Deepfakes are cheap, convincing, and widely accessible. For businesses that rely on digital onboarding, they have created a new attack surface in the onboarding process. But do you really know how well your detector works, and for whom?
This session takes a critical look at what deepfake detectors can and cannot do in real-world KYC. Detection performance can vary substantially across fake types, and models that perform well in the lab may struggle with generation methods they have never encountered. As new generation techniques emerge, generalization becomes a central challenge for reliable detection.
There is also a second, less visible risk: detectors do not perform equally across ethnic groups. For identity verification, this means two failures at once. Legitimate customers from some groups are rejected or pushed into costly manual review, while fraud from others slips through. The result is a gap in your fraud prevention, a broken customer journey, and growing compliance and reputational risk.
Attendees will learn:
1) How deepfake detectors perform across fake types and unseen generation methods
2) Where demographic and ethnic bias can emerge in detection, and how unequal performance can translate into customer friction and detection gaps
3) How to evaluate deepfake detectors in practice to build KYC processes that are both secure and fair
