Examinations serve as a cornerstone of educational assessment, rigorously evaluating students’ comprehension, skills, and abilities across diverse subject areas. Traditionally administered within controlled, in-person settings, exams have been widely recognized as a credible measure of academic achievement, as well as a crucial determinant of students’ academic and professional pathways [1]. With the recent, unprecedented shift to online learning [2], educational institutions worldwide have been compelled to reimagine and adapt these evaluative frameworks. This transition, often to digital assessment platforms, has fundamentally transformed the nature of the assessment process, introducing critical challenges to maintaining fairness, equity, and academic honesty [3–5].

The COVID-19 pandemic accelerated the widespread adoption of remote examinations, exposing inherent vulnerabilities within the online assessment environment [6,7]. Studies have highlighted a marked increase in cheating behaviors during online exams, underscoring the limitations of remote proctoring in preventing dishonest practices [8,9]. In the absence of physical oversight, combined with the high-stakes nature of assessments, students may be incentivized to exploit digital tools for assistance. The pervasive availability of internet resources, mobile devices, and screen-sharing technologies compounds this issue, posing significant risks to the credibility of exam outcomes and the integrity of academic assessments [10].

Several studies attribute the surge in online exam cheating to a range of factors, including environmental variables, institutional pressures, and the distinct challenges inherent to remote assessments [11]. As online education expands, ensuring the reliability of remote assessments has become a critical priority. However, existing solutions such as browser lockdown tools [12,13], identity verification processes [14], live proctoring [15], and 360 degree monitoring face notable limitations [16]. These methods often suffer from restricted dataset diversity, lack of real-time adaptability, and limited flexibility in addressing the wide array of cheating tactics that students may employ. Additionally, these approaches can be intrusive and challenging to implement consistently, particularly when external environmental factors are difficult to control, thus highlighting persistent gaps in the current strategies for online exam proctoring.

To address these challenges, the present study conducts an in-depth evaluation of various pre-trained deep learning models, leveraging two benchmark datasets specifically tailored for online proctoring contexts. By critically analyzing the performance, accuracy, and real-time adaptability of these models, this research aims to identify optimal approaches that achieve a balance between detection efficacy and computational efficiency. The findings presented herein offer valuable insights into the development of robust, scalable models that are capable of strengthening academic integrity within remote assessment environments.

The contributions of this paper are as follows: •

Comprehensive Comparison of Deep Learning Models for Cheating Detection: This study provides an in-depth comparative analysis of advanced deep learning models, including EfficientNet, MobileNetV2, and ResNet variants, specifically evaluating their effectiveness in detecting cheating behaviors in online exams.

The structure of this paper is organized as follows: Section 2 reviews related works, providing an overview of existing methodologies and their limitations. Section 3 details the datasets and classification methods used in this study. In Section 4, we present evaluation metrics, experimental results, and an analysis of our findings. Then, we discuss the results in Section 5, followed by the presentation of our work’s limitations in Section 6. Finally, Section 7 offers conclusions and suggests directions for future research.

This section provides an overview of recent systems for cheating detection in online examinations, highlighting the application of machine learning, deep learning, and computer vision methodologies for visual analysis and behavioral surveillance.

Liu et al. [8] proposed a framwork named CHEESE for identifying academic dishonesty in online examinations by multiple instance learning (MIL) and spatio-temporal graph analysis. They introduced an innovative weakly supervised method that combines body posture, gaze, head movement, and face data, obtained from OpenFace 2.0 and 3D convolutional networks, to identify spatio-temporal abnormalities in video recordings. Evaluated on the Online Exam Proctoring (OEP) dataset of 65 aberrant and 69 normal videos, CHEESE attained an AUC of 87.58%, surpassing leading methodologies. The model exhibited robust performance on additional anomaly detection datasets, achieving 80.56% AUC on UCF-Crime and 90.79% on ShanghaiTech Dataset.

