A Two-Stage, Object-Centric Deep Learning Framework for Robust Exam Cheating Detection

1. Problem Statement

Academic integrity is increasingly threatened by examination cheating, particularly in remote and hybrid learning environments. Traditional invigilation is inefficient, costly, and prone to human error at scale. While AI-powered monitoring systems exist, they face three critical limitations:

• Background Noise: Existing "full-frame" classification approaches treat the entire image as a single input. In multi-person exam settings, the actual cheating action occupies a small fraction of pixels, causing models to learn spurious correlations from irrelevant background elements.
• Lack of Granularity: Holistic models cannot localize who is cheating or where the action is occurring, making outputs non-actionable for proctors.
• Data Scarcity: The field lacks large-scale, diverse, and publicly available datasets. Most studies rely on small, private, or staged datasets, leading to overfitting and an inability to benchmark models fairly.

2. Methodology

The authors propose a Two-Stage, Object-Centric Framework that decouples scene understanding into localization and classification.

Stage 1: Object Localization (Detection)

• Model: YOLOv8n (Nano version).
• Function: Processes the input exam room image to detect human-like objects.
• Filtering: The system retains only bounding boxes labeled as "person" and discards other objects.
• Preprocessing: The detected regions are cropped to create Regions of Interest (ROIs). This step isolates the student from the background, significantly reducing noise.
• Rationale: YOLOv8n was chosen for its optimal trade-off between accuracy and computational efficiency (low latency), making it suitable for edge deployment in educational institutions.

Stage 2: Behavioral Analysis (Classification)

• Model: RexNet-150 (a variant of ResNet).
• Input: The cropped ROIs from Stage 1 are resized to 224x224 pixels.
• Training Strategy:
  • Transfer Learning: Initialized with weights pre-trained on ImageNet.
  • Fine-tuning: The model is fine-tuned on the curated dataset to distinguish between "cheating" and "not cheating."
  • Hyperparameters: Adam optimizer, learning rate of 3e-4, batch size of 16, trained for 10 epochs.
• Output: A binary classification label for each student.

Dataset Construction

To address data scarcity, the authors consolidated and standardized labels from 10 independent public sources (e.g., Roboflow datasets).

• Total Samples: 273,897 images.
• Standardization: Labels were mapped to a binary class: 1 (Cheating) and 0 (Not Cheating).
• Split: 80% Training (219,117), 10% Validation (27,389), 10% Testing (27,391).

3. Key Contributions

1. Novel Two-Stage Framework: A shift from holistic scene analysis to an object-centric approach. By isolating students before classification, the system eliminates background noise and enables granular, per-person assessment.
2. Large-Scale Standardized Dataset: The creation of a unified benchmark dataset from 10 diverse sources, facilitating robust training and reproducible evaluation for the research community.
3. Comprehensive Validation: Extensive experimental validation, including ablation studies and backbone comparisons, establishing a new state-of-the-art performance.
4. Ethical Considerations: The system is designed to deliver results privately (e.g., via email) to prevent public shaming, allowing students to reflect on their behavior.

4. Experimental Results

The framework was evaluated on a test set of 6,895 samples (5,057 normal, 1,838 cheating).

• Performance Metrics:

  • Accuracy: 95.16%
  • Precision (Cheating): 0.91
  • Recall (Cheating): 0.91
  • F1-Score (Cheating): 0.91
  • Inference Time: 13.9 ms per sample (demonstrating scalability).

• Ablation Study (Two-Stage vs. Full-Frame):

  • A baseline model using the same RexNet-150 architecture but trained on full, uncropped images achieved only 78.52% accuracy and an F1-score of 0.63.
  • Improvement: The proposed two-stage method yielded a 16.64% absolute increase in accuracy and a 28% increase in F1-score, proving that isolating subjects is critical for performance.

• Backbone Comparison:

  • RexNet-150 outperformed ResNet-18 (93.81% acc) and EfficientNet-B0 (94.55% acc), justifying the choice of the larger model for this nuanced task.

• Error Analysis:

  • False Positives: Often caused by ambiguous gestures (e.g., resting chin on hand) lacking temporal context.
  • False Negatives: Occurred when cheating evidence (e.g., a phone on a desk) fell outside the tight "person" bounding box, meaning the classifier only saw the student's face/head.

5. Significance and Future Work

• Significance: This study demonstrates that focusing computational resources on specific subjects rather than the entire scene significantly improves detection reliability. It provides a scalable, open-source foundation for ethical, real-time proctoring systems that balance high accuracy with privacy.
• Future Directions:
  • Temporal Analysis: Incorporating consecutive frames to distinguish between innocent gestures (scratching head) and suspicious ones (whispering).
  • Expanded ROI: Widening the bounding box to include hands and desk areas to capture peripheral evidence (e.g., phones on the desk).
  • Multi-Class Classification: Moving beyond binary labels to categorize specific types of cheating (e.g., phone usage vs. peer collaboration).
  • Audio Integration: Adding audio data to detect whispering or unauthorized communication.

In conclusion, the paper presents a robust, high-accuracy solution to exam cheating that overcomes the limitations of previous holistic approaches through a strategic, object-centric deep learning pipeline.