How Does Deepfake Detection Fit Video KYC, and Can Agents Spot a Deepfake on the Call?

How does deepfake detection fit video KYC?

Deepfake detection fits video KYC at three points: while the call is live, at the moment the agent captures a face snapshot, and after the call on the full recording. Each point answers a slightly different question, and most programmes use at least two of them.

During the call, the platform samples frames and short audio segments and sends them for analysis, so a high-risk score can reach the agent before they approve the customer. At capture, the still image that will be matched against the ID photo gets its own check, because that frame becomes the evidence of record. After the call, the stored recording can be scored again with more context, which helps with quality assurance, disputes and regulatory sampling.

Video KYC differs from a selfie flow because a person is watching. Banks often assume the agent is the deepfake control. The agent is a strong control against replayed video and clumsy fakes, and a weak one against a polished real-time face swap, which is where software earns its place.

Which deepfake attacks reach a video KYC call?

Four attacks show up in video KYC. A real-time face swap maps the attacker's movements onto a target face and streams the result through a virtual camera, so the attacker can answer questions live. A prerecorded deepfake is injected as the camera feed and works only if the questions are predictable. A replay uses genuine footage of the real customer. Synthetic or cloned audio can accompany any of these, or stand in when the customer claims their camera is broken.

The routes matter because the controls differ. Virtual camera injection and emulators are device problems that capture-integrity tooling should catch. Replays are a freshness problem. Face swaps and synthetic faces are a media problem, and that is what deepfake detection measures.

FinCEN's November 2024 alert on deepfake media lists behaviours that agents can see during a live check: the customer uses a third-party webcam plugin, or tries to switch communication method citing excessive or suspicious technical glitches. The alert also lists a photo or video flagged by deepfake detection software as a red flag in its own right.

Are human video KYC agents better at catching deepfakes than software?

Human agents are not reliably better than software at catching deepfakes, and software is not reliably better than humans; the two fail on different cases, which is the argument for using both. Groh and colleagues tested 15,016 participants against a leading computer vision model on deepfake videos and found ordinary observers and the model roughly equally accurate, with different kinds of errors. The study was published in PNAS in January 2022.

The same study found that people who saw the model's prediction became more accurate than either people or the model alone, but inaccurate predictions often pulled their accuracy down. For a video KYC desk, that means the score shown to an agent must be well calibrated, and the agent's procedure should say what to do when their judgement and the score disagree.

Attention is the other weakness. In an iProov consumer study of 2,000 people in the UK and US, published in February 2025, only 0.1% correctly classified every real and synthetic image and video, even though participants were told to look for deepfakes, and they were 36% less likely to identify a synthetic video than a synthetic image. iProov sells biometric verification, so treat the figure as vendor research. Agents handling dozens of calls a shift are less primed than quiz takers.

NIST now treats agent training as a requirement. SP 800-63A-4 says providers shall train proofing agents and trusted referees to look for indications of manipulated media, giving examples such as high latency, synchronisation issues and inconsistent skin tone or resolution. The companion guide on human deepfake detection covers what people can and cannot see.

What questions should video KYC agents ask to expose a deepfake?

The most useful agent prompts ask the customer to do something physical that a face-swap model handles badly, rather than to answer a fact question. The GOTCHA study from NYU, accepted at IEEE Euro S&P 2024, recorded 47 participants performing eight challenges and generated deepfake versions with three pipelines. Measured by how much each challenge degraded the deepfake, the ranking was clear.

Occluding the face with a hand degraded deepfakes most, at 91.4% averaged across generators, followed by manual facial deformation such as pressing a cheek (82.7%), putting on a face mask (82.6%) and sticking out the tongue (82.0%). Sunglasses (67.7%), changing expression (61.3%) and a flash of light on the face (49.1%) were weaker. A head turn, the prompt most liveness flows use, degraded deepfakes least at 26.5%. The pipelines tested date from 2022, and newer tools handle some occlusions better, so treat the ranking as a guide.

A practical prompt set for agents, drawn in random order each session:

• Pass your hand slowly across your face, then hold it over your mouth while you say your date of birth.
• Press a finger into your cheek and hold it there for two seconds.
• Put on or take off your glasses, or hold your ID card partly in front of your chin.
• Turn your phone's torch or a lamp toward your face, then away.
• Read out a short random phrase shown on screen, so audio and lip movement must stay in sync.
• Repeat one earlier answer after an unrelated question, which catches prerecorded footage.

Where NeuroVerify fits in a video KYC platform

NeuroVerify is the software half of that pairing. The video KYC platform sends sampled frames, the face capture or a clip to the NeuroVerify API, and receives one unified risk score with signal-level evidence that can be shown to the agent and stored with the session.

One request runs six signals: Visual Safety Screening, Live Presence Verification, Face Manipulation Detection, AI Generative Model Analysis, AI Pixel-Level Manipulation and AI Content Compliance. Eligible image requests return in under 200 ms p95 under defined test conditions, so frame checks can keep pace with a live conversation. NeuroVerify also accepts audio, which covers synthetic voice in the same call.

Video KYC vendors and banks can integrate through Android, iOS and Flutter SDKs, backend SDKs for Python, TypeScript, Go and Java, or the REST API, and deploy in managed cloud, private cloud or VPC, a secure container, on premises or air-gapped. Neural Defend holds ISO/IEC 27001, ISO/IEC 27701, ISO/IEC 42001 and SOC 2 Type II certifications.

The models build on ARCAS 1B, Neural Defend's deepfake detection foundation model. On the public VendorBench-100 benchmark, Neural Defend ranked first of 36 systems with an MCC of 0.876 against 0.646 for the next vendor; that test uses 100 still images, so it says nothing direct about live video, and one of its authors also co-authored the ARCAS paper.

What deepfake detection cannot do in video KYC, and what to record

A manipulation score cannot flag a genuine recording of the real customer injected into the call; varied questions and session freshness handle that. It does not detect virtual camera drivers or emulators either, so the platform needs device and capture-integrity checks, as the virtual camera injection guide explains. Poor bandwidth, heavy compression and low light reduce signal quality for both agents and models, and some customers cannot perform physical prompts because of disability, so procedures need an accessible alternative.

In India, RBI's V-CIP rules already require variable questions, liveness or spoof detection, recording and an activity log; the India KYC guide maps those clauses. Elsewhere, NIST SP 800-63A-4 sets similar expectations for remote attended proofing.

For each session, store the frames that were scored, the request identifiers, model version, signal results, the prompts the agent used and the agent's decision. That record lets a reviewer reconstruct why a session was approved and supports retesting when attack tools change.