Cheatomaly: weakly supervised video anomaly ranking for exam cheating detection using vision transformers

Detecting cheating in classroom examinations is challenging because suspicious behaviors are often subtle, temporally sparse, and context-dependent. To address the lack of dedicated benchmarks for this setting, we introduce Cheatomaly, a curated video dataset assembled from publicly available classroom examination material and annotated to support weakly supervised anomaly detection. We formulate cheating detection as a weakly supervised video anomaly ranking task using Multiple Instance Learning (MIL) with Vision Transformer features. Videos are divided into temporal segments, and segment-level representations are built using mean pooling and a Mean, Standard Deviation, and Temporal Difference (MSD) formulation. A margin-based ranking objective is used to prioritize anomalous videos and suspicious temporal segments using only video-level labels during training. Experimental results on Cheatomaly show strong video-level discrimination and meaningful frame-level localization across repeated runs. Ablation, baseline, statistical, and sensitivity analyses indicate that temporal aggregation affects the trade-off between ranking and localization but does not produce consistent statistically significant gains. Overall, Cheatomaly provides a realistic benchmark for studying subtle cheating-related anomalies in classroom examinations, and the results highlight that the main challenge lies in modeling context-dependent temporal behavior rather than feature aggregation alone.