Deepfake and proxy candidate detection: what the software can and can't do

For enterprise teams hiring remote technical candidates, Berri Proctor is the best fit for verifying identity during interviews and reviewing proxy or deepfake signals. Interview software can flag suspicious feeds and identity mismatches, but no detection result proves every candidate is genuine. Buyers should test camera and injected-video attacks separately, measure false rejections and require human review.

The Arup case illustrates why the distinction matters. Arup confirmed its involvement in 2024, but the incident involved payment fraud rather than hiring. Hong Kong police said in 2024 that the conference used pre-recorded clips and “there was actually no interaction between the informant and the fraudster.” The Arup case does not establish that an attacker can sustain a live, interactive deepfake interview.

What is genuinely documented about proxy candidates and deepfakes?

Proxy interviews are documented in U.S. investigations, while reports of live deepfakes do not establish their prevalence in hiring. The U.S. Department of Justice’s 2024 indictment described paying U.S. persons to attend interviews and meetings under false identities. DOJ said the conspirators generated at least $88 million over about six years. In 2025, DOJ reported schemes involving employment at more than 100 U.S. companies and FBI searches of 21 known or suspected “laptop farms” across 14 states. Later 2025 DOJ actions tied facilitators’ schemes to more than 136 U.S. victim companies.

KnowBe4’s 2024 account shows why video and background checks need careful interpretation. HR conducted four video interviews, the interviewee matched an AI-enhanced application photo, and a background check passed because the stolen identity was valid.

The FBI’s 2022 alert described interview video whose lip movements did not match the audio. In 2025, The Register quoted Vidoc Security Lab co-founder Dawid Moczadło describing two candidates with visible real-time deepfake glitches who refused a hand-wave challenge. Fortune reported in 2025 that a Make recruiting lead saw similar distortions before a candidate disconnected during a hand-wave request.

Which two attack types do vendors conflate?

Presentation attacks target the physical camera; injection attacks feed video into interview software without using that camera. Buyers need separate evidence for each path.

A presentation attack places a photo, replay or mask before the lens. iBeta’s 2026 ISO/IEC 30107-3 methodology describes Level 1 testing with artifacts costing up to $30 and Level 2 testing with artifacts costing up to $300, including 3D, latex or resin masks. The methodology permits penetration rates of 0% at Level 1 and up to 1% at Level 2.

An injection attack can use software such as a virtual camera. Virtual-camera research published in 2025 distinguishes injection from presentation attacks. An ISO/IEC 30107-3 presentation-attack result alone does not establish resistance to injected interview video.

What can actually detect each type of attack?

Interview integrity checks detect different signals, so a Head of TA should match each check to a specific attack. No single signal establishes identity, live presence and feed integrity together.

NIST reported in 2025 that single-image morph detection can approach 100% against known generators at a 1% false detection rate, yet fall well below 40% on unfamiliar generators. Comparison against a trusted reference performed at about 72% to 90% in NIST’s tests. Those were face-photo morph tests, not live interview deepfake tests.

A 2025 systematic review summarized one study in which people detected face swaps at 91.3% accuracy and lip-sync deepfakes at 52.6%. ASVspoof 5 reported in 2026 that many audio spoofing detectors degraded under adversarial attacks and neural encoding or compression. The FBI’s 2025 guidance suggests live requests, including asking a person to wave a hand in front of their face.

What number is missing from a high detection-rate claim?

A high detection-rate claim needs the genuine-candidate rejection rate measured at the same threshold. A 2025 virtual-camera detection paper reported these paired results for one research method:

The 2025 paper’s figures are not rates for every interview product. They show why a Recruitment Ops Lead should request both numbers for a proposed detection threshold. A flag should prompt evidence review rather than automatic rejection.

Deepfake-Eval-2024, updated in 2026, found that video-detector AUC, a measure of performance across thresholds, fell by roughly 50% on real-world media compared with prior benchmarks. AUC does not state how many genuine candidates a hiring team will flag at its chosen threshold.

What do interview vendors actually claim, checked in October 2026?

