Abstarct: One of the fundamental challenges in remote exams is proctoring exam scenes to prevent and detect cheating. The use of intelligent video systems for the automatic detection of cheating in remote exams is critical. Cheating can occur overtly (openly) or covertly. Overt cheating typically involves the visible use of cheat tools such as notes, which can be easily detected by scene inspection. However, covert cheating involves hidden tools or is done without specific tools, making it less visible to the proctor. During covert cheating, the examinee exhibits unusual behavior, such as looking around frequently and not paying attention to exam questions. In this study, the detection of suspicious scenes indicative of covert cheating in electronic exams is addressed through behavioral analysis of the examinee of interest. This can be considered as an anomaly detection problem in video. Since human behavior is manifested through body postures and movements, examining these plays a crucial role in identifying abnormal behaviors. Therefore, detecting body postures and movements is highly significant in this context. In contrast to previous research where the problem was treated as supervised (classification) for simplicity, focusing on predefined patterns of abnormal behaviors, this thesis considers the rarity of suspicious scenes compared to normal scenes, thus adopting a semi-supervised approach. Given that the aim of this thesis is to detect abnormal behaviors without the use of unauthorized tools during exams, relying solely on participant actions and behaviors, the use of skeleton data is proposed. By extracting features baxsed on skeletal data, video data is transformed into temporal sequences of body joint features. Consequently, the problem of cheating detection transforms into an anomaly detection problem in time series data, for which numerous algorithms have been proposed for various applications. This thesis evaluates the effectiveness of several semi-supervised anomaly detection methods baxsed on deep learning. In addition to the conventional approach where training data from various individuals are uniformly used (referred to as the global training approach), training baxsed on individual-specific models (referred to as the video-specific training approach) and its combination with the global training approach are also considered. The investigation into the video-specific training approach demonstrates that it can be equally effective for detecting suspicious behaviors of each individual as the global approach, which requires more data. To evaluate the proposed approaches, a dataset of 91 videos of remote exam sessions was collected. The results obtained from various experiments were evaluated using metrics of average precision (AP) and area under the receiver operating characteristic curve (AUC), yielding values of 0.66 and 0.82, respectively, indicating more than 10% improvement over the compared methods.