Multicenter Hierarchical Federated Learning With Fault-Tolerance Mechanisms for Resilient Edge Computing Networks

In the realm of federated learning (FL), the conventional dual-layered architecture, comprising a central parameter server and peripheral devices, often encounters challenges due to its significant reliance on the central server for communication and security. This dependence becomes particularly problematic in scenarios involving potential malfunctions of devices and servers. While existing device-edge-cloud hierarchical FL (HFL) models alleviate some dependence on central servers and reduce communication overheads, they primarily focus on load balancing within edge computing networks and fall short of achieving complete decentralization and edge-centric model aggregation. Addressing these limitations, we introduce the multicenter HFL (MCHFL) framework. This innovative framework replaces the traditional single central server architecture with a distributed network of robust global aggregation centers located at the edge, inherently enhancing fault tolerance crucial for maintaining operational integrity amidst edge network disruptions. Our comprehensive experiments with the MNIST, FashionMNIST, and CIFAR-10 datasets demonstrate the MCHFL’s superior performance. Notably, even under high paralysis ratios of up to 50%, the MCHFL maintains high accuracy levels, with maximum accuracy reductions of only 2.60%, 5.12%, and 16.73% on these datasets, respectively. This performance significantly surpasses the notable accuracy declines observed in traditional single-center models under similar conditions. To the best of our knowledge, the MCHFL is the first edge multicenter FL framework with theoretical underpinnings. Our extensive experimental results across various datasets validate the MCHFL’s effectiveness, showcasing its higher accuracy, faster convergence speed, and stronger robustness compared to single-center models, thereby establishing it as a pioneering paradigm in edge multicenter FL.

I. Introduction

Traditional machine learning frameworks, predominantly reliant on cloud-based data centers, necessitate the transmission of annotated training datasets to centralized servers for processing and analytical computations [1], [2]. However, this centralization paradigm is impeded by limitations in network bandwidth and substantial geographical distances between clients and cloud infrastructures [3]. Such constraints are particularly disadvantageous for burgeoning real-time applications, including autonomous vehicle navigation and live video broadcasting [4]. In the wake of advancements in 5G communication technologies, edge computing has surfaced as an efficacious alternative and adjunct to traditional cloud-based approaches [5], [6]. This paradigm shift leverages the computational proficiency and data storage capabilities of mobile edge computing (MEC) servers, effectively bridging the chasm between computational models and data origination points [7]. While MEC servers, in close proximity to end-users, can expedite the data aggregation process to satisfy real-time processing requisites, the delegation of computational tasks and data management to these servers still necessitates the transfer of potentially sensitive personal information [8]. This raises substantial privacy concerns among users involved in model training activities and might be in conflict with the progressively stringent norms of privacy legislation [9], [10].