New Federated Learning Protocol Tackles Dynamic Edge Computing

Researchers Mahyar Hassanpour and Reza Ramezani have proposed a novel adaptive semi-synchronous federated learning protocol, designated ASSFL, designed to address specific inefficiencies in dynamic edge computing environments. Published in the journal Cluster Computing, the study introduces a semi-decentralized topology and local server grouping mechanisms to improve the performance of distributed machine learning systems. The work aims to resolve critical challenges associated with traditional federated learning architectures, particularly those involving the coordination of multiple edge servers with a central aggregation server.

Federated learning is a computing paradigm developed to enhance privacy in edge computing by utilizing model parameters for communication rather than transmitting raw data. A central challenge in this field is the selection of an appropriate learning protocol. In synchronous learning protocols, each edge server must complete a single learning round within a specified time interval before participating in the aggregation process. This approach ensures that even the slowest server can finish its learning round within the given time, but it introduces the straggler effect, where faster servers must wait for all slower servers to finish, leading to potential inefficiencies.

In contrast, asynchronous learning protocols allow each edge server to communicate with the central server immediately after completing its learning round, with the central server performing aggregation each time it receives new parameters. While this avoids waiting for slower devices, the asynchronous protocol presents its own set of challenges, including an increased number of connections to the central server and the staleness of local models on slower servers. The negative impact of these outdated models on the global model is a significant concern that limits the effectiveness of purely asynchronous approaches in many scenarios.

The semi-synchronous protocol emerged as a hybrid solution intended to harness the advantages of both synchronous and asynchronous protocols. However, this approach faces several distinct challenges, including server idleness, slow convergence rates, local model staleness, and communication overhead. Additionally, dynamic environmental conditions, device mobility, and edge server failures further complicate the implementation of semi-synchronous protocols in real-world edge computing settings. The ASSFL protocol is specifically designed to address all these challenges through its adaptive nature and structural innovations.

The proposed ASSFL protocol incorporates a comprehensive client selection mechanism aimed at reducing communication overhead and accelerating convergence. It considers the geographical distribution of servers and features dynamic grouping capabilities to effectively manage server failures and device relocations. By utilizing a semi-decentralized topology and local server grouping, the protocol seeks to mitigate the straggler effect and the issues of model staleness that plague other federated learning architectures.

Thorough experiments were conducted to evaluate the performance of the ASSFL protocol using the CIFAR10, MNIST, Boston Housing, and KDD CUP ‘99 datasets. The results demonstrate superior performance of the proposed protocol in terms of global model accuracy and local model approximation compared to existing methods.

Subject of Research: Adaptive semi-synchronous federated learning protocol design for dynamic edge computing environments

Article Title: Adaptive semi-synchronous federated learning with semi-decentralized topology in dynamic environments

Article References: Hassanpour, M., & Ramezani, R. (2026). Adaptive semi-synchronous federated learning with semi-decentralized topology in dynamic environments. Cluster Computing, 29(16), Article 907. https://doi.org/10.1007/s10586-026-06738-x

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06738-x

Keywords: federated learning, edge computing, semi-synchronous protocol, ASSFL, distributed machine learning, cluster computing, privacy preservation, model aggregation, dynamic topology, client selection, Adaptive, semi-synchronous