A Survey on Efficient Federated Pruning: Progress, Challenges, and Opportunities

Federated Learning (FL) has emerged as a cornerstone for privacy-preserving decentralized machine learning, yet its practical deployment is frequently hindered by a Triple Constraint: the limitation of memory, computational, and communication costs on edge devices. While conventional model compression techniques primarily target post-training efficiency or isolated resource bottlenecks, federated pruning has gained traction as a holistic solution capable of addressing all three constraints throughout the training stage. This paper provides the first comprehensive survey dedicated specifically to federated pruning, distinguishing between Communication-Centric Pruning, which sparsifies only transmitted updates, and On-Device Efficient Pruning, which reduces the memory, computations, and communication costs to achieve training efficiency. We propose a novel taxonomy that categorizes the field based on topology heterogeneity, constraint dynamics, and optimization objectives, spanning Federated Homogeneous Pruning, which includes sparse training and iterative pruning, and Federated Heterogeneous Pruning, which includes split learning and personalized pruning to accommodate device and data heterogeneity. By systematically reviewing methodologies and performance across vision and language tasks, this survey identifies critical gaps in the current literature and outlines a research roadmap focusing on the interplay between sparsity and the fundamental FL challenges of privacy, robustness, and hardware acceleration.