Cheat detection in digital exams with artificial intelligence: Iğdır University example

In this study, artificial intelligence models Random Forest, XGBoost, K-Nearest Neighbors (KNN), and Logistic Regression were employed to analyze students' behavior during online exams to detect potential cheating. The primary goal was to classify behaviors as indicative of cheating (positive) or not (negative) by evaluating similarities in student actions during exams. The analysis revealed that Random Forest achieved the highest performance with an accuracy of 66.29%, followed by XGBoost at 58.98%, while KNN and Logistic Regression demonstrated lower accuracy and struggled to distinguish between classes effectively. According to ROC curve analysis, Random Forest and XGBoost excelled in differentiating positive and negative classes, whereas KNN and Logistic Regression showed weaker performance. Additionally, the impact of individual features on the models was examined, with factors such as "number of logins," "correct answer ratio," and "incorrect answer ratio" found to significantly influence model performance. Random Forest and XGBoost remained robust even when certain features were removed, whereas KNN and Logistic Regression models were more sensitive to such changes. Recommendations to enhance performance include exploring more complex deep learning models, refining feature engineering, and expanding the dataset. Ultimately, Random Forest and XGBoost emerged as the most reliable and effective models for ensuring online exam security and detecting cheating, offering valuable insights for developing strategies to prevent academic dishonesty in remote education platforms.