Scholarlink

<p>Online assessments have become a cornerstone of modern education, but ensuring fairness and academic integrity remains a significant challenge. Traditional proctoring systems rely on handcrafted features or manual supervision, which often fail to capture complex multimodal patterns of cheating behavior. This study addresses the problem of accurate and real-time cheating detection by proposing ViProctor-Transformer, a transformer-based multimodal framework that leverages cross-modal attention and long-range temporal encoding to model complex behavioral cues during online exams. The novelty of this work lies in introducing the first unified transformer-based proctoring framework that jointly performs cross-modal audio–visual–behavioral reasoning with boundary-aware temporal localization, enabling real-time, interpretable cheating detection beyond the capabilities of existing CNN-driven and heuristic fusion approaches. The model employs a Video Swin-Transformer for visual stream encoding and an Audio Spectrogram Transformer for audio feature learning. Cross-modal attention modules integrate these streams, while a temporal transformer head predicts cheating events with boundary-aware localization. Unlike prior CNN-based systems, this is the first integrated transformer framework that jointly model’s audio–visual–behavioral cues in real time, providing interpretable, boundary-aware cheating detection for scalable online proctoring. Training uses focal loss to address class imbalance and IoU-based boundary loss for precise event segmentation. The MSU Online Exam Proctoring (OEP) dataset, developed at Michigan State University, provides a unique resource for this problem. It contains synchronized webcam and wearable camera video, along with microphone audio, across 24 student sessions where five distinct cheating behaviors were systematically enacted and annotated with precise start and end times. Experimental evaluation under subject-independent protocols demonstrates that ViProctor-Transformer achieves superior performance compared to conventional CNN-based baselines. Specifically, the model achieved a macro-F1 score of 0.94, an AUROC of 0.97, and a 15.00–20.00% reduction in false positives compared with the ResNet50 baseline. In addition, the architecture maintains real-time efficiency, processing at ≥ 25 fps on a single GPU, and offers interpretability by distilling gaze, speech, and phone-related cues through auxiliary attention heads.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multimodal Transformer Framework for Real-Time Cheating Detection in Online Assessments and E-learning Platforms

  • 期刊论文

  • 发表日期：

    2026年1月13日

查看全文
  • Edrees A. Alkinani

摘要

Online assessments have become a cornerstone of modern education, but ensuring fairness and academic integrity remains a significant challenge. Traditional proctoring systems rely on handcrafted features or manual supervision, which often fail to capture complex multimodal patterns of cheating behavior. This study addresses the problem of accurate and real-time cheating detection by proposing ViProctor-Transformer, a transformer-based multimodal framework that leverages cross-modal attention and long-range temporal encoding to model complex behavioral cues during online exams. The novelty of this work lies in introducing the first unified transformer-based proctoring framework that jointly performs cross-modal audio–visual–behavioral reasoning with boundary-aware temporal localization, enabling real-time, interpretable cheating detection beyond the capabilities of existing CNN-driven and heuristic fusion approaches. The model employs a Video Swin-Transformer for visual stream encoding and an Audio Spectrogram Transformer for audio feature learning. Cross-modal attention modules integrate these streams, while a temporal transformer head predicts cheating events with boundary-aware localization. Unlike prior CNN-based systems, this is the first integrated transformer framework that jointly model’s audio–visual–behavioral cues in real time, providing interpretable, boundary-aware cheating detection for scalable online proctoring. Training uses focal loss to address class imbalance and IoU-based boundary loss for precise event segmentation. The MSU Online Exam Proctoring (OEP) dataset, developed at Michigan State University, provides a unique resource for this problem. It contains synchronized webcam and wearable camera video, along with microphone audio, across 24 student sessions where five distinct cheating behaviors were systematically enacted and annotated with precise start and end times. Experimental evaluation under subject-independent protocols demonstrates that ViProctor-Transformer achieves superior performance compared to conventional CNN-based baselines. Specifically, the model achieved a macro-F1 score of 0.94, an AUROC of 0.97, and a 15.00–20.00% reduction in false positives compared with the ResNet50 baseline. In addition, the architecture maintains real-time efficiency, processing at ≥ 25 fps on a single GPU, and offers interpretability by distilling gaze, speech, and phone-related cues through auxiliary attention heads.