Communication-efficient Federated Learning
Abstract
Advances in deep learning have revolutionized numerous fields, yet deploying these models often requires aggregating massive datasets in a central location, raising critical privacy and scalability concerns. Federated Learning addresses this by enabling collaborative model training directly on distributed client devices without sharing private data. This thesis addresses communication overhead, computational inefficiency, client heterogeneity, and personalization through three contributions: EvoFed, an evolutionary-strategy-based method that exchanges compact fitness similarity metrics; Model-Agnostic Projection Adaptation, a unified low-rank factorization method that compresses the model parameter space; and Principal-Aligned LoRA, a personalized approach that uses singular value decomposition to align client-specific updates. Together, these methods improve the efficiency, scalability, and personalization of federated learning.
Canonical citation
Rahimi, M. M. (2025). Communication-efficient Federated Learning [Ph.D. dissertation, Korea Advanced Institute of Science and Technology].
@phdthesis{rahimi2025communication,
title = {Communication-efficient Federated Learning},
author = {Rahimi, Mohammad Mahdi},
school = {Korea Advanced Institute of Science and Technology},
year = {2025},
type = {Ph.D. dissertation}
}
Archive note: this author-hosted PDF preserves the visible pages of the official version. Only embedded document metadata was normalized so title and author fields match the title page.