Coding-Enforced Robust Secure Aggregation for Federated Learning Under Unreliable Communication

Shudi Weng, Chao Ren, Yizhou Zhao, Ming Xiao, Mikael Skoglund

Abstract

This work studies privacy-preserving federated learning (ppFL) under unreliable communication. In ppFL, zero-sum privacy noises enables privacy protection without sacrificing model accuracy, effectively overcoming the privacy-utility trade-off. However, in practice, unreliable communication can randomly disrupt the coordination of zero-sum noises, leading to aggregation errors and unpredictable partial participation, which severely harm the model accuracy and learning performance. To overcome these challenges, we propose a robust coding-enforced structured secure aggregation method, termed secure cooperative gradient coding (SecCoGC), which enables exact reconstruction of the global model under unreliable communication while allowing for arbitrarily strong privacy preservation. In this paper, a complete problem formulation and constructions of real-field zero-sum privacy noise are presented, and fairness is introduced as a privacy metric. Privacy across all protocol layers in SecCoGC is evaluated, accounting for the correlation among privacy noises and their linear combination under unreliable communication. Moreover, a distinct convergence analysis for the FL algorithm with a binary outcome for global model recovery is provided. Experimental results demonstrate that SecCoGC achieves strong resilience to unreliable communication while maintaining varying levels of privacy preservation, yielding test accuracy improvements of up to 20%-70% over existing benchmark methods.

Introduction

Federated Learning (FL) allows multiple clients to collaboratively train a machine learning model without sharing their raw data. However, the model updates themselves can leak sensitive information. To mitigate this, researchers often use privacy-preserving FL (ppFL), typically by injecting random noise into updates.

A powerful technique in ppFL is the use of zero-sum noise. When clients add noises that sum to zero, the server recovers a clean, accurate global model upon aggregation. However, real-world networks are often unreliable. If a client's update is lost, the corresponding zero-sum noise remains in the aggregate as an error, which can ruin model accuracy and cause training to fail. This paper introduces SecCoGC (Secure Cooperative Gradient Coding) to maintain perfect noise cancellation and model accuracy even in unpredictable communication environments.

Methods

SecCoGC combines coding theory with privacy-preserving techniques to ensure robustness. The approach relies on three key pillars:

  1. Cyclic Gradient Coding (GC): Clients act as relays, sharing masked updates with their neighbors. They transmit encoded "partial sums" to the server. The coding structure is designed so that the server can reconstruct the exact global update as long as a minimum threshold of packets is received.
  2. Real-Field Zero-Sum Noise: Unlike previous methods limited to discrete noise, this work provides a mathematical construction for zero-sum noise in the real-number field, making it practical for standard machine learning gradients.
  3. Binary Recovery Mechanism: The system is designed with an "all-or-nothing" outcome. The server either receives enough data to reconstruct the exact global model (canceling all noise) or receives nothing meaningful, preventing the "client drift" caused by biased partial updates.
SecCoGC architecture showing cooperative gradient coding and zero-sum noise cancellation.

As illustrated in Figure 1, the cooperative nature of SecCoGC allows clients to share information (indicated by dotted lines) to create redundant paths for the global model reconstruction. This ensures that the sum of gradients $\Theta_i$ can be recovered even if some transmissions fail, while the privacy noises $N_i$ are structured to cancel out perfectly.

Experimental Results

The researchers evaluated SecCoGC against standard private FL and Diversity Network Codes using the MNIST and CINIC-10 datasets. They simulated both symmetric and asymmetric network environments to test resilience against varying levels of unreliability.

Key Findings

  • Significant Accuracy Improvements: In highly unreliable networks, SecCoGC improved test accuracy by 20% to 70% compared to existing private FL methods.
  • Utility Preservation: SecCoGC consistently approached the performance of "Ideal FL" (perfect communication with no privacy), proving that high security does not require sacrificing model utility.
  • Robustness to Stragglers: The framework successfully mitigated the impact of slow or dropping clients, maintaining stable convergence where other methods fluctuated or diverged.

Insights and Future Work

SecCoGC introduces Fairness as a privacy metric, ensuring that every client receives an identical level of privacy protection (noise power). By using coding as a privacy shield, the system prevents the server from ever seeing a partial sum that might leak individual client data.

While promising, the method introduces computational overhead for clients who must encode, decode, and relay data. Future research will focus on adapting the system for heterogeneous learning weights and more complex, non-cyclic network topologies.

Summary updated

Abstract

This work studies privacy-preserving federated learning (ppFL) under unreliable communication. In ppFL, zero-sum privacy noises enables privacy protection without sacrificing model accuracy, effectively overcoming the privacy-utility trade-off. However, in practice, unreliable communication can randomly disrupt the coordination of zero-sum noises, leading to aggregation errors and unpredictable partial participation, which severely harm the model accuracy and learning performance. To overcome these challenges, we propose a robust coding-enforced structured secure aggregation method, termed secure cooperative gradient coding (SecCoGC), which enables exact reconstruction of the global model under unreliable communication while allowing for arbitrarily strong privacy preservation. In this paper, a complete problem formulation and constructions of real-field zero-sum privacy noise are presented, and fairness is introduced as a privacy metric. Privacy across all protocol layers in SecCoGC is evaluated, accounting for the correlation among privacy noises and their linear combination under unreliable communication. Moreover, a distinct convergence analysis for the FL algorithm with a binary outcome for global model recovery is provided. Experimental results demonstrate that SecCoGC achieves strong resilience to unreliable communication while maintaining varying levels of privacy preservation, yielding test accuracy improvements of up to 20%-70% over existing benchmark methods.
BibTeX
@misc{weng2025coding,
  author = {Shudi Weng and Chao Ren and Yizhou Zhao and Ming Xiao and Mikael Skoglund},
  title = {{Coding-Enforced Robust Secure Aggregation for Federated Learning Under Unreliable Communication}},
  year = {2025},
  eprint = {2507.07565},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.07565},
}
Figure