The problem of cheating in education is becoming more and more serious today. This is because educational platforms are becoming more diverse and rapidly developing for online learning. This research conducts a small-scale comparative study on five popular machine learning models to detect cheating behavior in online learning systems. We use the Junyi Academy dataset to serve this study. The dataset contains up to 12,537 student interactions, which are evaluated by models including XGBoost, LightGBM, Random Forest, AdaBoost, and TabNet under different noise conditions to assess their robustness in real-world scenarios. Our results demonstrate that AdaBoost achieves the highest overall detection capability with over 95% recall and 97% accuracy under noise conditions under 30%. On the other hand, Random Forest exhibits exceptional noise resistance, maintaining strong performance even at 40% label noise with 92% accuracy. Our superior performance may come from large-scale behavioral datasets rather than synthetic grade patterns. We conclude that AdaBoost and Random Forest will bring great potential for applying Machine Learning model building to anti-cheating features for online learning systems. This research also indicates that deep learning models like TabNet may be less suitable for noisy educational tabular data than ensemble methods. However, we will need to conduct more research on other deep learning models to clarify this point.