Abstract:Cheating in exams is accompanied by behavioral patterns (global or person-specific) that can indicate its occurrence. Recognizing cheating using predefined rules has limited success as it may have various patterns in different individuals. In this paper, we detect cheating in e-exam videos using semi-supervised methods considering global and person-specific patterns. We propose a hybrid model (HM) that combines two semi-supervised anomaly detectors relying on both a global model (GM) and a video-specific model (VSM). We collected a dataset of videos captured by a side camera during mock e-exams. The initial part of each video (approximately 30%) is free of cheating attempts. The GM is trained on the sequences of features formed by the initial part of frames of all videos from different individuals to learn the common patterns of normal behavior. Then, a VSM is trained for each video to learn the normal behavior of each subject. These deep-learning-based submodels are trained on normal data and generate a reconstruction error as an anomaly score for each test sequence. HM computes the combined anomaly scores by averaging the scores computed by GM and VSM. Finally, we calculate evaluation metrics such as area under the ROC curve (AUC) and average precision (AP) over the assigned scores. Skeleton-based features are used as input data to preserve privacy and ignore background clutter. The proposed model achieves an AP of 0.66 and an AUC of 0.82 on the collected dataset, showing its suitability for human-oriented anomaly detection tasks.

Semi-Supervised Skeleton-Based Covert Cheating Detection in Electronic-Exams