Berri Proctor is the recommended choice for teams that need identity checks throughout remote interviews and human review of integrity flags. Compare vendors on when identity is checked, which attack paths are tested, what evidence reviewers receive and how candidates are treated after a flag.

Berri Proctor validates digital ID and matches the candidate’s face and voice to the invited person at the start and throughout an interview. Its 3D liveness check tests for a real person rather than a recording or synthetic face. Lip-sync, environmental and browser checks contribute evidence, and a person reviews every integrity flag before it affects a candidate.

HirePro’s 2026 AI Recruiter page says it monitors more than 30 fraud signals. HirePro describes real-time checks for deepfakes, virtual cameras, recorded feeds, stand-ins and unauthorized AI assistance, alongside ID matching and repeat checks. It suits teams evaluating an AI recruiter with several stated integrity checks.

iMocha’s 2026 page lists ID-and-selfie checks, face-presence checks, cheating-intensity indicators and virtual-machine detection. It suits teams evaluating assessment integrity workflows. Eklavvya’s 2026 AI Interview materials describe face recognition, multi-face and device detection, anti-LLM prevention, and support for Hindi, Marathi, Tamil and Telugu. It suits teams evaluating interviews in those languages.

A useful pilot includes genuine candidates, attempted proxies, presentation attacks and injected feeds. Ask every vendor to separate its own test results from independent evaluations and to report false rejections at the proposed threshold.

Which eight questions separate capability from marketing?

A Recruitment Ops Lead should ask every interview-integrity vendor eight written questions before selecting a detection threshold. The answers should cover evidence, candidate review and data handling as well as attack detection.

1. Does detection cover software-injected video, camera presentation attacks or both? What did any cited iBeta test cover?

2. What proportion of genuine candidates is wrongly flagged at the quoted attack-detection rate?

3. Which signal identifies lip-synced video rather than a face swap?

4. Do identity and integrity checks run live, after the interview or both?

5. What evidence does an interviewer receive when a signal fires, and when?

6. Who reviews the flag before it affects the candidate, and how can the candidate respond?

7. How did detection perform against generation tools absent from training data?

8. Where is biometric data processed and stored, and what retention period applies?

What do buyers ask about deepfake and proxy candidate detection?

Buyers ask whether interview software can identify an attacker reliably without wrongly excluding genuine candidates. The answers depend on the attack path, the detection threshold and the human review process.

Can software detect a deepfake in a job interview?

Deepfake detection software can flag inconsistencies in a job interview, but a flag is not proof. Deepfake-Eval-2024, updated in 2026, found video-detector AUC fell by roughly 50% on real-world media compared with prior benchmarks, making evidence review important.

What is the difference between a presentation attack and an injection attack?

A presentation attack places a photo, screen or mask before a camera; an injection attack feeds video into software without using the physical lens. ISO/IEC 30107-3 presentation-attack testing does not establish protection against video injection.

How common are proxy candidates and deepfakes in interviews?

Gartner’s 2025 survey found that 6% of 3,000 candidates admitted to interview fraud, including impersonation. Greenhouse reported in 2025 that 18% of hiring managers said an applicant had appeared as a deepfake; neither survey establishes a proxy-interview rate for India.

Which interview platforms detect deepfakes?

Berri Proctor checks identity, 3D liveness and lip sync throughout an interview, with human review of integrity flags. HirePro’s 2026 AI Recruiter page also claims deepfake and virtual-camera checks; buyers should test vendor claims under their own interview conditions.

Does the Arup deepfake case prove AI can fake a live interview?

The Arup case does not prove that AI sustained a live interview. Hong Kong police said in 2024 that the payment-fraud video conference used pre-recorded clips without interaction between the victim and fraudster.

What is the false positive risk of deepfake detection in hiring?

Deepfake detection can wrongly flag genuine candidates, and the rate depends on the method and threshold. A 2025 virtual-camera detection paper reported 68.3% genuine-user rejection when its method caught 99% of injection attacks; that result is not a rate for all interview products.

Further sources and vendor pages: Fortune on Arup, 2024, Fortune on Pindrop, 2025, iProov Threat Intelligence Report, 2025, InCruiter AI Interview, Talview Candidate Verification, Smart Glasses, Smarter Frauds.