Go to ICLR 2026 Conference homepageFedSumUp:Secure Federated Learning Without Client-Side Training for Resource-Constrained Edge Devices
Keywords: Federated Learning, Data Condensation, Server-Side Optimization, Privacy-Preserving, Edge Devices, Variational Autoencoder
Abstract: Horizontal Federated Learning (HFL) enables multiple clients with private data to collaboratively train a global model without sharing their local data. As a research branch of HFL, Federated Data Condensation with Distribution Matching (FDCDM) introduces a novel collaborative paradigm where clients upload small synthetic datasets instead of gradients and parameters. FDCDM faces two key challenges: privacy leakage risk, where synthetic data may leak the privacy of real data; and high computational cost on the client side, which limits the deployment capability of FDCDM on resource-constrained devices. To address these challenges, we propose FedSumUp, an improved FDCDM method. The core designs of FedSumUp include: generating initial data templates based on a Variational Autoencoder (VAE); and migrating the entire synthetic data optimization process to the server side, requiring clients only to upload distilled synthetic data and the mean of raw data features without exposing the original data itself. Experimental results on multiple real-world datasets demonstrate that FedSumUp achieves notable advantages in the following aspects: drastically reducing the visual similarity between synthetic and real data, and effectively resisting membership inference attacks; significantly lowering client-side computational overhead, making it deployable on edge devices. FedSumUp is the first work to systematically analyze privacy risks in FDCDM from the perspective of data similarity, providing a new direction for building efficient and privacy-preserving federated learning frameworks.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Submission Number: 23402
Loading