are taking a selfie, but in fact we are examining micro movements before and after the selfie, and those movements are enough for us to make an “active” determination of their liveness. The technology works with very little movement.
Our liveness is certified to ISO 30107-3 Presentation Attack Detection Level 2 by both iBeta and BixeLab, NIST accredited biometric testing labs.
It boils down to results
The guide should focus on the outcome of the liveness tests, rather than one part of the possible liveness tests contained within the liveness solution. Leading liveness detections look at skin texture, blood flow, lighting across the face and the background, depth and micro movements—to name but a few factors. The outcome is that they can detect all types of attacks from paper printouts, to advanced silicon masks, and also sophisticated deep fakes videos.
Further, what the ENISA good practice guide fails to consider is that mandating active movements discriminates against those who cannot easily perform these checks either because of a disability, age, or lower cognitive abilities.
Do we really still need humans?
The guide claims that human reviewers can spot fraud attempts which automated systems alone cannot. We do not agree with this statement, as our fully automated fraud solution detects all types of fakes.
We do not think that the guide takes into account the modern tools of fraudsters; deep fake documents made by generative AI. We are tracking over 110 websites that offer deep fake documents from under $10 that can defeat all manual reviewers.
We advised Interpol of this increasing risk as they look to provide further guidance in policing these new sites which have been established since the publication of large language models such as OpenAI.
Defending against deep fakes
The ENISA guide does recognise the threat of deep fakes, but claims it will take five years to evaluate the technology on the market that counters it. We would encourage ENISA to test our technology and compare it to human reviewers.
Our clients, including global retail banks, have tested our technology against the latest and greatest fakes, including deep fakes, and have found that its next-generation technology can detect all types of fakes whereas human reviewers simply cannot.
Whilst it is true that GANs (generative adversarial networks) are being used to produce deep fakes, they are also being used to train AI models that can detect these deep fakes. IDVerse has been using GANs for over three years to train our market-leading fraud detection tools.
We have staff constantly scanning the dark web for the latest frauds, and if we find new attacks we can retrain our algorithms very quickly to respond to the threat. We are using AI to detect AI frauds. It requires AI to detect AI.
Keeping up with the tech
I again emphasize that the guide should focus on the outcomes of the solutions, rather than the methods used. Instead of recommending (with no evidence) that manual reviewers are still required, the guide could recommend that buyers test the solutions with their own deep fakes and other fraudulent artifacts.
The framework should aim to follow a technology-agnostic approach, where possible, to be able to keep pace with the rapidly evolving technical landscape. This will enable organizations to embrace new technologies while upholding human rights and responding to ethical risks.
Consulting the experts
We invite ENISA to spend some time with our AI engineers to understand the advancements in remote identity detection technology over the past few years, and what the future in this area holds.
We would happily help ENISA ensure its next edition of the guide is up to date and reflects the latest and greatest technologies.
About the post:
Images are generative AI-created. Prompt: A brightly colorful map of Europe comprised of various human faces of different ethnicities. Tool: Midjourney.
About the author:
Peter Violaris is Global DPO and Head of Legal EMEA for IDVerse. Peter is a commercial technology lawyer with a particular focus on biometrics, privacy, and AI learning. Peter has been in the identity space for 6 years and before that worked for London law firms.