Ramzan et al. [17] evaluated the efficacy of pre-trained convolutional neural networks in identifying anomalous behaviors during online examinations, specifically cheating behaviors. Key video frames were recovered by motion-based approaches, and deep learning models, including YOLOv5, DenseNet121, and Inception-V3, were utilized to categorize cheating behaviors, such as head movements or the use of external devices. The YOLOv5 model attained superior performance, with a precision of 95.54%, a recall of 93.16%, and a mean average precision (mAP) of 95.40%, outpacing alternative methods in identifying cheating behaviors during online examinations.

Kaddoura and Gumaei [18] suggested a deep learning method for detecting cheating in online examinations using video frames and speech analysis. Their approach incorporates three essential modules: front camera detection, rear camera detection, and speech-based detection, utilizing CNN and Gaussian-based DFT for real-time cheating classification. The system was tested on the OEP dataset, which comprised data from 24 students, achieving an accuracy of 99.83% for front camera recognition and 98.78% for back camera detection.

Nurpeisova et al. [19] created the Proctor SU system, an automated online proctoring technology aimed at improving the integrity of examinations in the higher education sector in Kazakhstan. The system incorporates AI technologies like face detection, facial recognition, and behavioral analysis, employing models such as CNN, R-CNN, and YOLOv3 for real-time surveillance. The system, evaluated at Seifullin Kazakh Agro Technical University, analyzed 5000 photos with a detection rate of 97.12%.

Ozdamli et al. [20] created a facial recognition system to assess student emotions and identify cheating in remote learning settings. Their approach employed computer vision and deep learning technologies, such as CNN, to assess face expressions and actions during online examinations. The system, evaluated using datasets such as FER and FEI Face Database, attained an accuracy of 99.38% for student recognition and 66% for emotion detection. Furthermore, it monitored gaze direction and head motions to identify cheating, achieving a gaze tracking model accuracy of 96.95%.

Hu et al. [21] devoloped a multi-perspective adaptive cheating detection system for online examinations, employing three cameras (overhead, horizontal, and frontal views). The system integrates a gaze recognition model, detects cheating tools, and behavior analysis to enhance monitoring. The system employs datasets of 3000 photos for gaze direction and 7600 overhead and 8899 horizontal perspective images for detecting cheating tools, dynamically adjusting perspectives according to gaze direction. It attained 95% accuracy, providing an efficient and immediate method for mitigating cheating in digital examinations.

Dang et al. [22] developed an AI-driven auto-proctoring system intended for incorporation into MOOCs, emphasizing the identification of dishonest conduct during online assessments. The system integrates facial recognition, mobile device detection, and head posture estimation, employing deep learning models like RetinaFace and YOLOv10 for precise identification. Their system, evaluated on 4311 films from actual examinations, attained a remarkable accuracy of 95.66%, with an average reaction time of 0.517 s.

Potluri et al. [23] presented an automated AI-driven online proctoring system utilizing Attentive-Net to identify student misconduct during tests. The system has four modules: facial detection, multiple individual detection, facial spoofing, and head posture estimation. The authors employed the CIPL dataset alongside a bespoke dataset of over 200 movies exhibiting diverse behaviors to assess their model. Their technique attained an accuracy of 87% utilizing the Attentive-Net in conjunction with the Liveness Net and head position estimation.

Several studies have concentrated on cheating detection systems for online examinations, incorporating facial recognition, object detection, and head and gaze tracking methods. Authors in [24,25] both employed YOLOv3 for object detection in their systems. They introduced a system that combines facial recognition, tracking head movements, identifying objects, gaze and mouth tracking, background analysis, etc. Achieving an accuracy surpassing 80% in gaze tracking, mouth supervision, and object detection. Singh et al. [26] similarly utilized YOLO to detect dishonest activities, including face recognition, mouth tracking, and mobile device detection, to ensure academic integrity during virtual assessments.

