Sumsub has launched an Adaptive Deepfake Detector that uses continuous online machine learning to identify emerging fraud patterns within hours rather than the weeks or months required by conventional periodic-update models. The product targets a structural weakness in existing deepfake detection systems, where gaps between scheduled model updates leave enterprises exposed to novel attack methods.

How the self-learning model works

Unlike traditional detection tools that rely on periodic retraining cycles, Sumsub’s system continuously learns from fraud signals across multiple layers — including documents, geolocation, IP address, device signals, and facial biometrics. The multilayered approach means the detector does not rely solely on visual content inspection; it also monitors injection methods used by fraudsters to bypass liveness checks, with the model’s decision boundary adjusting automatically as new threats are observed.

The company says this approach pushes average detection accuracy close to 100%, with no manual retraining required and no waiting period for scheduled update cycles.

Fraud escalation driving urgency

Sumsub’s own platform data shows the share of multi-step attacks increased by 180% in 2025, reaching 28% of all fraud detected globally. The company notes that AI-generated deepfake activity has grown steadily since 2023, with no signs of slowing across key Asia Pacific markets.

“In 2026, the threat landscape has evolved, demanding risk management teams to respond with 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.” — Nikita Marshalkin, Head of Machine Learning, Sumsub

Sumsub serves over 4,000 clients across financial services, gaming, transport, and other regulated sectors. The company is recognised as a Leader by Gartner, Forrester, and IDC in the identity verification and fraud prevention space.