Thank you so much for having me here. Good afternoon. What you're seeing right now is not Martin Kuppinger. This is a deepfake created by John Erik using widely available technology. No special equipment, no Hollywood budget. And as you can see, it's already hard to tell the difference. Deepfakes can be used to impersonate people during digital identity verification, creating accounts, bypassing checks, authenticating as someone else entirely. That's what John Erik is here to talk about. Over to you, John Erik.
Thank you, Martin. And this is easy to create. It does not take a lot of effort, and this is the problem. I'm going to talk about onboarding. And we see, I'm just going to give you this number, during peak attacks, because the attacks of deepfakes go in waves, during a peak attack, as many as 10% of the verified identities can be a deepfake.
10%, 1 in 10. Okay, a little walk back through identity assurance. It's about two things. Is this a real person? And is that person present? That's the two things we want to establish. And in the good old days, in the village, it was easy.
You know, oh, you're Tom's son, right? People knew each other and trust was easy. It was small. Everybody knew each other. So trust was established that way. But when society grew, we moved to the passport ritual. You needed some sort of document that you proved you exist, and that one had a picture that bound you to you. So you had the trust in the document and that you showed up physically.
Then, of course, you had a human aspect. If you were female in the old days, you needed your husband to come with you. If you had a different color skin, you would look down on, etc. So you had a lot of bias and so on in there. But this also shows vouching for someone else. So the husband had to vouch for his wife. Then we had early internet banking. And I guess a lot of you remember that, or have you forgotten already, where you signed up on a form and they sent you a letter to your mailbox and they asked you some questions. We called it knowledge-based authentication.
So if you know this information, you must be you, right? Typically in Norway, where I come from, you were asked for your national identification number. If you knew those 11 digits, well, you had to be you, right?
Well, it worked sort of. I mean, you onboarded people in that area. And this was Kambushama. It took a long time. So what happened in the Nordics, again, was the bank ID disappeared. They were reusable. You onboarded once to that bank ID and you could reuse that. And it was very convenient, because then you didn't have to go through that process every time to get that letter and to ask all these questions and so on. So where are we now? Now we are scanning an identity document and we're taking a selfie. That's pretty much the onboarding process we're doing right now.
So we're saying you scan the identity document. That's proved that I really exist. That's issued by some trusted entity, some government. It proves I exist. I take a selfie. We compare the picture we pull, typically from the NFC chip in your document, with the selfie. And if they compare, well, the user is present. So we have proof that the user exists by the document, proof that the user is present by this comparison.
However, this introduces friction. People have to do all this stuff. And banks lose clients due to this friction, right? So they focused on, you know, making this faster, less friction. They wanted higher completion, higher rates of completion of this. But at what cost? What about those that made it through here?
I still, and those who have seen me before, have seen this many times. This is 1993. This is more than 30 years ago. He said on the Internet, nobody knows you're a dog. A couple of years back, I added this one.
With AI, you can be anyone on the Internet. And that is so true, and that's also what Peter just talked about, right? You can be anyone in a call like that. And then with AI, you can deepfake the document, and you can deepfake the person. And what can you really trust? And is this a real problem?
Well, this case from the Netherlands. Yes, it's a real problem. This is just one case of showing this. He opened 46 bank accounts in this. When they investigated, they found that he had been using chattypt to learn how to do this. So he didn't have any specific skills or anything, but he used that technology. And how was he caught? He was caught by his own stupid mistake. He used a deepfake of a male on a document that had a female name. And that was caught by a random check by a human being in the bank. So he wasn't caught by the automatic processes or anything.
All the deepfakes went through here without any problems. 46 bank accounts. As I mentioned, attack fluctuate. You have peaks of attacks going on here, and in peak times up to 10% of your identities may be fake. And that should really have you worried. And we don't know when these attacks hit. I'd like to mention another kind of attack that's sort of different, but it's important anyway. That's the morphing. In most countries, when you go to issue your identity document, you bring your own photo. I live in one of the very few countries where the police actually take my photo when I'm there.
