Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Age Aware Scheduling for Differentially-Private Federated Learning

Kuan-Yu Lin, Hsuan-Yin Lin, Yu-Pin Hsu, Yu-Chih Huang

TL;DR

This work studies differential privacy in federated learning with time-varying, Markovian client data and introduces age-aware scheduling to manage the privacy-accuracy trade-off. By deriving an upper bound on the loss difference caused by data aging and modeling data evolution via birth-death Markov chains, the authors formulate an optimization to schedule data collection and adapt noise levels under a global privacy cap. The proposed protocol jointly optimizes scheduling and adaptive DP noise, yielding tighter privacy-utility bounds than non-scheduling baselines and demonstrating superior performance in simulations. The results provide practical scheduling strategies for DP-FL in dynamic databases, with implications for privacy-aware FL deployments in real-time or streaming data settings.

Abstract

This paper explores differentially-private federated learning (FL) across time-varying databases, delving into a nuanced three-way tradeoff involving age, accuracy, and differential privacy (DP). Emphasizing the potential advantages of scheduling, we propose an optimization problem aimed at meeting DP requirements while minimizing the loss difference between the aggregated model and the model obtained without DP constraints. To harness the benefits of scheduling, we introduce an age-dependent upper bound on the loss, leading to the development of an age-aware scheduling design. Simulation results underscore the superior performance of our proposed scheme compared to FL with classic DP, which does not consider scheduling as a design factor. This research contributes insights into the interplay of age, accuracy, and DP in federated learning, with practical implications for scheduling strategies.

Age Aware Scheduling for Differentially-Private Federated Learning

TL;DR

This work studies differential privacy in federated learning with time-varying, Markovian client data and introduces age-aware scheduling to manage the privacy-accuracy trade-off. By deriving an upper bound on the loss difference caused by data aging and modeling data evolution via birth-death Markov chains, the authors formulate an optimization to schedule data collection and adapt noise levels under a global privacy cap. The proposed protocol jointly optimizes scheduling and adaptive DP noise, yielding tighter privacy-utility bounds than non-scheduling baselines and demonstrating superior performance in simulations. The results provide practical scheduling strategies for DP-FL in dynamic databases, with implications for privacy-aware FL deployments in real-time or streaming data settings.

Abstract

This paper explores differentially-private federated learning (FL) across time-varying databases, delving into a nuanced three-way tradeoff involving age, accuracy, and differential privacy (DP). Emphasizing the potential advantages of scheduling, we propose an optimization problem aimed at meeting DP requirements while minimizing the loss difference between the aggregated model and the model obtained without DP constraints. To harness the benefits of scheduling, we introduce an age-dependent upper bound on the loss, leading to the development of an age-aware scheduling design. Simulation results underscore the superior performance of our proposed scheme compared to FL with classic DP, which does not consider scheduling as a design factor. This research contributes insights into the interplay of age, accuracy, and DP in federated learning, with practical implications for scheduling strategies.
Paper Structure (19 sections, 3 theorems, 5 equations, 4 figures)

This paper contains 19 sections, 3 theorems, 5 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Notation
  4. System Model
  5. Federated Learning for Time-Varying Data
  6. Empirical risk minimization
  7. Age-dependent Differential Privacy
  8. Problem Formulation
  9. Main Results
  10. Loss Difference Analysis
  11. Mean Estimation for Birth-and-Death Markov Chain
  12. Privacy Analysis
  13. Accuracy Analysis
  14. Proposed Federated Learning Protocol
  15. Discussion and Simulation
  16. ...and 4 more sections

Key Result

Proposition 1

Consider an $\epsilon_{\textnormal{c}}$-DP mechanism for the $i$-th client, $i\in[m]$. Then, it is also $\epsilon(t,{\epsilon_\textnormal{c}})$-age-dependent DP, where $\epsilon(t,\epsilon_{\textnormal{c}})$ is defined as where $\Delta_i(t)$ is the total variation distance for the $i$-th client, given by and $d_{\textnormal{TV}}(\cdot {,} \cdot)$ is the total variation distance between probabili

Figures (4)

  • Figure 1: Illustration of the time-varying FL procedures for the case of three clients. $t_{\textnormal{c},i}$ denotes the scheduled time to collect data for each client, and $t_\textnormal{agg}$ is the aggregation time by the central server.
  • Figure 2: Accuracy loss difference versus data collection time. In each subfigure, we plot the accuracy loss versus the data collection time of a particular client, while fixing that of the other two clients to be the freshest time. For the simulations, relative DP noises are added to ensure a specific privacy level for each client. The simulation settings will be detailed in Section \ref{['sec:discussion_simulation']}.
  • Figure 3: Loss difference/Noise versus DP requirement.
  • Figure 4: Loss difference versus DP requirement.

Theorems & Definitions (5)

  • Definition 1: Classical $\epsilon$-Differential Privacy ($\epsilon$-DP)
  • Definition 2: Age-Dependent DP ZhangWeiBerryHuang24_1
  • Proposition 1: Mechanism-Dependent Guarantee
  • Theorem 1: Loss Difference Between Global Practical Weight under Privacy Guarantee and Ideal Weight
  • Theorem 2