How Do You Spot a Deepfake on a Live Video Call?

What is a live video impersonation attack?

A live video impersonation attack is a real-time call in which the attacker appears as someone else. The tools range from crude to sophisticated: a still photo with the camera "frozen", a pre-recorded clip played through a virtual camera, a real-time face swap that maps another person's face onto the attacker's, or a fully animated avatar driven by the attacker's expressions. A cloned or converted voice usually accompanies the video.

The aim is the same as other impersonation fraud: a payment, credentials, access or sensitive information. The FBI Internet Crime Complaint Center reported on 3 December 2024 that criminals use real-time video chats with purported company executives, law enforcement or other authority figures, and send private videos to convince online contacts they are real people. An earlier IC3 warning, in February 2022, described attackers showing a still image of a CEO with no audio, or using deepfake audio, in video meetings before moving the transfer request to chat or email.

Live attacks also target identity verification. A real-time face swap streamed through a virtual camera can attempt to pass a video KYC session or a remote hiring interview, which is why NIST and FinCEN guidance addresses manipulated media during live checks.

How do you spot a deepfake on a live video call?

You spot a deepfake on a live video call by testing the conditions real-time models handle poorly and by watching for technical tells. Ask the person to turn their head slowly to a full profile, to pass a hand in front of their face, to put on or take off glasses, or to hold an object next to their cheek. Face swaps often blur, flicker or show the original face at the edges during these moves.

NIST SP 800-63A-4, finalised in July 2025, requires remote proofing services to train agents to spot manipulated media, naming high latency, synchronisation issues and inconsistent skin tone or resolution as signs, and to introduce random human-in-the-loop cues such as requesting specific movements or asking the user to move objects between the camera and their face. Those cues work just as well in a business call.

None of these checks is decisive. Poor connections, cheap webcams and background blur cause similar glitches, and better tools pass some tests. Treat a failed test as a reason to stop and verify, and a passed test as no proof at all.

• Lip movements that drift out of sync with speech, or audio that lags the face.
• A face that is sharper, smoother or differently lit than the neck, ears and hair.
• Flicker or warping at the face edge when the person turns, touches their face or moves quickly.
• Refusal to turn on the camera properly, repeated 'connection problems' when challenged, or an insistence on moving to chat.
• Participants who never speak, or who all join and leave together.

Where does real-time face-swap technology still break down?

Real-time face-swap technology still breaks down wherever the model has to invent information it was not trained on or cannot compute fast enough. Profile views are the classic weakness: many models are trained mostly on frontal faces, so a 90-degree turn exposes stretching or a sudden switch back to the attacker's real features. Occlusion is another: a hand, cup or pen passing in front of the face forces the model to decide which pixels belong to the face, and errors show as smearing or the object vanishing.

Speed limits quality. To run at video frame rates, live tools use smaller models and lower resolutions than offline deepfakes, so fine detail in teeth, eyes and hair is weaker and fast motion causes lag or flicker. Lighting changes, such as a lamp switched on or a phone screen lighting the face from below, can leave the swapped face lit from the wrong direction. Accessories, beards and hairlines that differ between the attacker and the target are also hard to reconcile.

These gaps narrow with each tool generation, and attackers avoid triggering them by keeping calls short, blaming bad connections and steering away from requests to move. That is why procedural checks matter more than visual inspection.

Controls that work even when the deepfake looks perfect

The strongest control is never to approve consequential actions inside the call. Payments, bank detail changes, credential resets and data releases should require a callback on a known number and approval in the system of record, after the meeting ends. The Arup fraud, confirmed by the company in May 2024, shows why: a finance employee joined a call where the CFO and several colleagues appeared, then made 15 transfers totalling HK$200 million. Several familiar faces made the request feel verified.

Meeting hygiene helps. Use authenticated corporate accounts for internal meetings, be wary of unfamiliar meeting platforms, as IC3 recommends, and treat external invitations claiming to come from executives with suspicion. For identity verification, combine deepfake detection on the stream with controls that detect virtual cameras and emulators, since an injected feed may never show a physical glitch.

How Neural Defend helps with live video impersonation

For investigations, NeuroDetect takes an authorized meeting recording into a case and shows results down to frames and regions, with explainability heatmaps that highlight where the evidence sits, audio-segment views for the voice, and a downloadable forensic report recording uncertainty and the model version. That helps a security team establish what happened and link a campaign across incidents.

For identity flows such as video KYC or remote onboarding, NeuroVerify scores submitted video and images with six signals in one API call, including Face Manipulation Detection, Live Presence Verification and AI Generative Model Analysis, and returns a risk score with signal-level evidence for your policy engine. It does not detect virtual cameras itself, so pair it with device and session integrity checks. On the public VendorBench-100 benchmark of 100 adversarial still images, including face swaps and avatar composites, Neural Defend ranked first of 36 systems with an MCC of 0.876; one VendorBench author also co-authored Neural Defend's ARCAS paper, the corpus is public, and it does not test live video.