How to verify candidate identity in a live interview (2026 guide)
Sep 15, 2026
How to verify a candidate's identity during a live interview
In July 2024, KnowBe4 hired a Principal Software Engineer who turned out to be a North Korean operative. They published the whole thing themselves, which is why we can quote it:
"Our HR team conducted four video conference based interviews on separate occasions, confirming the individual matched the photo provided on their application."
Four video interviews. The photo matched. The photo was an AI-enhanced stock image. The background check came back clean because the stolen identity was a real person's. Malware started loading onto the company laptop on 15 July at 21:55 EST.
That's the clearest available evidence for something most hiring teams still assume the opposite of: seeing a face on a call is not identity verification. It's a face on a call.
Three different problems get lumped together and they need different answers.
Someone sitting off camera feeding answers. This is the least technical version and, in our experience, the most common. For scale, HirePro's analysis of 900,000 assessments in India put cheating at 30 to 50 percent of candidates for entry level roles and 10 to 25 percent for lateral hires, though that study covers assessments rather than interviews and does not break the figure down by method.
A different person on camera entirely. A paid proxy who takes the interview so the real candidate gets the job. The US Department of Justice's December 2024 indictment of fourteen North Korean nationals describes the tactic in plain terms: "paying U.S. persons to attend job interviews and work meetings remotely under fake identities." That scheme generated $88 million over about six years.
A synthetic face over a real person. A deepfake, either a face swap or a lip-synced overlay. This is the rarest and the one that gets all the coverage.
The scale is real but the good numbers are narrower than the headlines. Gartner's July 2025 survey of 3,000 candidates found 6 percent admitted to interview fraud, either posing as someone else or having someone pose as them. That's self reported, so it's a floor. Gartner also predicts one in four candidate profiles worldwide will be fake by 2028, though that's an analyst prediction with no published methodology sitting inside a release about candidate trust in AI. Treat it as a forecast, not a finding.
In June 2025 the DOJ announced actions covering more than 100 US companies, 21 laptop farms searched across 14 states and roughly 137 laptops seized. CrowdStrike's 2025 threat hunting report linked the same actor to more than 320 companies over twelve months, a 220 percent year on year increase.
The free things that work
Before any tooling, there's a set of checks that cost nothing and have caught real cases. The FBI published them in July 2025, in an alert about North Korean IT workers. They're worth quoting directly because they're operational rather than theoretical:
"Mandate video and request that their backgrounds be unobscured" "Have the individual point the camera out a window and ask questions about their claimed current location" "Ask the individual to wave their hand in front of their face as it may prompt a malfunction in AI generated video" "When possible, mandate in-person drug tests or fingerprinting to verify identity and claimed location" "Capture images for comparison with future meetings"
The hand wave is the interesting one because it's been tested in the wild. Dawid Moczadło at Vidoc Security Lab asked a suspected deepfake candidate to do it and the candidate refused. A recruiting lead at Make asked the same thing of a candidate whose face edges were distorting, and they disconnected after 30 seconds. Occlusion breaks face mapping models, which is why it works.
One caveat we'd rather state than have you discover. That's a 2026 weakness, not a permanent one. Occlusion handling improves with every model generation, and the hand wave will stop working at some point.
Two more that cost nothing. Ask about something on the resume that can't be looked up, then ask a follow up to the answer, because a person reading a script can't handle the second question. And note whether responses arrive as structured bullet points instead of conversation, which is what Moczadło reported.
What automated verification actually adds
Four signals do real work and they don't do the same work.
Face matching against a government ID. This is the baseline and it's genuinely useful, with one important qualification from NIST's August 2025 work on morph detection: comparing a face against a trusted reference image performs at 72 to 90 percent, and is far more consistent than analysing a single image in isolation, which drops below 40 percent against generators the detector hasn't seen. The lesson is that verification against a document beats trying to spot a fake from the image alone.
Liveness detection. Confirming there's a real three dimensional person in front of the camera rather than a photo, a mask or a screen. The standard here is ISO/IEC 30107-3, and iBeta runs conformance testing against it at two levels. Level 1 caps artefact cost at $30 and allows a 0 percent penetration rate. Level 2 allows artefacts up to $300, including 3D printed and resin masks, and up to a 1 percent penetration rate.