So it's a genuine photo. But in most places you can bring your own and you can have a morphed photo. You can put four, maybe ten people, you can morph into the same photo. So you have several people that can use the same identity document. This is another attack vector that is difficult to catch without checking for deepfakes. So you can pass all the checks and balances on this and still you may not exist. And just to talk on the identity wallet, since we are here on an identity wallet conference. If you cannot trust the identity, you cannot trust any of the attributes.
I should have a countdown timer, it's still set 20 on here, so I have a lot of time left, I guess. Can you give me a hint?
Okay, good, yeah. Okay, now I got 14, 15.
Okay, so if you cannot trust the identity, you cannot trust any of the attributes either, right? It all depends on that trust of the identity that the wallet is owned by someone and you can trust that identity. And we're basing everything on that. And if it fails on the onboarding step, then we have a problem. A report from Kapinko and Bailey said, you know, it's an arms race. We're getting continuously better tools for generating AIs and the attacks are getting much more realistic and more sophisticated.
I mean, the deepfake I started with Martin in the beginning, it wasn't very good. I mean, you could pick up, there were some peaks and stuff there. The reason for that, I picked it out from one of those webinars and I didn't find a long enough video of him, so there were clips in there. But I mean, it's easy to improve on that. If I spent some more time, I could have made one that was very good. But the attacks are coming better and it's about weaponization of AI for deepfakes. Gartner also say that they're working with vendors.
You need to work with vendors that are specialized in these deepfake attacks. You need, and as Peter mentioned, it's moving so fast. It's changing all the time. You have to be on top of this every time. And Gartner predicts that 30% of enterprises will consider unreliable due to AI by 2026. That's now. This was two years ago. Rhetorical question. We are there already. Deepfakes are so realistic now, they can fool anything. Just one example.
I mean, three years ago, it took weeks to generate something I made here on Martin. I mean, it took me just, I think I spent half an hour creating Martin here, the whole thing. Now you can do it in minutes. The cost was enormous. Now it doesn't cost hardly anything and you don't need the skills.
As I said, the guy that attacked the bank in the Netherlands, he just asked Chattopadhyay, how can I do this? Last quarter of 2025, 55 new synthetic metagenders were graced. Almost one every second day, there was a new generator for synthetic media. This should have us worried. Hugging Face, which is an AI community, have more than 95,000 models for generating video and images. This is what we're up against. Dr. Guus in Lelles, they did a stress test the first half of last year. So this is some time ago, but still.
Out of 60 million identity checks, there were 500,000 synthetic identities that were blocked. OK, and then they claim, well, we have liveness detection, so that's going to get it, right? No. There are three layers here and we need to consider all layers. One is the presentation attack detection. That's where you do the liveness check that it's a real person sitting there. Then you have the injection attack where you inject the video stream. That's what you showed by using OBS, injecting a new video stream. And of course, there is software on phones now that makes this difficult, not impossible.
And you need deepfake detection to analyze the data stream to check for deepfakes. So is a live person present?
Well, the pad will detect that. It looks like a live person. Did it come from a real camera? The injection attack will supposedly attack that. And is the content itself genuine? That's the deepfake detection. The challenge is the speed at which things are moving. The first two things are typically happening in the app on the phone. You need to have something installed on your device to do that. That means when a new attack appears, it's going to take a complete release cycle before you get those updated, right? While the deepfake attacks typically takes place in server.
When you identify a new attack, you add it and it's available immediately. So you need all of these three. Then what we typically do when there's a suspicious flag, we send it to this poor guy. That's what we traditionally have done, right? We send it to a human because we assume, okay, a human can see this a lot better.
Well, poor guy. I mean, we are at the place now. It's impossible to see the deepfakes. There are so many good examples of that. So this is a challenge. And also this is made a challenge because this is a service. Crime as a service. You can buy everything you need. You don't need to know anything. You go to the marketplace. This is like going to Amazon. They have recommendations. They have reviews. They have escrows. They even have money-back guarantees if it doesn't work. You can go and buy this stuff. You don't really need to know anything.
