Abstract
This article presents a comparative study of crossmodal attention architectures for intelligent AI-based proctoring. Our microservices-oriented multimodal system integrates video inputs (gaze, facial dynamics, head position) and audio features (paralinguistic anomaly signals) to detect cheating during online exams. Six attention mechanisms are evaluated: baseline concatenation, scalar product attention, joint attention, cross-modal transformer (CMT), extended memory attention, and gated crossattention, using a curated dataset of remote exam recordings. The CMT model achieves the highest macro F1 score of 0.935, outperforming the baseline concatenation approach (0.872) and other attention variants (0.905-0.928), with an absolute improvement of 0.063. Visualization of attention weights offers interpretable evidence by highlighting key temporal segments and modalities that trigger alerts, enhancing system transparency (XAI) and auditability. The findings demonstrate that explicitly modeling intermodal dependencies through cross-modal attention leads to higher accuracy and fewer false positives than unimodal or latefusion methods, confirming its practical value for scalable and secure remote assessment systems.