What question did this study set out to answer?

This research aims to develop a federated learning architecture that addresses user privacy while meeting latency demands in edge computing environments.

April 17, 2026Open Access

A Novel Federated Learning Architecture for Preserving User Privacy in Latency-Sensitive Edge Computing Environments

Key Points

This research aims to develop a federated learning architecture that addresses user privacy while meeting latency demands in edge computing environments.
Proposed a non-blocking asynchronous aggregation protocol for federated learning
Implemented local differential privacy using gradient clipping and additive Gaussian noise
Maintained a global momentum buffer to enhance stability under Non-IID distributions
Utilized sparse Top-k compression to optimize communication overhead
Conducted a theoretical analysis to ensure convergence with bounded excess risk.
Achieved sustained convergence without strict synchronization
Reduced communication overhead with the sparse Top-k method
Showed improved stability against stochastic fluctuations
Provided tighter privacy guarantees via Rényi differential privacy
Identified key challenges for future work such as robustness against adversarial attacks.

Abstract

The rapid growth of the Internet of Things (IoT) and the adoption of edge computing have intensified the need for real-time intelligent services that also protect user privacy. Conventional cloud-centric learning pipelines and synchronous federated learning (FL) protocols (e.g., FedAvg) often fail to meet strict latency requirements in heterogeneous networks due to straggler effects and global synchronization delays. This paper proposes a novel FL architecture tailored to latency-sensitive edge environments that combines non-blocking asynchronous aggregation with explicit privacy preservation. The core contribution is an asynchronous streaming aggregation protocol equipped with an adaptive damping function that reduces the adverse impact of stale client updates and sustains convergence without reverting to strict synchronization. To improve stability under Non-IID data distributions, the server maintains a global momentum buffer that smooths stochastic fluctuations across client updates. Communication overhead is further reduced through sparse Top-k compression with error-feedback accumulation, enabling frequent transmission of the most informative gradient components while limiting accuracy degradation. Privacy is ensured via local differential privacy (LDP) using gradient clipping and additive Gaussian noise, while privacy loss accounting is performed with Rényi Differential Privacy (RDP), providing tighter composition bounds and better alignment with continuous asynchronous updates. A theoretical analysis establishes convergence with bounded excess risk for convex objectives, supporting the practical feasibility of high-performance, privacy-preserving Edge AI. The paper also highlights remaining challenges – including non-convex optimization in deep models, robustness against poisoning/Byzantine behaviors, and extreme network instability—and outlines directions for future work.

A Novel Federated Learning Architecture for Preserving User Privacy in Latency-Sensitive Edge Computing Environments

Key Points

Abstract

Cite This Study