Novosibirsk Researchers Teach an Algorithm to Tell Real Images From AI-Generated Ones
Researchers at Novosibirsk State Technical University have developed a system designed to distinguish real photographs from images created by generative neural networks.
Today, the idea that “seeing is believing” has finally become obsolete. In February 2026, the Bank of Russia officially identified deepfakes as one of the leading threats to individuals and businesses. Criminals use AI to steal identities, bypass biometric security and extort money. Against that backdrop, academic research is fighting back. Novosibirsk State Technical University (NSTU) has introduced a method for automatically detecting synthetic images.
Frequency “Fingerprints” and a Sharp-Eyed Transformer
The approach is based on a simple premise: neural networks that generate images leave artifacts invisible to the human eye in the form of hidden frequency “fingerprints.” The detector’s job is to read them.
NSTU’s innovation is its use of the Vision Transformer (ViT) architecture. Unlike conventional algorithms that scan an image in local patches, ViT analyzes the entire image, mapping complex mathematical relationships among pixels. Tests of the compact ViTTiny model showed 89% accuracy.
One point matters: this is still a research prototype, not a commercial product. But the value of the approach lies in its methodology: the model does not require enormous volumes of labeled data for training, a critical advantage when high-quality datasets are scarce.
An Arms Race: Who Fools Whom?
The Novosibirsk research fits into a global trend. The market for anti-fraud and fact-checking technologies is booming. As early as 2024, Honor began integrating deepfake detectors directly into smartphones. Russian IT companies followed in 2025: Kontur.Tolk introduced video-stream analysis, RVB launched a user-facing service with a claimed accuracy of 95%, and MTS Web Services reported a 98% detection rate.
Experts caution, however, that these accuracy figures cannot be compared directly. Each company tests its algorithms on different datasets and against different generative models.
In parallel, researchers are working on a kind of “preventive medicine” for the internet. The Ivannikov Institute for System Programming of the Russian Academy of Sciences is developing digital watermarks for labeling AI-generated content. Ideally, the future lies in combining the two approaches: invisible markers embedded during generation and independent detectors on the user side.
From Digital Forensics to BRICS Exports
Where could the Siberian technology be used? Digital forensics and bank anti-fraud systems are the most immediate applications. Imagine your “boss” calls and urgently asks you to transfer money, sending a photo of a document or a selfie as proof. A hybrid detector that NSTU researchers plan to develop could identify the fake in a fraction of a second.
The technology is also needed by news organizations to verify images, by social networks for automated moderation, and by digital services to protect copyrights.
Its export potential remains an open question. NSTU has announced no direct contracts, but demand exists in this niche. Russian company Smart Engines is already successfully marketing its anti-fraud technologies in CIS, BRICS and MENA countries. After undergoing an independent audit, the Siberian technology could potentially become part of this exportable layer of protection.
The Eternal Contest Between Shield and Sword
The main conclusion researchers draw is straightforward: this is an endless race. Generative models are improving at the same pace as the tools designed to expose them.
The value of NSTU’s work lies not in the headline 89% figure, but in the foundation it establishes. The researchers’ next step is to develop hybrid detectors that remain robust after compression, cropping and post-processing in messaging apps. The winner in this contest will not be whoever builds a detector for one specific neural network, but whoever can teach an algorithm to adapt quickly to new threats.
For now, research is giving society a new shield. The question is how quickly it can move from university laboratories into our smartphones and banking apps.
