KYC Deepfake Attacks: Liveness Alone Isn’t Enough
András Somkuti ·
Remote identity verification was built around a relatively intuitive proposition: if an organisation can examine reliable identity evidence and confirm that the person presenting it is the legitimate holder, much of the confidence once associated with physical presence can be recreated digitally. Generative AI has not invalidated that proposition, although it has made the word “presenting” considerably more complicated.
Identity verification is consequently moving away from the idea that one highly accurate biometric comparison can settle the question. The relevant challenge has become the integrity of the entire capture and verification environment.
Public discussion naturally gravitates towards deepfakes because they are visually striking. From the perspective of identity proofing, however, a convincing synthetic face is only one possible attack.
A presentation attack may involve a printed photograph, replayed video, screen presentation, mask or another artefact physically shown to a biometric sensor. Presentation Attack Detection, or PAD, attempts to distinguish these manipulations from the genuine presence of a person.
Modern guidance increasingly treats digital injection separately. In an injection attack, manipulated media can enter the process between the original capture point and the component performing the comparison.
The distinction is important. If a fabricated video is inserted directly into a digital stream, the attacker may not need to fool the camera at all.
NIST’s current Digital Identity Guidelines explicitly address both presentation attacks and forged-media or injection attacks, noting that remote identity proofing is vulnerable at several points and that biometric comparison alone cannot prevent those forms of manipulation.
Liveness is valuable, but the term is often asked to carry too much weight
“Liveness detection” has become one of the most familiar phrases in remote onboarding. In practice, it belongs to the broader discipline of presentation-attack detection, where systems examine signals intended to establish whether the biometric sample comes from a real person present at the point of capture.
Some techniques are active, requiring an action from the applicant, while others analyse passive characteristics. The sophistication of the attacker changes the value of individual signals, however. A control designed to distinguish a photograph from a live face addresses a different threat from an injected synthetic video that was generated specifically to satisfy expected liveness behaviours.
Consequently, the meaningful question is not whether a KYC product has “liveness”. It is what attack model the verification architecture assumes and how several independent controls interact when one of them is defeated.
Fraud detection is becoming contextual
A technically convincing identity session can still contain behavioural anomalies. Just a few examples: the same device may appear across unrelated applications, network characteristics may contradict the claimed location, repeated attempts may show systematic variation, or transaction patterns may indicate an organised attack rather than an individual applicant.
These signals do not establish identity by themselves, and over-reliance on opaque risk scoring introduces its own problems. They nevertheless provide context that a single face comparison cannot. The emerging model resembles layered fraud analysis in other security disciplines: each control contributes evidence, while confidence comes from the combined pattern.
The future is not a perfect deepfake detector
There is an appealing idea that identity providers will eventually build a detector capable of reliably labelling every synthetic face, fabricated document or manipulated video, but that expectation misunderstands the nature of adversarial technology. As generation improves, detection improves as well; attackers then alter their methods, and detection systems adapt in response. This cycle continues indefinitely.
A more resilient architecture assumes that no single classifier will remain decisive indefinitely. Strong identity proofing therefore depends on several independent relationships being difficult to counterfeit at the same time: the person and their document, the document and authoritative information, the biometric sample and the physical capture environment, the device and the session, and the session and the broader behavioural context.
Deepfakes have changed KYC, but perhaps in a useful way. They have forced the industry to confront something that was always true: identity cannot be reduced to whether two images look alike.
See how Netlock verifies identity before it issues a qualified certificate: video identification on any device with a camera, or the Netlock mobile app with a chip identity card. The identification is automatic, and a member of our staff checks it when needed. If you need to verify your own customers or companies, KYC and KYB verification are offered as a service, and AML screening is available too. Sign up for free and start with 15 free signatures.