A Unified Framework for Automated Proctoring under Overhead Surveillance and Imperfect Supervision

Abstract

Automated exam proctoring under overhead surveillance is challenging due to desk occlusion, privacy concerns, rare cheating events, and limited manual annotations. To address these issues, we propose a noise-aware proctoring framework based on Temporal Kinematic Network (TKNet), a topology-independent backbone that focuses on the temporal evolution of upper-body kinematics through a perspective-robust spatial-temporal representation, enabling fine-grained cheating micro-motion recognition under severe overhead occlusion. The framework then uses entropy-based routing to select ambiguous windows for pseudo-labeling with Multimodal Large Language Models (MLLMs). To reduce pseudo-label noise, we introduce a dual-smoothing strategy that combines trust-aware soft labeling for MLLM annotations and self-distillation for hard-negative normal behaviors. Due to the lack of standardized overhead proctoring benchmarks, we evaluate our approach on a private in-the-wild dataset and a modified public benchmark. Compared with graph-based and 3D-CNN baselines, the standalone TKNet achieves competitive Cheating-class performance with 0.48 Recall and 0.53 F1-score, while the full framework further improves Recall to 0.57 and F1-score to 0.56. TKNet also requires only 0.71M parameters and 0.044 GFLOPs per candidate window, supporting efficient window-level inference after pose extraction. These results suggest that topology-independent kinematic modeling with noise-aware pseudo-label refinement is a promising direction for privacy-aware overhead behavior recognition.