There's a limitation in that standard you should know about before a vendor quotes it at you. iBeta's published methodology covers presentation attacks, meaning artefacts physically shown to a camera. It does not cover injection attacks, where a synthetic video stream is fed straight into the pipeline through a virtual camera driver and never passes in front of a lens at all. A vendor advertising iBeta Level 2 has demonstrated resistance to masks and printouts. That isn't the same as resistance to a deepfake injected into a video call.
Lip sync and audio-visual consistency. This is where automation earns its place, and the research is unusually clear about why. A 2025 systematic review of the research reports one study in which humans detected face swap deepfakes at 91.3 percent accuracy and lip sync deepfakes at 52.6 percent. That second number is a coin flip. Humans are good at spotting a wrong face and useless at spotting a real face driven by synthesised speech, which is exactly the attack shape in a video interview.
The FBI's original 2022 alert describes the signal in the same terms: "the actions and lip movement of the person seen interviewed on-camera do not completely coordinate with the audio of the person speaking."
Voice verification. Cross-checking vocal patterns against earlier interactions with the same candidate. Useful, and worth pairing with the fact that people are poor at this unaided: a 2023 UCL study of 529 participants found humans identified deepfake speech correctly 73 percent of the time, missing more than a quarter, and training barely helped. The caveat on the automated side is that the ASVspoof 5 evaluation found anti-spoofing performance degrades under neural encoding and compression. Zoom, Teams and Meet apply exactly that to every call, so accuracy measured on clean audio doesn't transfer straight to a conference stream.
At BerriBot, Berri Proctor detects micro-signals through 3D liveness and lip sync analysis as part of the interview product rather than as a separate module. Our reasoning was that the decision gets made in the room, so that's where the check belongs. Other vendors do this differently and some do parts of it better.
The false positive problem, which nobody puts on a slide
Detection rates are half a number. Here's the other half, from published research on virtual camera detection:
Share of injection attacks caught
Share of genuine candidates wrongly rejected
90 percent
14.6 percent
99 percent
68.3 percent
99.9 percent
91.7 percent
To catch 99 percent of injection attacks at that operating point, you reject roughly two thirds of real people.
That's one method and one paper, and the numbers will differ across systems. The shape of the trade off won't. Any vendor quoting a detection rate without a false rejection rate at the same operating point is showing you one column of a two column table.
In hiring this matters more than in fraud prevention generally, because a false rejection isn't a declined transaction. It's a qualified candidate who was told they cheated. Ask every vendor what happens to that person, who reviews the flag, and how the candidate appeals.
Related, and worth building into your process regardless of tooling: an integrity signal should trigger a human review, not an automatic rejection. Deepfake-Eval-2024, a benchmark built from deepfakes actually circulating in the wild rather than lab-generated ones, found detector AUC dropped by about 50 percent for video compared with previous benchmarks. Commercial models beat open source ones and none reached the accuracy of human forensic analysts. The tooling narrows the field. It doesn't make the call.
What the law lets you collect
Short version, and we're a software company rather than your lawyers, so confirm this with counsel.
India. The DPDP Act 2023 creates no special category for biometric data, which surprises people. Biometric verification sits under the general regime, not a heightened one. Section 5 requires notice before or with the consent request, available in English or any of the 22 scheduled languages. Section 6 requires consent that is free, specific, informed, unconditional and unambiguous, separately for each purpose, withdrawable as easily as it was given. Section 7 lists legitimate uses that don't need consent, including employment purposes, but it's written in terms of "a Data Principal who is an employee." Whether a job applicant who hasn't been hired falls inside that is an open question Indian counsel are still arguing about, and the Rules notified in November 2025 don't clarify it. The safe position is to take consent from candidates. Penalties reach ₹250 crore for security safeguard failures.
Illinois. BIPA applies to a face match on an Illinois candidate. Section 15(b) requires written notice that biometrics are being collected, disclosure of the specific purpose and the retention term, and a written release before collection. An August 2024 amendment made two things easier: electronic consent is expressly valid, and repeated collection from the same person now counts as a single violation rather than one per scan, which had been the position after Cothron v White Castle. Statutory damages are $1,000 per negligent violation and $5,000 per intentional one. Texas and Washington have similar statutes.
