Communication-Efficient Federated Point Cloud Analysis via Kronecker Product Factorization
Year: 2026
Type: article
Abstract: Federated Learning (FL) is a vital paradigm for privacy-preserving 3D perception. However, the substantial communication overhead remains a critical bottleneck, hindering FL deployment on resource-constrained devices. In this work, we propose FedKP, a communication-efficient FL framework based on Kronecker Product (KP) factorization for point cloud analysis. By pre-decomposing model parameters into Kronecker factors, FedKP mitigates the communication burden during both uploads and downloads. To further enhance scalability, we extend KP factorization to Sum-of-Kronecker-Products (SKP) factorization, giving rise to FedSKP. Our analysis reveals that FedSKP implicitly achieves multi-branch feature aggregation, enabling the model to capture more diverse local geometric features. Consequently, FedSKP maintains competitive performance while substantially minimizing communication cost. Furthermore, we develop a reconstruction-free computation acceleration strategy to enhance the viability on edge devices. Extensive experiments are conducted on the ModelNet40, ShapeNetPart, and S3DIS datasets. Experimental results show that these frameworks exhibit competitive communication efficiency and model performance compared to relevant baselines.
Source: IEEE Transactions on Multimedia
Institution: Xidian University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7203468604