Sumsub launches Adaptive Deepfake Detector for fraud

Sumsub has launched Adaptive Deepfake Detector.

The new model is designed to identify emerging deepfake scams faster than conventional systems.

The release comes as fraud and compliance teams face growing pressure from AI-generated deception that is increasingly difficult to spot through visual checks alone. Newer attacks often combine manipulated images, voices or video with other tactics that can evade tools reliant on periodic updates.

According to company figures, the share of multi-step attacks rose by 180% in 2025 and accounted for 28% of all fraud detected on Sumsub's platform worldwide. That trend has increased the need for systems that can adapt more quickly as fraud patterns change.

Traditional offline models are often updated on fixed cycles, creating a gap of weeks or months between retraining and deployment. During that time, newly developed attack methods can slip through controls and affect app users and businesses.

Sumsub said its new system uses online learning, allowing the model to update itself as new fraud signals appear instead of waiting for scheduled training cycles. This approach lets the detector adapt within hours, according to the company.

How it works

The model draws on signals from several sources rather than relying on a single anomaly. These include documents, geolocation, IP address, device data, facial biometrics and verification information checked across multiple users to identify suspected fraudulent network activity.

This broader approach reflects a shift in attack methods. Fraudsters are not only creating synthetic media but also using injection techniques that add a separate layer of data for prevention systems to assess.

With each new observation, the model adjusts its parameters without manual retraining, according to Sumsub. The detector's decision boundary shifts as threats evolve, pushing average detection accuracy close to 100%, it said.

Nikita Marshalkin, Head of Machine Learning at Sumsub, said the threat environment had changed sharply.

"In 2026, the threat landscape has evolved, demanding risk management teams to respond with the next-generation fraud prevention models. Modern deepfakes can no longer be detected by the human eye, and decision-making should be based on multiple signal analysis in real time," said Nikita Marshalkin, Head of Machine Learning at Sumsub.

He said the tool is intended to sit within a broader fraud screening framework.

"That's why we launched our upgraded Deepfake Detector, offering clients not just a tool, but rather an online learning system that combines advanced document checks, device intelligence, and fraudulent networks analysis to complement deepfake detection capabilities. When the price of failure is too high, a comprehensive approach to the increasing AI-driven fraud challenge is the answer we need," said Marshalkin.

Fraud pressure

The launch comes as digital fraud prevention providers face pressure to respond to increasingly sophisticated attacks across financial services, consumer platforms and other online sectors. Deepfake technology has advanced quickly since 2023, widening concerns from isolated identity spoofing to coordinated attacks that exploit weaknesses in onboarding and verification processes.

For companies that rely on remote identity checks, the challenge is no longer limited to spotting manipulated visuals. Risk teams must also assess the broader context of a user session, including device behaviour, location signals and links between accounts.

Sumsub offers identity and business verification, ongoing monitoring and fraud prevention services to more than 4,000 clients, including firms in financial technology, telecoms, travel and gaming. Its customer list includes Bitpanda, Wirex, Avis, Bybit, Vodafone, Duolingo, Kaizen Gaming and TransferGo.

The latest detector was built to reduce reliance on regular human review to keep models current. Instead, it is intended to incorporate new deepfake types and injection methods as soon as they are observed.

That marks a shift from static detection methods to systems that continuously revise risk thresholds as attack patterns change. In a market where fraud techniques can spread quickly, speed of adaptation has become a central issue for suppliers and their customers.

Sumsub said the model continuously learns new patterns, including emerging deepfake types and injection methods, and immediately adds them to its list of known threats.