Roy and Chanda [27] focused on developing a cost-effective, webcam-based eye-gaze estimation system for human-computer interaction (HCI) using a convolutional neural network (CNN) and Mediapipe for real-time facial identification. Parkhi et al. [28] developed a comprehensive examination monitoring system leveraging deep learning, incorporating face detection, face spoofing detection, object detection (YOLO), eye tracking, and head-pose estimation. Their system achieved a 90% success rate in tracking face and head movements. Gadkar et al. [29] presented an automated proctoring system that incorporates facial recognition, mouth aspect ratio analysis for speech detection, and audio detection to identify suspicious phrases or multiple individuals in the frame, using Haar cascade for face detection.

In terms of anomaly detection, Atabay and Hassanpour [30] proposed a semi-supervised method for online exam proctoring based on skeletal similarity. Their system segments exam videos by evaluating skeletal features and computing the similarity between training and test segments to identify abnormal behavior.

Other approaches focus on human pose estimation. Samir et al. [31] utilized TensorFlow PoseNet to monitor head positions and hand motions in real-time, achieving detection accuracies of 94% for head posture and 97% for hand motions.

This section outlines the proposed benchmarking approaches. To clearly illustrate the implementation flow of our proposed cheating detection framework, we provide the following pseudocode (see Algorithm 1), which outlines the key steps involved—from data preprocessing and annotation to model training and evaluation. The process begins by extracting frames from the input videos, followed by manual annotation of those frames based on observed behaviors. The data is then split into training, validation, and testing subsets. Finally, a selected deep learning model (e.g., EfficientNet, ResNet, or YOLOv5) is trained using the labeled data and evaluated on unseen samples to detect cheating behavior. This structured pipeline ensures a reproducible and transparent methodology.

The evaluation of deep learning models on the OEP and OP datasets uncovers substantial insights into their efficacy for detecting online test cheating, highlighting the trade-offs among accuracy, computational efficiency, and resilience. EfficientNetB0 and ResNet50-CBAM continuously rank as superior models on the OEP dataset, exhibiting elevated test accuracy alongside balanced precision, F1 scores, and AUC, making them highly reliable choices for accurately detecting cheating behaviors. The ConvNext model, while not achieving the maximum accuracy, exhibits the quickest inference time, indicating its appropriateness for real-time applications. In the OP dataset, EfficientNetB1 excels in accuracy and F1 scores, demonstrating a balanced and effective performance in precision and recall measures, but with increased processing requirements during training and efficient test performance ensures reliable detection with minimal delay, aligning well with the demands of real-time applications (see Figs. 11 and 12).

Concerning YOLOv5 on both datasets highlights the disparities in dataset architecture and its impact on model convergence. The OEP dataset seems to promote more efficient model convergence, probably owing to balanced feature patterns and class distributions, leading to consistently reduced test losses. In contrast, models trained on the OP dataset exhibit elevated loss values and variability, suggesting potential issues such as class imbalance or noisy features. Despite these limitations, YOLOv5 attains elevated precision-recall regions and mAP scores across both datasets, confirming its formidable detection capability.

The findings indicate that EfficientNetB0, EfficientNetB1, and ResNet50-CBAM are superior models for achieving a balance between accuracy, computational efficiency, and generalization, whereas YOLOv5 has significant potential for high-stakes cheating detection applications. The findings underscore the significance of choosing the suitable model and dataset, taking into account model convergence, efficiency, and class distribution to get dependable performance in identifying cheating behaviors.

To ensure practical applicability, the models evaluated in this study were specifically chosen and fine-tuned to handle a wide range of real-world cheating tactics, including behaviors involving second devices, behind-the-scenes assistance, and deceptive strategies like imitation. In the OP dataset, for instance, cheating behaviors include the examinee interacting with hidden mobile phones, receiving whispered answers, or mimicking natural behavior to avoid detection—such as looking forward while covertly using peripheral vision to check another screen. Our use of object detection models like YOLOv5 enables direct localization of external devices (e.g., phones or tablets) and subtle cues such as head turns or partial hand gestures, making the system responsive to behind-the-scenes activity.

