🧠 Online Exam Proctoring – Cheat Detection (OEP Dataset)

This project builds a Cheating Detection System using the Online Exam Proctoring (OEP) dataset.
The model classifies each webcam frame into:

• Cheating
• Not Cheating

A 2D Convolutional Neural Network (CNN) is trained on extracted grayscale frames to detect suspicious behavior automatically during online examinations.

---

📂 Dataset Overview

Dataset Used: Online Exam Proctoring (OEP) Dataset (from Kaggle)
Link: https://www.kaggle.com/datasets/raajanwankhade/oep-dataset

Each subject folder contains:

• Webcam video files (*1.avi)
• Ground truth file gt.txt with cheating time intervals

Expected structure:

OEP database/
├── subject1/
│ ├── subject1_1.avi
│ └── gt.txt
├── subject2/
│ ├── subject2_1.avi
│ └── gt.txt
└── …

Ensure the dataset folder name matches this in the code:
data_path = "OEP database"

---

🏗️ Project Structure

project/
├── main.ipynb # Main code: Data preprocessing, training & evaluation
└── README.md # Documentation

---

⚙️ Workflow

✔ Extract frames from videos (1 FPS sampling)
✔ Convert to grayscale & resize to 48×48
✔ Frame-wise label creation using time intervals
✔ Subject-based train/test split
✔ CNN training for binary classification
✔ Model evaluation using standard metrics

---

🧠 Model Architecture

• Input: 48 × 48 grayscale frame
• Layers:
  • Conv2D → ReLU → MaxPool
  • Conv2D → ReLU → MaxPool
  • Fully Connected layers with dropout
  • Softmax for 2-class output

Lightweight for quick experiments & educational use.

---

🚀 How to Run

1️⃣ Place dataset next to notebook:

project/
├─ main.ipynb
└─ OEP database/

2️⃣ Open & Run All Cells in main.ipynb

3️⃣ Metrics and plots will be displayed at the end

Note: First run may take longer due to frame extraction

---

🔧 Installation / Requirements

Install via notebook or manually:

pip install torch torchvision torchaudio
pip install opencv-python numpy pandas scipy
pip install scikit-learn matplotlib seaborn scikit-plot

GPU Recommended (CPU works but slower)

📊 Evaluation Metrics

The following are displayed:

Accuracy

Precision

Recall

F1-Score

Confusion Matrix

ROC Curve

Performance depends on subject split & number of frames extracted.

🔮 Future Enhancements

Temporal models (LSTM / GRU / 3D CNN / Transformers)

Use additional camera views (wearable/desktop)

Stronger backbones (ResNet, EfficientNet)

Imbalance handling strategies (weighted loss / oversampling)

📌 Disclaimer

This project is made only for academic and research purposes and should not be used in real-world proctoring without proper validation.

📚 Credits

Dataset: Online Exam Proctoring (OEP) Dataset
Developed by: Anas Multani
B.Tech – Computer Science & Engineering
Nirma University
Roll No: 23BCE188