Sumsub has released an upgraded Adaptive Deepfake Detector that uses online self-learning updates to respond more quickly to new deepfake fraud tactics.
The upgraded system replaces static, rule-based detection with a continuously retrained architecture that adapts to emerging deepfake techniques in near-real time. Rather than relying on periodic manual updates, the detector incorporates signals from across Sumsub’s verification pipeline to improve accuracy as new attack vectors appear.
The multi-layered learning approach works across several detection dimensions. When device signals flag a session as potentially fraudulent, the liveness check is re-examined. If a deepfake passed detection, the system logs it as a false negative and uses that signal to retrain. Similarly, when document fraud correlates with a face that passed deepfake screening, the detector gains another data point for continuous improvement.
Nikita Marshalkin, Head of Machine Learning at Sumsub, said the threat landscape now demands next-generation fraud prevention models, noting that modern deepfakes can no longer be detected by the human eye and that decision-making should be based on multiple signal analysis in real time. The company says the architecture supports continuous retraining, adaptive model updates, and real-time risk scoring rather than fixed detection rules.
Sumsub, which recently partnered with crypto exchange MEXC to combat AI-driven fraud, processes identity verifications across financial services, cryptocurrency, and regulated industries. The company’s annual fraud report documented a shift toward complex multi-step schemes in 2025 and predicted a surge in agentic AI scams for 2026.
Other vendors have also moved toward adaptive fraud detection. Fingerprint recently upgraded its Suspect Score with adaptive AI-based detection, reflecting a broader industry consensus that static anti-fraud models cannot keep pace with generative AI-enabled attack methods.
Sources: Sumsub Newsroom
–
By the ID Tech Editorial Team
