Scientific research project

Abstract

Exam cheating is the use of prohibited actions to gain illegal advantage during testing and examination. Cheating in exams has far-reaching consequences, detrimental to the quality of training, causing negative consequences and damaging the reputation of educational institutions, hindering the development of education. country. It makes learners less capable of learning and creating, leading to the training of unqualified individuals, thereby creating a gap in knowledge and skills in the workforce in society. This research project uses a deep learning approach to use an estimate of the candidates' posture to determine if they are cheating and will alert the proctor when it detects that a student is acting inappropriately often. In this paper, we propose action recognition methods based on LSTM, VGG19, Faster R-CNN, Mask R-CNN and Vision Transformer network architectures to classify them as normal actions or actions anomalies such as viewing friends, passing documents, jumping up, exchanging articles based on the collected data set. Through the training results, we found that the Vision Transformer model when detecting fraudulent and normal actions has an accuracy of more than 98%. This contributes to supporting teachers in managing candidates in exams, contributing to creating fairness and transparency in education in the country.

Introduction

Today, exams and assessments are integral components of the educational curriculum in schools. In the process of teaching and learning, modern technologies have been applied extensively. However, a significant issue that arises is the phenomenon of cheating during exams. It is common to observe that students often smuggle in materials to cheat during exams, which sometimes leads to their being caught and reprimanded by teachers or invigilators. Additionally, students may secretly discuss or exchange answers during classes or exams when invigilators are not paying attention. In Vietnam, according to Circular No. 15/2020/TT-BGDĐT, which outlines the regulations for high school graduation exams as of May 26, 2020, cheating is defined as actions that violate the exam council's regulations, such as copying answers or bringing unauthorized materials into the exam room. Cheating is now more overt and not hidden; many are aware of it but choose not to speak out, despite various measures and penalties implemented by schools. Cheating has detrimental effects on students, fostering a reliance on dishonest practices and discouraging them from striving for genuine achievements. The issue of cheating is prevalent across all educational levels, from primary to higher education, and includes alarming levels of grade manipulation, degree fraud, and title buying. Cheating not only increases in scale but also becomes increasingly sophisticated and harder to detect.

Proposed Method

In this study, we utilize machine learning methods based on deep learning techniques and Vision Transformer networks to detect patterns of cheating behavior during exams. The proposed model consists of two stages: training and testing. The details of these stages are illustrated in the diagram in Figure 1.

Figure 1. Proposed modell

Data Preprocessing: The training dataset consists of real-world videos captured during student exams or tests. Each video has a duration of 5 to 10 seconds with a frame rate of 30 FPS. The videos are converted into images or frames for data preprocessing. Many pre-trained models, including Vision Transformer (ViT), utilize the ImageNet dataset for initialization or pre-training. Consequently, input images are often resized to 224 x 224 pixels to ensure compatibility and consistency across different models.

Feature Extraction and Training: Leveraging the advantages of deep learning networks previously discussed, we extract features of different types of actions from the preprocessed images using three models: LSTM, VGG19, and Vision Transformer. Specifically, for the Vision Transformer model, the process involves preprocessing the input images and then dividing them into patches. These patches are passed through linear projection, with each patch assigned specific positional and label information to ensure that the features extracted from the image remain intact and undistorted. The patches are then fed through a Transformer encoder for encoding and classification. The outcome of this process is the classification of the images into their respective labels.

To conduct experiments for the proposed model, we implement two scenarios with the training parameters outlined in Table 1.

Table 1. Proposed Scenarios

Results

Table 2 presents experimental images illustrating the performance of the three proposed models. For the information exchange scenario, all models successfully identify the action of exchanging information. However, Scenario 1 achieves a high accuracy rate in detecting the interaction between two students, whereas Scenarios 2, 3, and 4 correctly identify one student but misclassify the second student as rise_up (standing up). In Scenario 5, only one student engaged in cheating is detected. In the document transfer scenario, all scenarios successfully identify the action with relatively good results, with Scenario 1 providing the best prediction accuracy, approaching 99%. For the card looking scenario, Scenario 1 achieves absolute accuracy with 100%. However, Scenario 2 fails to detect the cheating action, possibly due to lower accuracy in the recognition and classification process. The remaining scenarios all successfully identify the action with relatively good results. For the rise_up and normal scenarios, Scenario 2 fails to recognize the action, possibly due to less distinct action representation. Conversely, Scenarios 1, 3, 4, and 5 achieve relatively high accuracy, with Scenario 1 providing the highest accuracy among them. These experimental results demonstrate that Scenario 1 outperforms the other scenarios in detecting and classifying cheating actions effectively.

Figure 3. Testing results

Conclusions

Cheating in exams is a common issue among students, with significant consequences such as fostering bad habits and ethical issues. This paper introduces new contributions by creating a training dataset for detecting cheating behaviors, demonstrating the effectiveness of deep learning models in identifying such behaviors. Utilizing advanced architectures like LSTM, VGG19, Faster R-CNN, Mask R-CNN, and Vision Transformer, we achieved over 85% accuracy in detecting cheating actions like looking at another's paper, standing up, exchanging papers, and passing documents. The Vision Transformer model, despite longer training times, provided the highest accuracy at over 98%, allowing for effective and timely detection. Future work will focus on recognizing a broader range of cheating behaviors and incorporating additional features, such as audio and gesture data, to enhance detection capabilities.

References