Dissertation Title: Detect Cheating from Multi-modalities
Abstract:
When online assessments became more common after COVID, greater challenges arose for cheating detection without in-person monitoring. This paper proposes a machine learning system for detecting cheating in online assessments with multi-modal data collected from the computer achieving an accuracy of 86% on the test set.
We conducted a laboratory experiment with 103 subjects, of which 102 were usable, and efficient data cases (31 cheating class 1, 71 non-cheating class 0). The standardized multi-modal response data (i.e., mouse movement, keyboard dynamics, audio, video, and results reported) were collected.
We initially computed ’delta (∆)’ features representing the features’ difference between baseline tasks and cheating detection tasks to measure the change in behaviors. Furthermore, we first integrated multiple modalities into cheating detection. A stacking model was trained by the predictions of single-modal models and outperformed other approaches to integrating multi-modal data comprising majority voting and concatenation of multi-modal features.