SDFLMQ: A Semi-Decentralized Federated Learning Framework over MQTT

TL;DR

The paper addresses scalability bottlenecks in Federated Learning for IoT edge environments by proposing SDFLMQ, a semi-decentralized FL framework that runs over MQTT. It leverages topic-based role management, hierarchical clustering, and dynamic aggregation placement to distribute the aggregation load across edge devices while avoiding a single central server. The framework architecture comprises a coordinator, client logic, and a parameter server, with MQTT-based broker bridging to regionalize clusters. Preliminary runtime evaluations show that SDFLMQ can achieve performance close to centralized approaches and offers practical edge deployment benefits, highlighting its potential for large-scale, heterogeneous IoT deployments.

Abstract

Federated Learning is widely discussed as a distributed machine learning concept with stress on preserving data privacy. Various structures of Federated Learning were proposed. Centralized Federated learning for instance has been the primary structure that suits cloud computing. Decentralized Federated learning also has been proposed for ecosystems where communication is dominantly peer-to-peer. Semi-Decentralized Federated Learning (SDFL) has emerged recently as a new concept where the interconnected nodes are clustered, and each cluster is managed independently. The potential of SDFL lies in its clustering feature, which distributes the load of the global model update down onto multiple nodes. Since the concept is fairly new, much can be done to render this FL model a reliable, efficient, and real-time service at the edge. In this paper, we propose SDFLMQ, a semi-decentralized Federated learning framework at the Edge that uses MQTT as the communication protocol. We demonstrate how the publish/subscribe communication model is used to facilitate the clustering and load balancing in SDFL. We also demonstrate how SDFLMQ can use some of the core MQTT features to expand its capacity with no significant costs. Based on some primary evaluations, we demonstrate how SDFLMQ can efficiently distribute the load of aggregation, and potentially save unnecessary memory allocation, all with no requirement for a powerful central unit for aggregation and global model update. We also disclose some of the key future expansions of SDFLMQ with a focus on the operation of large deep neural network models at the edge.

Figures (8)

• Figure 1: Federated Learning Topologies
• Figure 2: overview of semi-decentralized federated learning (SDFL) over MQTT communication with broker bridging enabled.
• Figure 3: SDFLMQ framework architecture. Both the aggregation and training services are included in the client model. However, based on the role arbiter, either one, or both of them can run and contribute to the FL round. The clustering engine and Task manager services on the coordinator model communicate with the role arbiter on the client system to choose their aggregation and training roles. The aggregation pipeline running on the root node will store the updated model on the parameter server. The global update synchronizer on the parameter server updates the local model in each client after an FL round.
• Figure 4: Illustration showing a) interaction between client and coordinator for a new session establishment for up to 8 contributors, and b) interaction between a new client and coordinator for joining a session.
• Figure 5: Illustration showing a) interaction between clients and the coordinator, where the coordinator is publishing clustering topology on session topic to all clients b) A clustered network of clients, where A/T5 is the head of cluster 1, A/T4 is the head of cluster 2, and A/T6 is the head of cluster 3.