Furthermore, to simulate practical diversity, the datasets inherently contained variations in environmental conditions—such as differing lighting setups (natural light, artificial overhead, or low-light rooms), background complexity, and camera angles (e.g., slightly above, straight-on, or tilted webcam views). Models like EfficientNetB1 and ResNet50-CBAM demonstrated strong resilience to these conditions due to their ability to learn fine-grained spatial features and adaptively focus on relevant regions through attention mechanisms. CBAM, in particular, improved robustness by dynamically emphasizing spatial and channel-level features, helping detect unusual posture shifts or light-induced image distortions.

The models also addressed imitation behavior—where users attempt to mimic “normal” behavior—through temporal consistency and anomaly detection across frames. For example, even if a student momentarily behaves correctly, sudden or inconsistent gestures, frequent gaze switching, or atypical face orientations across time frames are flagged as suspicious. Additionally, ResNet and ConvNeXt models’ deep feature extraction capabilities helped distinguish genuine from feigned attentiveness.

Together, this model ensemble supports real-time, context-aware classification, accounting for varied cheating strategies and ensuring adaptability across diverse testing environments. These robustness features are essential for deploying reliable proctoring systems in practical academic settings.

Table 5 presents a detailed comparative analysis between our experimental outcomes and prior research on cheating detection in online examination contexts. The results demonstrate that our evaluated models generally exceed or closely align with state-of-the-art findings reported in existing literature. Specifically, our CNN implementation achieved an accuracy of 97.19% on the OP dataset, significantly outperforming the 93% reported in [19]. Similarly, our ResNet50 model achieved 93.87% on the OEP dataset and 93.75% on OP, surpassing the previously documented 87% in [23].

While ConvNext and Swin Transformer have been investigated in earlier works, our study examines them using more comprehensive datasets (OEP and OP), attaining 92.75% and 90.94% accuracy on OP for ConvNext and Swin Transformer, respectively. These results represent an improvement over earlier findings (88.82% for ConvNext [43] and 90.42% for Swin Transformer [44]). Furthermore, YOLOv5 continued to perform among the top models, consistent with prior benchmarks (95.70%), achieving 94.50% and 94.90% accuracy on the OEP and OP datasets, respectively [17].

In addition to revisiting previously studied classifiers, our work introduces and thoroughly assesses several advanced deep learning models not previously explored in the literature, including EfficientNetB0/B1, MobileNetV2, ResNet50-CBAM, and SE-ResNet. These models exhibited strong and dependable performance across both datasets, with EfficientNetB0 and EfficientNetB1 consistently surpassing 93% accuracy. Notably, ResNet50-CBAM reached 93.70% accuracy on OEP and 93.23% on OP, underscoring its effectiveness in balancing accuracy with architectural complexity.

Recent studies such as Automated Smart Artificial Intelligence-Based Proctoring System Using Deep Learning and Smart Artificial Intelligence-Based Online Proctoring System [45] have made significant strides in addressing cheating detection during online exams by developing a multi-modal system with pre-defined behavioral rules or heuristic methods to detect anomalies such as gaze deviation, head pose changes, or presence of multiple faces. While these approaches are effective in controlled environments, they often rely on limited behavior patterns, fixed camera conditions, and constrained datasets, which can reduce adaptability to real-world scenarios with high variability.

In contrast, our approach advances beyond these frameworks in several key ways. First, we introduce a diverse ensemble of advanced models—including EfficientNetB0/B1, ResNet50-CBAM, YOLOv5, ConvNeXt, and Swin Transformer—each offering unique architectural strengths such as compound scaling (EfficientNet) or attention-based refinement (CBAM). This allows our models to capture subtle cheating behaviors such as screen-looking, device use, and background interference with greater sensitivity and generalization.

