Collaborative Cheating Detection in Remote Interviews Audio-Visual Fusion Techniques

Cheating in remote interviews poses a serious threat to the fairness and credibility of candidate evaluations, particularly in the growing landscape of online recruitment and virtual assessments.This work presents an enhanced, lightweight, real-time cheating detection system that operates entirely on standard webcams without the need for specialized hardware or high-end GPUs.Building on a previously established faciallandmark-based pipeline, the proposed system now incorporates two additional visual cues-phone-in-frame detection and multiple-face presence-using a combination of MediaPipe Face Mesh and YOLOv8 object detection.The system extracts a 192-dimensional feature vector per frame by tracking 32 key facial landmarks, capturing both static positions and frame-to-frame motion dynamics.These features are fed into a Random Forest classifier trained on labeled video data representing both cheating and normal behaviors.To further improve robustness, a YOLOv8-nano model identifies mobile phones in the frame, and a simple face-count logic flags instances of multiple individuals appearing in view.Alerts are only triggered when these signals persist over defined time thresholds, reducing false positives.Extensive testing across real and augmented datasets demonstrates that the multi-cue system significantly improves event-level recall (from 76% to 92%) while maintaining real-time inference speeds on consumer-grade CPUs.This practical, interpretable, and privacyconscious approach offers a scalable solution for integrating cheating detection into remote interview platforms, AI-assisted proctoring systems, and HR tools, without relying on opaque deep learning video models.