Sumsub's Adaptive Deepfake Detection Aims to Outpace Fraud Arms Race

Event summary

  • Sumsub launched 'Adaptive Deepfake Detector,' a new fraud prevention model featuring instant online self-learning updates.
  • The share of multi-step fraud attacks increased by 180% in 2025, reaching 28% of all fraud detected by Sumsub globally.
  • Traditional deepfake detection methods are failing due to the speed of evolving fraud techniques, with a gap between model updates lasting weeks or months.
  • Sumsub's new detector aims to adapt to emerging threats within hours, leveraging multiple data layers beyond visual content.
  • Nikita Marshalkin, Head of Machine Learning at Sumsub, emphasizes the need for real-time, multi-signal analysis in fraud prevention.

The big picture

The rapid evolution of AI-generated deepfakes is creating a significant challenge for digital businesses, forcing a shift from reactive, periodic fraud prevention to proactive, real-time solutions. Sumsub's Adaptive Deepfake Detector represents a move towards a more dynamic security posture, but the ongoing arms race between fraudsters and security teams will require continuous innovation and investment. The increasing prevalence of multi-step attacks underscores the need for holistic verification approaches that extend beyond simple visual checks.

What we're watching

Adoption Rate
The speed at which Sumsub's existing client base adopts the Adaptive Deepfake Detector will be a key indicator of its perceived value and impact on fraud reduction.
Competitive Response
Other verification platform providers will likely accelerate their own real-time adaptation capabilities, potentially leading to a price war or feature parity.
Efficacy
The long-term effectiveness of the 'online learning' model will depend on its ability to stay ahead of increasingly sophisticated fraud techniques and avoid false positives.

Related topics

Technology