You just need to have an idea how you want to do that. And this includes also the money laundering when you actually manage to get something out of it. Perfect. Kaspersky found that there were about 47 tools specifically designed to bypass KYC. Sold for as little as five to twenty dollars. Verifuse is like a tool set for doing this. And Thomas from Bycatch said, you know, the assumption is that if the camera is live, the user is there. But with AI, that is not true anymore. And I especially love this one. Bypass KYC. Educational use for education only.
Yeah, right. This is the tool. It's not on the dark web. You can look it up. It's available. You can buy the source code or you can subscribe for a monthly price. And they will do the deepfake attacks on the KYC process for you. And this is what we're up against.
Then, how aware are organizations about this problem? Well, how many organizations have deepfake protocols? 13. Have you heard of deepfakes? Only 22% of the people that answered to this survey had even heard about it. And on the executive level, are you familiar with deepfakes? Only 25%. So we do have a problem. I would hope that with banks, this is higher, but we don't know. So we are in a deepfakes arms race. Peter showed a good timeline on this. This is just a different way of showing that. We have already passed the point where we have to have deepfake detection as part of the solution.
Because the attacks are so good. The criminals are so professional. And there are so many tools available that are moving all the time. So you really need to have some good tools in there. So Gartner said deepfake detection will become a standard non-negotiable feature. This was in December. I would say it should already be that available. And then consider that building this is hard or impossible. You need to leave these to the experts that really know what they're doing in the space. This is really complex. It needs continuous monitoring because there's so much stuff happening.
Development is moving so fast. You really need to get on top of it. And you need to do continuous updates. All the time you need to update the solution. And even though you have a solution that's good enough today, well, it's going to be insufficient when the next wave comes. And it's also difficult to prove how good it is. We don't have any standards for deepfake detection because everything is moving so fast. So you're dependent on knowing how good you are at detecting this current wave and building it from there. So that was my story. Thank you. Thank you very much, Jan-Erik.
So do we have any questions? Yes, there's a question. Thank you for the presentation. I have a question. What early indicators or anomalies can we look or detect to prevent the rise in the graph that you show us? How to detect that?
I mean, in the device or maybe, I don't know, in the business indicator? I mean, I guess the business indicator is, and of course that's a slow one, knowing when you're getting a lot of fake accounts, a bank is getting a lot of fake accounts, but often you don't know that until much later, right? So this is about monitoring what's going on, what's being released, which new models are available, testing them against the deepfake solution.
So again, this is what requires the very specific expertise on deepfakes and understanding it and building them deepfake detection models continuously. Yes. So I have a question. Do you believe that the banks are aware of this? I think they are aware. I think a lot of them are still underestimating the consequences of this. And as I showed in an early slide, they're afraid of too much dropout. It's important to have enough customers signed up, etc. And that's often more important. And this one doesn't show up until much later, maybe when you're rolling up something.
So, yeah, I think awareness is there. The banks are also so busy with, you know, this tsunami of regulations coming in now that they need to add, you know, the new regulations that they need to comply with, including EIDIS, of course, that maybe this is not high enough up on their radar. Thank you. Yes. I asked that question because personally, my bank decided to have a courtesy call with me via Teams, which I thought was very good. Right. So I then phoned them up and said, I don't believe that it's going to be you.
So here via this safe phone number, I have given you a second factor, which is I will only believe it if you tell me this word, whereupon there was complete amazement that this might be a problem. Right. And I wondered how common you think that is.
I mean, the problem is that, you know, everything we've added so far of identification, because that's what you're touching, is one way. I need to prove who I am towards the bank. If I ask the bank, well, can you prove that you are really my bank? They don't have a way to do that. Mutual authentication. Right. And that's one of the things I think the wallet is going to solve, which is part of my next presentation later on today. OK. Just come and see me talking about the EIDIS. Don't forget to go to your, when is it this afternoon? It's in B5. B5. B there or B square. Something like that. Yes.
All right. Thanks for joining, everyone. Thank you very much.
