Ibrahim Aslam - Thesis Defence
A FRAMEWORK FOR STORAGE REDUCTION OF VEHICULAR IOT DATA
Master of Science (Computer Science): Ibrahim Aslam
02 December 2025
3:00 PM Atlantic
Join the meeting now
Meeting ID: 275 019 616 589 3
Passcode: A6iJ3rQ6
Thesis Committee:
Dr. Sazia Mahfuz, Supervisor
Dr. Farhana Zulkernine, Queen’s Univ., External Examiner
Dr. Lydia Bouzar-Benlabiod, Internal Examiner
Dr. Darcy Benoit, Director of School
Dr. Erin Crandall, Chair of the defence
Abstract
Vehicular Internet-of-Things (IoT) systems generate large multivariate sensor streams that are valuable for road condition monitoring and predictive analytics, yet the continuous accumulation of high-frequency data creates practical challenges for storage, transmission, and real-time use. This thesis proposes a storage reduction framework that integrates learning based encoding and classification to enable efficient storage and real-time interpretation of vehicular sensor data.
For developing the storage reduction component of our framework, we explored different deep learning architectures. The framework uses autoencoder-based designs and evaluates four model types: Transformer, Convolutional Neural Network (CNN), Convolutional Neural Network-Long Short Term Memory (CNN-LSTM), and Multi-Layer Perceptron (MLP). Results show that the Transformer based supervised autoencoder achieves the lowest reconstruction error (≈ 0.229), outperforming CNN (≈ 0.263), CNN-LSTM (≈ 0.294), and MLP (≈ 0.373). Reconstructed signals maintain high classifier accuracy, demonstrating that the reduction process does not erase task relevant information. These outcomes validate the proposed framework’s ability to achieve storage reduction without affecting information quality or classification accuracy. Statistical tests confirm that performance differences between models are significant and repeated run-wise evaluations illustrate the stability of these results.
Beyond individual model comparisons, the thesis contributes an end-to-end vehicular data reduction pipeline that implements the trained encoders for continuous, real-time encoding and class-aware logging. This pipeline compresses multi-sensor data streams by over 95-98% while maintaining near perfect road surface recognition, offering a deployable foundation for efficient data transfer and improved privacy in vehicular IoT systems.
In summary, the work goes beyond testing models to build a working system that reduces data size and improves storage efficiency for connected vehicles. The proposed framework also has strong potential for deployment on edge devices where low-latency and energy-efficient processing are important.
About Ibrahim…
I’m a Master’s student in Computer Science at Acadia University. My research focuses on developing deep learning frameworks for efficient data reduction and road-surface recognition in vehicular IoT systems. The goal of my work is to reduce sensor data storage. I’m passionate about artificial intelligence, machine learning and cybersecurity and I enjoy building systems that connect research with real-world applications. Outside of research I like mentoring students, working on technology projects and exploring new advances in AI.