EU. Two things to separate. Using AI to infer emotions of a person in a workplace context is prohibited under Article 5(1)(f) of the AI Act, in force since 2 February 2025, with a narrow exception where the system is put in place for medical or safety reasons. That's a ban rather than a compliance checklist, and it's likely to catch products that score a candidate's confidence, enthusiasm or stress. Biometric verification, meaning one to one matching to confirm a person is who they claim, is explicitly carved out of the remote biometric identification category in Annex III point 1(a). Recruitment and candidate evaluation systems are separately classified as high risk under Annex III point 4(a). The date those high risk obligations bite has moved under the Digital Omnibus, so check the current position before you rely on a date you read somewhere.
A process that works without buying anything
If you're starting from zero, this order gets you most of the benefit.
Take a government ID and a selfie at application, and keep the selfie as the reference image for every later session. Run the interview on video with backgrounds unobscured. Ask one unscripted follow up to a resume answer, early. Use the hand wave when something feels off, and treat a refusal as the answer. Capture a frame from each session and compare against the reference and against prior sessions, because the North Korean cases consistently involve identity switching between rounds. Get consent in writing before any of the biometric parts.
Then, if volume makes that impossible to do by hand, buy tooling for the parts that don't scale: face matching against the reference, liveness, and lip sync analysis. Keep the human judgement for what happens after a flag.
FAQ
Is a video interview enough to verify a candidate's identity?
No. KnowBe4 conducted four video interviews with a North Korean operative in 2024, confirmed the person matched the photo on the application, and hired them. The photo was an AI-enhanced stock image over a stolen but real identity. Video confirms a face is present, not that the face belongs to the applicant.
How do you tell if a candidate is using a deepfake in an interview?
The FBI recommends asking the candidate to wave a hand in front of their face, which can break AI generated video, and asking them to point the camera out a window and answer questions about their location. Other reported signals include lip movement not matching the audio, distortion at the edges of the face, misalignment between head and neck, and answers delivered as structured bullet points rather than conversation. Automated lip sync analysis helps most here, because in the research humans detect lip sync deepfakes at around 52.6 percent accuracy, against 91.3 percent for face swaps.
What is liveness detection and does it stop deepfakes?
Liveness detection confirms a real three dimensional person is in front of the camera rather than a photo, mask or screen. It's tested against ISO/IEC 30107-3, with iBeta running conformance at Level 1 and Level 2. That standard covers presentation attacks, meaning artefacts shown to a camera. It does not cover injection attacks, where synthetic video is fed into the pipeline through a virtual camera and never passes a lens. So liveness certification is meaningful but does not by itself answer the deepfake question.
Can we legally require biometric identity verification of job candidates in India?
Under the DPDP Act 2023 biometric data has no special category and falls under the general regime. You need notice under Section 5 and consent under Section 6 that is free, specific, informed and withdrawable, taken separately for each purpose. The Section 7 employment exemption is written in terms of employees and its application to pre-hire candidates is unsettled, so the safe position is to take consent. Confirm with counsel.
Does the EU AI Act ban AI interviews?
It bans one specific thing: inferring emotions of a person in a workplace or education context, prohibited under Article 5(1)(f) since February 2025, with a narrow exception for medical or safety purposes. Recruitment and candidate evaluation systems are classified high risk rather than prohibited, with obligations rather than a ban. Biometric verification, matching a person against their own ID, is explicitly excluded from the remote biometric identification category.
What should we do when an integrity check flags a candidate?
Review it with a person before acting. Published benchmarks on deepfakes circulating in the wild show detector AUC falling about 50 percent for video compared with lab benchmarks, and no commercial model reaching the accuracy of a human forensic analyst. Decide in advance who reviews a flag, what evidence they see, and how the candidate can respond. A wrongly flagged candidate in hiring is a discrimination complaint, not a declined transaction.
Deepfake and proxy candidate detection: what the software can and can't do
What interview fraud detection software actually claims, what the research says it delivers, and the eight questions that separate real capability from marketing.
