Research Foundation · Ege & Ceyhan, Huawei Turkey R&D (2023)

Privacy-First Client-Side Proctoring

Scaling online interviews requires trustworthy anti-cheat measures without invasive surveillance. We treat the browser-based approach pioneered by Huawei researchers as a blueprint—keeping inference on the candidate device and logging only anonymized events.

Detection capabilities

Device detection

The study reports 75.4% and 72.0% accuracy for detecting phones and laptops using COCO-SSD MobileNetV2.

Voice classification

Local inference achieved 97.1% accuracy distinguishing human voices from background audio in the published results.

Face verification

Face-api.js embeddings enable continuous presence validation paired with privacy-preserving segmentation in the reference design.

Environment privacy

BodyPix segmentation is applied locally to blur non-candidate regions before any frames are eligible for upload.

Privacy and governance

Compliance guardrails

Enterprise teams require clear data boundaries, explicit consent flows, and auditable logs.

  • GDPR/PDPA aligned: inference happens locally and only anonymized events are stored.
  • Configurable escalation: customers define thresholds for auto-flag, manual review, or ignore.
  • Transparency: candidates can preview captured evidence and request review via self-service portal.
  • Cost efficiency: zero server GPU cost unless a high-risk event needs human escalation.

Operational takeaways from the research

Challenges highlighted by server-side approaches

Centralized inference increases infrastructure cost and complicates cross-border data transfer.

Manual review backlogs grow when reviewers triage ambiguous clips without context.

Opaque monitoring erodes candidate trust and creates compliance risk.

How we incorporate the published guidance

Favor on-device inference so costs scale with review volume rather than total traffic.

Keep blurring and evidence handling local until escalation criteria are met.

Provide reviewers with contextual timelines and candidate notifications to reinforce transparency.

Implementation highlights

Local-first architecture

TensorFlow.js runs directly in the candidate browser, with WASM fallbacks for edge devices.

Risk-based escalation

Weighted signals (device, voice, face) feed a risk score that determines whether to store obfuscated evidence or simply log metadata.

Candidate experience

Inline notifications explain detections in plain language and provide remediation guidance before penalties apply.

Resources