The real cost of crypto fraud may be seven times what official figures show. On July 29, 2026, the Consumer Federation of America put 2025 crypto-linked losses at roughly $80.7 billion, scaling the $11.37 billion in complaints the FBI’s Internet Crime Complaint Center actually logged by a 7.1x under-reporting multiplier drawn from federal survey data on how rarely fraud victims come forward.
The sevenfold difference between reported and real is exactly the kind of hidden pattern Dr. Anton Korshunov has spent nearly two decades learning to surface. His doctoral research at the Ivannikov Institute for System Programming (ISP RAS), a leading Russian research institution, mapped hidden community patterns inside online social graphs, and his earlier applied work included identifying concealed user attributes and tracing how influence spreads through message streams. Today, he continues to conduct research in his areas of specialization, social network analysis, graph theory, and natural language processing, in collaboration with leading U.S. universities, while also developing Quantum.uk, a cryptocurrency platform that he plans to bring to the international financial market. In addition, he actively participates in professional conferences and seminars and is involved in teaching and volunteer activities in the fields of artificial intelligence and blockchain technology. We spoke with him about what network analysis can reveal about how fraudulent operations are structured and what this means for anyone deciding whether to trust an online investment group.
Dr. Korshunov, part of your published research at ISP RAS examined how influence moves from person to person across a network, the same mechanism a fake investment community exploits long before it asks for money. Why does that growth curve keep bending upward, even as public warnings get louder?
Awareness campaigns usually describe the outcome, the fake exchange, the frozen withdrawal, and almost never the process that gets you there, which is where the real damage is done, because by the time the outcome arrives, the psychological work is already finished. What mattered most in my research was the sequence of trusted-looking contacts that delivered a message, more than the message itself, since people do not evaluate information in a vacuum; they evaluate the person it came from, and scammers understood that long before academics wrote it down.
A scam group rarely approaches you cold. It builds a small, consistent presence first: a few active members, repeated interactions, a shared vocabulary, long before any investment pitch shows up, and that is the same pattern legitimate online communities form on their own, which is exactly what makes the whole thing work. So telling people to distrust strangers online does not help much when the group has spent weeks making itself stop feeling like strangers.
Some of your own applied work personalized content from users’ behavioral signals. This year the FBI counted AI-enabled crime separately for the first time, logging $893 million in losses. What changes when generative AI picks up the tools scams once needed whole teams to run?
Mostly, it compresses the timeline: convincing fake communities used to need a whole team managing dozens of accounts around the clock, which put a floor under how many operations could run at once, and now language models and voice-cloning tools let a much smaller group produce that same volume of personalized, believable interaction without breaking a sweat. Some of my earlier work was about personalizing content from behavioral signals, and the same technique, pointed at you without your consent, lets a scam address you by name, bring up a past conversation, adjust its pitch on the spot. The mechanics of persuasion have not changed. What collapsed is the labor cost of running them at scale.
Your doctoral thesis on “Community Structure Analysis of Online Social Network Graphs” was essentially a study of what holds an online community together mathematically. From that standpoint, how does a scam operation stay coherent enough to feel like one consistent community to a victim?
That coherence is manufactured, and it is one of the more detectable features of these networks. A real community forms unevenly, right? Some users are highly active, most stay peripheral, and the connections between them reflect genuine, varied interaction, while a scripted operation does the opposite: near-identical response times, the same phrasing surfacing across supposedly different members, interaction patterns that cluster too tightly. Graph-based anomaly detection, the kind used in academic community-structure research, is built to catch exactly that artificial uniformity, well before you send any money.
Law enforcement’s recent gains have come largely from reaching people before they pay: the FBI’s Operation Level Up says it has notified 8,000 targets and prevented some $500 million in losses that way. Your ISP RAS research explored identifying newsworthy events inside message streams in near real time. Does early-signal detection beat flagging the transaction after the fact?
It works because it interrupts the relationship-building stage, not the transaction stage. By the time you are ready to send funds, you have usually spent weeks inside what feels like a real community, and a warning at that point is competing with real social investment, which it mostly loses. Reaching you earlier, while that network around you is still forming, is far more effective, and that is why identifying the network early matters more than catching any single transfer.
You spent seven years as a research scientist at ISP RAS and have continued your research while also building your own business and developing a cryptocurrency platform. Does combining scientific research with entrepreneurship give you a different perspective on where responsibility lies?
Definitely, it sharpens it, because detection research gets to identify patterns after the fact, publish, and move on, while a platform has to decide in real time, with money on the line, whether this specific transaction goes through or gets held, and there is no reviewer you can appeal to when you get it wrong. Licensing and verification requirements exist exactly so that platforms carry a baseline of accountability before they start making those calls.
That is why Quantum.uk’s operations are paused right now, while we go through a new licensing process, and I will not pretend that is a comfortable position for a business, because it is not. But speaking as a researcher, I would rather the industry treat compliance friction as a legitimate cost of cutting down fraud.
Some of your graph-detection methods were patented. For a reader with ten minutes to spare, what should they check before trusting an investment group?
Three things are worth those ten minutes. Check how long the group or profile has existed and whether its activity has been continuous or arrived in one sudden burst. See whether other members interact with each other in ways that predate you joining, because genuine communities have history you were not part of. And be skeptical of answers about withdrawing funds that come back fast and polished, since that kind of responsiveness shows up far more in scripted operations than in organic ones, where answers tend to be slower and less uniform.
Where do you see the contest between AI-generated scam networks and detection research heading over the next few years?
Toward the same arms race you see across network security: attackers automate what used to take manual effort, and defenders automate detection of the traces that automation leaves behind. AI personas mimic individual conversations well, and they are still bad at faking the messy, uneven way real communities grow over time. That weakness is where detection will focus next.
How do you see the next stage of your career, and what are your plans for continuing your research?
The research doesn’t stop, and neither does collaboration with U.S. universities, since the best open problems appear where academic detection methods meet production data. Once Quantum.uk’s relicensing is behind us, the priority is rebuilding the platform with compliance built in from the start, not added later. I’ll also stay visible in the field: Fintech Xchange, HumanX, The AI Summit New York, WAGMI, and ASONAM, the one closest to my training, plus teaching and volunteering in AI and blockchain. The throughline stays the same: understand how a network’s structure forms, then build for how easily it can be faked.
Featured image credit