Second, while previous smart AI-based systems typically focus on single-camera views and assume stable lighting or clean backgrounds, we benchmark our models using two robust, real-world datasets (OEP and OP) that include diverse lighting conditions, multiple angles, background noise, and imitation behaviors, making our evaluation more representative of actual online exam environments.

Third, our approach is deeply comparative and practical—we not only report accuracy and F1-score but also assess each model’s computational cost and inference time, which is crucial for deployability in real-time exam monitoring systems. This comprehensive evaluation makes our framework not only more scalable and accurate but also more adaptable to different institutional needs compared to existing smart AI-based systems.

Overall, the integration of these diverse and contemporary architectures into our study advances the field of deep learning-based cheating detection, providing compelling evidence for the adaptability and scalability of modern models in supporting academic integrity. Our findings underscore the value of examining a broad spectrum of architectures and datasets to identify models that best meet the operational demands of real-world proctoring environments.

The novelty of our proposed approach lies not only in the individual deep learning models used but also in the way we comprehensively benchmark and adapt them to the context of cheating detection under realistic, diverse, and challenging proctoring conditions. While some models like CNN and ResNet have been used in prior studies, our contribution advances the state of the art by introducing and evaluating more recent and powerful architectures—including EfficientNetB0/B1, ResNet50-CBAM, ConvNeXt, and YOLOv5—in a unified framework across two realistic datasets (OEP and OP). EfficientNet models introduce compound scaling, which allows for efficient depth, width, and resolution balancing, improving performance under constrained environments such as remote proctoring. ResNet50-CBAM integrates attention mechanisms at both the channel and spatial level, making it especially effective for detecting subtle and deceptive behaviors that might escape simpler models. YOLOv5, known for real-time object detection, is fine-tuned for binary classification of cheating behavior, providing both speed and high precision in detecting visual cues such as secondary devices or off-screen glances.

Furthermore, our work is novel in how it tests model robustness against variations such as lighting conditions, webcam angles, imitation behavior, and environmental distractions. Most prior works do not systematically evaluate model resilience in such diverse cheating scenarios. The comparative evaluation across architectures also provides practical insights into balancing accuracy and computational efficiency, making this work a step forward toward scalable, real-world deployment of AI-based proctoring systems.

While our study offers a comprehensive evaluation of deep learning models for cheating detection in online examinations, several limitations must be acknowledged. We primarily focused on CNN-based architectures such as EfficientNet, ResNet variants, and YOLOv5, without fully exploring advanced transformer-based models like Vision Transformers (ViT), DeiT, or hybrid CNN-transformer approaches, which have shown promising results in recent computer vision research. Additionally, our evaluation was limited to two datasets (OEP and OP), which, despite being relevant to online proctoring, may not fully capture the diversity of real-world examination settings in terms of behavior, demographics, and environmental context. Another limitation lies in the unimodal nature of our input data, as we relied solely on visual information from video frames. Incorporating multimodal inputs such as audio signals, gaze tracking, or keystroke dynamics could enhance model robustness and improve detection accuracy in complex scenarios.

This paper presents a comprehensive analysis of deep learning models for detecting cheating in online exams, identifying EfficientNetB0, EfficientNetB1, ResNet50-CBAM, and YOLOv5 as top-performing architectures across the OEP and OP datasets. EfficientNet models and ResNet50-CBAM demonstrated balanced accuracy, efficiency, and strong precision-recall performance, making them suitable for real-time proctoring applications, while YOLOv5 excelled in precision and mAP, showcasing potential for high-stakes, real-time detection. Future work should aim to enhance model robustness through hybrid and ensemble approaches, expand datasets to capture diverse cheating behaviors, and develop lightweight, privacy-preserving, multi-modal frameworks that integrate audio, gaze, and other behavioral cues. Addressing ethical concerns, refining adaptive feedback mechanisms, and optimizing for edge devices will support the deployment of scalable, effective, and user-centric cheating detection systems that uphold academic integrity in diverse online education settings.