Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Coding-Enforced Resilient and Secure Aggregation for Hierarchical Federated Learning

Shudi Weng, Ming Xiao, Mikael Skoglund

TL;DR

The paper tackles robust secure aggregation in hierarchical federated learning under unreliable communications and strong privacy constraints. It introduces H-SecCoGC, a coding-enforced framework that combines gradient-sharing with $(K,s)$-cyclic gradient codes to enable exact global model recovery while canceling privacy noises during aggregation. The authors provide a worst-case information-leakage analysis with optimal Gaussian secret keys and deliver a full local differential privacy (LDP) treatment across client, relay, and server layers. Simulations on CIFAR-10 with Dirichlet-distributed data show that H-SecCoGC robustly maintains accuracy under outages, approaching the ideal relay-connected performance and outperforming noncoding baselines.

Abstract

Hierarchical federated learning (HFL) has emerged as an effective paradigm to enhance link quality between clients and the server. However, ensuring model accuracy while preserving privacy under unreliable communication remains a key challenge in HFL, as the coordination among privacy noise can be randomly disrupted. To address this limitation, we propose a robust hierarchical secure aggregation scheme, termed H-SecCoGC, which integrates coding strategies to enforce structured aggregation. The proposed scheme not only ensures accurate global model construction under varying levels of privacy, but also avoids the partial participation issue, thereby significantly improving robustness, privacy preservation, and learning efficiency. Both theoretical analyses and experimental results demonstrate the superiority of our scheme under unreliable communication across arbitrarily strong privacy guarantees

Coding-Enforced Resilient and Secure Aggregation for Hierarchical Federated Learning

TL;DR

The paper tackles robust secure aggregation in hierarchical federated learning under unreliable communications and strong privacy constraints. It introduces H-SecCoGC, a coding-enforced framework that combines gradient-sharing with -cyclic gradient codes to enable exact global model recovery while canceling privacy noises during aggregation. The authors provide a worst-case information-leakage analysis with optimal Gaussian secret keys and deliver a full local differential privacy (LDP) treatment across client, relay, and server layers. Simulations on CIFAR-10 with Dirichlet-distributed data show that H-SecCoGC robustly maintains accuracy under outages, approaching the ideal relay-connected performance and outperforming noncoding baselines.

Abstract

Hierarchical federated learning (HFL) has emerged as an effective paradigm to enhance link quality between clients and the server. However, ensuring model accuracy while preserving privacy under unreliable communication remains a key challenge in HFL, as the coordination among privacy noise can be randomly disrupted. To address this limitation, we propose a robust hierarchical secure aggregation scheme, termed H-SecCoGC, which integrates coding strategies to enforce structured aggregation. The proposed scheme not only ensures accurate global model construction under varying levels of privacy, but also avoids the partial participation issue, thereby significantly improving robustness, privacy preservation, and learning efficiency. Both theoretical analyses and experimental results demonstrate the superiority of our scheme under unreliable communication across arbitrarily strong privacy guarantees
Paper Structure (18 sections, 7 theorems, 21 equations, 3 figures)

This paper contains 18 sections, 7 theorems, 21 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Coding Strategies in Straggler Mitigation
  3. Privacy-Preserving HFL/FL
  4. System Model and Preliminaries
  5. Hierarchical Federated Learning
  6. Communication Model
  7. $(K,s)$-Cyclic Gradient Codes
  8. The Proposed Method
  9. Description
  10. Optimal Secret Keys
  11. Privacy Analysis
  12. Client-level LDP
  13. Relay-level LDP
  14. Server-level LDP
  15. Simulations
  16. ...and 3 more sections

Key Result

Lemma 3.1

Any other distribution of the local model updates than the Gaussian vector with a diagonal covariance matrix $\zeta^2\boldsymbol{I}_D$ results in lower privacy leakage.

Figures (3)

  • Figure 1: The proposed robust secure aggregation based on hierarchical CoGC with $K=3$$s=1$ under unreliable communication condition.
  • Figure 2: Test accuracy comparison of H-SecCoGC with benchmark methods under varying levels of privacy noises over symmetric networks.
  • Figure 3: Test accuracy comparison of H-SecCoGC with benchmark methods under varying levels of privacy noises over unsymmetric networks.

Theorems & Definitions (13)

  • Remark 3.1: Communication Costs
  • Remark 3.2: Decoding Complexity
  • Lemma 3.1: Worst Input
  • Lemma 3.2: Optimal Keys
  • Lemma 4.1
  • Theorem 4.1
  • proof : Proof Sketch
  • Theorem 4.2
  • proof : Proof Sketch
  • Theorem 4.3
  • ...and 3 more