In an academic environment, adherence to the code of academic integrity stands as tangible evidence of a strong culture of honesty, competence, and the credibility of one's acquired knowledge. However, in realistic practice, there are various loopholes that can be exploited by certain individuals in pursuit of the best possible results — and this is particularly relevant in the context of examinations.

For this reason, up-to-date cheating detection methods are increasingly necessary to identify and address dishonest behavior during exams.

In response to this challenge, lecturers from the Informatics Engineering study program at Universitas Dian Nusantara (UNDIRA), Mr. Uus Rusmawan, S.Pd., M.Kom., and Mr. Imam Mulya, S.Kom., M.M.S.I., along with their Informatics Engineering students, developed an exam cheating detection system model based on SURF-CNN, which has been successfully published in a SINTA 4-accredited journal.

This research focuses on a cheating detection approach that leverages machine learning through the SURF (Speeded Up Robust Features) method, complemented by a deep learning algorithm specifically designed for visual image detection, namely CNN (Convolutional Neural Network).

It is worth emphasizing that this research centers on detecting cheating through the identification of suspicious objects and movements, rather than through an analysis of exam answer sheets.

Based on prior research, the implementation of CNN in visual data imaging has consistently yielded relatively stable results, with an average accuracy score of 87.9%. However, when used independently, CNN's limitations lie in its inability to effectively analyze dynamically moving objects, sudden changes in scale, and insufficient lighting conditions, all of which pose significant challenges for the Informatics Engineering team, as these factors can mislead the detector.

This is where the SURF model steps in to address CNN's shortcomings. SURF is responsible for detecting and describing interest points, alongside key points within visual image data. The strength of SURF lies in its robustness to changes in scale, rotation, and lighting variation. In other words, SURF first "filters" and "highlights" the most relevant visual information from each image frame before passing the data on to CNN for classification.

As a result, CNN no longer works from scratch with raw data; instead, it receives a far richer and more structured feature representation and this is ultimately what drives the overall improvement in model accuracy.

Building on this understanding, the data collection process began with the implementation of data protection and recovery using SnapMirror technology, followed by the recording of facial images, head posture, and eye gaze of exam participants using a mounted HIKVISION 5MP 3K camera capable of recording video at a resolution of 1920×1080p at 24 fps. To minimize data imbalance across classes, data augmentation was applied to prevent class imbalance. This process yielded approximately 1,200 samples, with an average of 200 image data per class.

The subsequent stage involved evaluation, data categorization, and a comparative analysis between the standalone CNN system and the SURF-CNN combination, across six behavioral movement categories: bored, focused, confused, dizzy, suspicious head-turning, and suspicious glancing.

Based on the evaluation results, the hybrid SURF-CNN model consistently demonstrated superior performance, recording an overall accuracy rate of 91.80%, while the standalone CNN model only achieved 88.20%. This advantage is most clearly observed in the detection of suspicious head-turning behavior, highlighting the model's enhanced capability in analyzing dynamic movements compared to the standalone CNN.

This achievement affirms that the hybrid approach, combining SURF's interest point-based feature extraction with CNN's deep classification capabilities, is the right direction for building a more reliable and accurate exam cheating detection system. This research also serves as concrete evidence that the academic community at UNDIRA is not only committed to upholding academic integrity on a normative level, but is also actively delivering real technological solutions to realize it.