Balancing Privacy, Performance, and Efficiency in Federated Learning for the Internet of Things - دانشکده فنی و مهندسی
Balancing Privacy, Performance, and Efficiency in Federated Learning for the Internet of Things
نوع: Type: Thesis
مقطع: Segment: masters
عنوان: Title: Balancing Privacy, Performance, and Efficiency in Federated Learning for the Internet of Things
ارائه دهنده: Provider: Mahnoush Sefidabian
اساتید راهنما: Supervisors: Dr. Mehdi Sakhaei-nia
اساتید مشاور: Advisory Professors: Dr. Moharram Mansoori Zadeh
اساتید ممتحن یا داور: Examining professors or referees: Dr. Morteza Yousef Sanati،Dr.Reza Mohammadi
زمان و تاریخ ارائه: Time and date of presentation: 2026
مکان ارائه: Place of presentation: seminar
چکیده: Abstract: Recently, federated learning (FL) has emerged as a transformative distributed paradigm that trains machine learning models without exposing raw data, offering a robust infrastructure for privacy-preserving collaborative intelligence. However, the iterative exchange of high-dimensional model parameters between clients and the central server incurs substantial communication overhead and leaves the system vulnerable to gradient-based inference attacks. These bottlenecks are severely exacerbated under heterogeneous, non-independent and identically distributed (Non-IID) data distributions. To resolve these challenges, this study proposes EA-SEC, a unified framework for secure and communication-efficient FL. The synergistic combination of the four mechanisms including adaptive sparsification, error compensation, homomorphic encryption, and Schnorr authentication works together to balance the model usefulness, communication effectiveness, and privacy.Within this architecture, adaptive sparsification dynamically restricts the volume of transmitted updates, error compensation mitigates the convergence degradation induced by severe compression, and the integrated cryptographic protocols guarantee strict confidentiality and integrity during global aggregation. Comprehensive empirical evaluations on the MNIST and CIFAR-10 datasets validate the framework’s superiority. On the MNIST dataset, EA-SEC achieved an accuracy of 98.40%under homogeneous conditions and 95.24%under heterogeneous (Non-IID) configurations. To reach target accuracy thresholds of 90%and 95%, the communication overhead was constrained to 70.03MB and 103.40MB (homogeneous), and 277.07MB and 494.07MB (heterogeneous), respectively. Ablated variants fundamentally failed to meet these targets under identical constraints. Furthermore, on the highly complex Non-IID CIFAR-10 dataset, EA-SEC achieved a final accuracy of 62.85%, outperforming the baseline (60.40%) and competing algorithms. To reach 50%and 60%accuracy, EA-SEC required only 740.99MB and 1551.19MB of data transmission, demonstrating a substantial reduction compared to the 986.70MB and 1842.50MB demanded by the baseline. Overall, the findings establish that the EA-SEC framework provides a highly scalable and robust solution for neutralizing the critical privacy-efficiency trade-offs in heterogeneous FL environments.