Joseph Tuazon
All work 05 / 05

SOAP

Deep-Learning Proctoring Research

Year

2025

Role

Researcher / Full-Stack

Stack

Deep LearningPyTorchPCSC 2026
01Overview

Integrating Deep Learning Models to Improve Cheating Detection in UPOU's Specialized Online Assessment Platform. Accepted to the 26th Philippine Computing Science Congress, 2026.

02Abstract

Rule-based proctoring struggles with subtle behavioral cues. I extracted facial landmarks, eyelid features, head pose and object-detection signals from the Online Exam Proctoring dataset into windowed time-series features, then trained and benchmarked five architectures (CNN1D, LSTM, BiGRU, TCN, CNN-LSTM) with subject-wise five-fold cross-validation and nested hyperparameter tuning. Every model beat the SOAP baseline; CNN-LSTM led overall.

0.474
F1 (CNN-LSTM)
0.656
ROC-AUC
+40%
F1 over baseline
+67%
PR-AUC over baseline
03What I built
  • Feature pipeline

    Built the extraction pipeline turning raw proctoring video into normalized, windowed behavior sequences for training.

  • Five architectures

    Engineered and benchmarked CNN, LSTM, BiGRU, TCN and CNN-LSTM with a faithful recreation of SOAP's rule-based baseline as the comparison point.

05Tech
PythonOpenCVMediaPipeKerasTensorFlowPyTorchPandasNumPy
·Access

Read the paper

The full paper is available on request. Email me and I'll send it over.

Request the paper
Next project — Pedestrian Mobility GIS
Tulay