Accuracy of Machine Learning Models for Anomaly Detection in Online Proctored Examinations: A Review of the Potential of Haar Cascades for Mobile Based Proctoring

Abstract

The rapid shift toward digital learning in higher education has made online examinations an everyday reality, increasing the need for reliable and trustworthy anomaly detection systems powered by machine learning (ML). While recent advances in deep learning have improved the technical capabilities of automated proctoring tools, many existing solutions remain difficult to deploy fairly and effectively; particularly in mobile-first environments that are common in low resource settings. Challenges related to computational demands, accessibility, and algorithmic bias continue to limit their practical impact. This review explores the accuracy and suitability of ML models used for anomaly detection in online proctoring, with particular attention given to Haar cascade classifiers; a classical yet computationally efficient approach that remains widely supported on mobile platforms. A structured literature search across seven major databases identified 150 relevant studies, from which 20 met the defined inclusion criteria. The findings show that deep learning models, especially convolutional neural networks (CNNs), consistently achieve the highest detection accuracy but often require substantial computational resources. In contrast, Haar cascades offer fast, low latency detection suitable for mobile devices, although their performance declines under challenging conditions such as poor lighting, pose variation, and facial occlusion. Notably, hybrid approaches that combine Haar cascades with lightweight CNNs emerge as a promising middle ground, balancing efficiency with improved robustness. However, the review also highlights important research gaps, including the scarcity of mobile-centric datasets, limited real world field evaluations, and insufficient testing for fairness across diverse demographic groups. Addressing these gaps will require future research to prioritize mobile optimized model design, standardized evaluation benchmarks, privacy aware computation strategies, and broader empirical validation in authentic exam settings.