فواد ياسر

This thesis provides a holistic solution for AI-assisted federated sensor fusion for privacy-preserving location context profiling on mobile devices. While the traditional GPS-based method is accurate, it is grossly energy-inefficient an‎d privacy-violating. To overcome this, the proposed method employs the low-energy IMU sensorsʹ accelerometers an‎d gyroscopes to derive semantic location contexts like Indoor-Home, Indoor-Public, an‎d Outdoor-Moving. The hybrid machine learning solution combines Ran‎dom Forest, Support Vector Machines, an‎d Multi-Layer Perceptrons. To encourage the minimal usage of computational resources, local Ran‎dom Forest models were optimized with a reduced tree depth (15) an‎d reduced numbers of estimators (50), which are mobile-friendly for CPUs. Federated Learning (FL) is utilized via the FedAvg algorithm to manage the joint training of 30 distributed clients without sharing raw data. Based on the UCI HAR dataset in a cloud simulation setup (Google Colab) with Python 3.12, the experimental outcome verifies that the framework is 99.84% accurate in centralized classification an‎d retains 96.7% accuracy in federated settings. The system provides 57.2% energy savings over GPS localization. This approach will empower future location-aware mobile systems to strike a balance between high accuracy, full data privacy, an‎d energy efficiency through on-device computation.

يادگيري فدرال , ادغام حسگرها , محاسبات آگاه از متن , حسگرهاي IMU

Federated learning , sensor fusion , text-aware computing , IMU sensors

Fouad Yasir

Dr.Abolfazl Diyanet