Sumsub released its upgraded deepfake detection solution with instant online self-learning updates, setting a new standard in catching sophisticated fraud online
Sumsub today launches Adaptive Deepfake Detector, a new model that tackles the prevailing issue of traditional offline solutions being incapable of catching the newest deepfake scams. Unlike its predecessors, Sumsub’s deepfake detector effectively spots emerging types of sophisticated fraud thanks to its ML-driven detection tool with instant online self-learning upgrades.
Periodic model updates reveal systemic vulnerability: between the upgrades (which can take weeks or months to launch and be implemented), the new threats manage to bypass the defenses and cause real damage to digital app users and companies. The key differentiator of the new tool offered by Sumsub lies in its detection accuracy, which stems from non-stop model learning from fraud signals across multiple layers, allowing it to adapt within hours, not weeks or months.
The need for such an approach gained urgency in the last couple of years, when it became obvious that the arms race between online fraudsters and security teams entered a new phase of high-level sophistication and speed. In 2025, the share of multi-step attacks soared by 180%, reaching 28% of all fraud detected by the Sumsub platform globally. The growth of AI-generated deepfakes has been noticeable since 2023, and various markets across the globe see no signs of mitigation.
“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. “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”.
In current deepfake detection, risk teams cannot rely solely on the visual content inspection. The full context of the user session should be taken into account. Apart from generating deepfake images, voices or videos, fraudsters also use various injection methods, thus providing a separate data layer for prevention systems to check and monitor.
From the technical standpoint, real-time detection based on the ‘online learning’ model implies no waiting time for scheduled training cycles and no need for regular human review to stay up-to-date. Instead, the new solution:
- Continuously learns new patterns, including emerging deepfake types or injection methods, immediately incorporating them into the known threats list;
- Signals are collected from multiple sources, not on a single anomaly vector. The multilayered fraud detection system analyzes documents, geolocation, IP address, device signals, facial biometrics (liveness) data, and cross-checks verification information from multiple users to spot fraudulent network activity;
- Within each new observation, the model adjusts its parameters with no manual retraining required;
- The detector’s decision boundary shifts to account for evolving threats, pushing the average detection accuracy close to 100%.
To learn more about Sumsub’s Adaptive Deepfake Detector, please go to https://sumsub.com/deepfake-detection/
