Congratulations to Mohammad Partohaghighi for passing EECS Ph.D. Proposal Defense
Mr. Mohammad Partohaghighi is now (as of 11/21/2025) a Ph.D. candidate at the EECS Graduate Program. He passed his Ph.D. proposal defence on 10/24/2025. The exam committee gave unanimous pass to Mohammad! Congratulations!!
Web: https://mechatronics.ucmerced.edu/fc4fl
Title: Smart Federated Learning and Unlearning: Theoretical Framework, Benchmark Studies, and Applications,
Abstract: Federated Learning (FL) enables collaborative machine learning across distributed devices while preserving data privacy, but challenges like client drift in non-independent and identically distributed (non-IID) settings, slow convergence, adversarial updates, and privacy-compliant data removal limit its efficacy. This thesis introduces and evaluates five novel algorithms—Roughness-Informed Federated Learning (RIFedAvg), Fractional Order Federated Learning (FOFedAvg), WeightWatcher Federated Learning (WWFedAvg), Roughness-Informed Federated Unlearning (RIFU), and Physics-Informed Federated Proximal (PI-FedProx)—to address these issues, using Federated Averaging (FedAvg) as a baseline. RI-FedAvg mitigates client drift through adaptive regularization based on a Roughness Index that captures data heterogeneity. FOFedAvg enhances convergence using memoryaware updates derived from fractional calculus. WWFedAvg strengthens robustness against adversarial updates via spectral analysis of model weights. RI-FU enables efficient, privacy-compliant unlearning through clustering-based selective data removal, ensuring compliance with regulations. PI-FedProx integrates partial differential equation (PDE) constraints into the FedProx framework, as described in Chapter 7, using adaptive sampling of collocation points to minimize PDE residuals. This physics-informed approach addresses non-IID data and data scarcity by enforcing physical consistency through PDEs, improving model accuracy and robustness in scientific applications like fluid dynamics and pollutant dispersion where physical laws govern system behavior. Evaluated across classification, energy consumption modeling for electric vehicles, heart disease prediction, and PDE-based scientific tasks, these algorithms improve robustness, convergence speed, privacy, and physical consistency. This work advances FL by providing a comprehensive framework that balances performance, security, privacy, and physical consistency, offering practical solutions for decentralized machine learning in real-world scenarios.
Created and Updated by Prof. YangQuan Chen 11/21/2025
