Summary
This paper (arXiv:2609.15950) by Yilin Xu, Chun Hei Michael Shiu, Chih Wei Ling, and Linqi Song addresses a structural misalignment in record-level differentially private personalized federated learning: client-specific variation is often low-dimensional, yet training repeatedly releases high-dimensional model updates. The authors propose releasing a private client context only once and confining repeated adaptation to a fixed coefficient space. Beyond dimensionality reduction, the factorized generator induces an adaptive optimization geometry that reshapes noisy updates; controlled ablations show that most of its private-training gain is retained through radial evolution. To cut communication cost, the Gaussian mechanism for coefficient updates is realized directly via variable-length quantization with finite expected code length, allowing quantization error to serve as the required privacy perturbation rather than extra distortion. Experiments on MNIST and CIFAR-10 show the design matches or outperforms full-model private adaptation across various privacy budgets and client heterogeneity settings, while reducing protected uplink communication by 2.67x on CIFAR-10 at epsilon = 16 with comparable accuracy for future clients.
Paper Overview
- Field: Machine Learning
- Authors: Yilin Xu, Chun Hei Michael Shiu, Chih Wei Ling, Linqi Song
- Published: 2026-09-14
- arXiv: 2609.15950
Abstract
Record-level differential privacy exposes a structural misalignment in personalized federated learning when client-specific variation is low-dimensional while training repeatedly releases high-dimensional updates. This paper addresses the misalignment by releasing a private client context once and confining repeated adaptation to a fixed coefficient space.
Beyond dimensionality reduction, the factorized generator induces an adaptive optimization geometry that reshapes noisy updates; controlled ablations show that most of its private-training gain is retained by radial evolution. To further reduce communication cost, the authors realize the Gaussian mechanism for coefficient updates directly through variable-length quantization with finite expected code length, so that the quantization error itself acts as the required privacy perturbation rather than additional distortion.
On MNIST and CIFAR-10, the proposed design matches or outperforms full-model private adaptation across various privacy budgets and client heterogeneity settings, while reducing protected uplink communication by 2.67x on CIFAR-10 at ε = 16, with comparable accuracy for future clients.
*Auto-collected on 2026-09-16.*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178634869