FedFog: Resource-Aware Federated Learning in Edge and Fog Networks

Somayeh Sobati-M · 2025

Read FedFog: Resource-Aware Federated Learning in Edge and Fog Networks by Somayeh Sobati-M (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-03952.txt.

Introduction

Somayeh Sobati-M introduces FedFog, a simulation framework for learning across edge and fog networks without centralizing data. Edge and fog computing now anchor many distributed systems. Interest is growing in pairing designs that run code on demand with learning that keeps data local. Existing simulators capture that pairing poorly. FedFog extends the FogFaaS environment to support such learning across edge-fog setups.

An adaptive scheduler, privacy-respecting data flow, and resource-aware control recreate shifting, realistic conditions for Internet of Things settings. Simulations on standard datasets showed faster model convergence, lower delay, and better energy use than standard setups. The author offers the tool to researchers building edge systems that must scale.

Catalog record

CREATORSomayeh Sobati-M
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-03952

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

FedFog: Resource-Aware Federated Learning in Edge and Fog Networks Somayeh Sobati-M Abstract As edge and fog computing become central to modern distributed systems, there’s growing interest in combining serverless architectures with privacy-preserving machine learning techniques like federated learning (FL). However, current simulation tools fail to capture this integration effectively. In this paper, we introduce FedFog, a simulation framework that extends the FogFaaS environment to support FL-aware serverless execution across edge–fog infrastructures. FedFog incorporates an adaptive FL scheduler, privacy-respecting data flow, and resource-aware orchestration to emulate realistic, dynamic conditions in IoT-driven scenarios. Through extensive simulations on benchmark datasets, we demonstrate that FedFog accelerates model convergence, reduces latency, and improves energy efficiency compared to conventional FL or FaaS setups—making it a valuable tool for researchers exploring scalable, intelligent edge systems. I Introduction The rapid growth of Internet of Things (IoT) applications has led to an increased demand for computing paradigms that process data near its source. Edge and fog computing offer promising solutions by distributing computation closer to end devices, reducing both latency and bandwidth usage [ 1 ] . Simulation frameworks such as iFogSim [ 2 ] and iFogSim2 [ 3 ] have proven valuable for evaluating resource management and scheduling in these distributed environments. However, these frameworks do not support serverless computing models, which are increasingly prevalent in cloud-native and event-driven architectures. Serverless paradigms, or Function-as-a-Service (FaaS), offer dynamic scalability and are well-suited for edge computing scenarios that involve sensitive or distributed data. While FL libraries such as Flower [ 4 ] enable flexible experimentation with learning algorithms, they do not simulate the underlying constraints of edge environments—such as network unreliability, limited compute, or energy-aware scheduling [ 5 ] . Moreover, existing tools lack orchestration support for FL in serverless environments. This omission is critical, as real-world FL deployments must address client heterogeneity, data drift, and dynamic participation, all while operating under resource constraints. This divergence between theoretical algorithmic capabilities and practical implementation requirements reveals a critical research void that must be addressed. In operational environments, FL systems encounter four interconnected challenges that collectively complicate deployment. First, the inherently dynamic nature of client participation stems from device mobility patterns and finite energy budgets, creating unpredictable availability windows. Second, the statistical distribution of data varies significantly across edge nodes, violating the independent and identically distributed (IID) assumptions common to many machine learning approaches. Third, the serverless execution model introduces unpredictable cold start latencies during function initialization, which can disrupt training timelines. Finally, practitioners must navigate fundamental tensions between model accuracy objectives and system efficiency metrics, requiring careful optimization across multiple competing dimensions. These operational realities remain largely unaddressed by current simulation tools, limiting their utility for real-world deployment planning. In this paper, we introduce FedFog —a modular simulation framework that integrates FL with serverless execution atop the iFogSim environment. FedFog supports dynamic function orchestration, resource-aware scheduling, node health monitoring, and data drift simulation. It enables detailed evaluation of FL systems under realistic edge conditions, including client dropout, cold starts, and non-IID data. Our work is motivated by the practical challenges faced when modeling FL in constrained and heterogeneous environments. FedFog aims to fill this simulation gap and serve both as a research tool and a testbed for robust and efficient FL algorithm design. Our key contributions are: • Design and implementation of FedFog, a serverless FL simulator built on iFogSim. • Support for cold start simulation, client heterogeneity, energy constraints, and data drift. • Comprehensive evaluation against FaaS and centralized learning baselines across multiple metrics (accuracy, latency, energy). The remainder of this paper is organized as follows. Section II surveys the most relevant simulation tools and federated learning frameworks, identifying the limitations that motivate our work. Section III introduces the architecture and design principles of the proposed FedFog framework, including its adaptive scheduling and cold start modeling. Section IV presents the experimental setup, evaluation metrics, and comparative results under various conditions, including adversarial scenarios and real-world deployment. Section V provides a theoretical analysis of FedFog’s complexity and trade-offs, supported by sensitivity and scalability experiments. Finally, Section VI concludes the paper and outlines future directions, including enhanced privacy mechanisms and trust-aware scheduling. II Related Work This section systematically reviews prior research in three key areas foundational to our work. First, we analyze iFogSim and iFogSim2 as seminal simulators for fog/edge environments, focusing on their resource modeling capabilities and limitations in handling machine learning workloads. Second, we examine advancements in Serverless Computing , particularly FogFaaS extensions, that enable efficient function orchestration across edge-fog hierarchies. Finally, we survey FL Frameworks , emphasizing their adaptation to distributed infrastructure constraints and privacy requirements. By synthesizing these strands of research, we identify critical gaps—especially in simulating FL-serverless integration—that motivate our FedFog framework. Serverless computing, or Function-as-a-Service (FaaS), has emerged as a paradigm for scalable and event-driven applications. OpenFaaS [ 6 ] is an open-source FaaS platform that enables developers to deploy functions over Docker and Kubernetes. While effective in production environments, OpenFaaS lacks simulation capabilities and is unsuitable for evaluating large-scale edge scenarios under controlled conditions. The iFogSim toolkit [ 2 ] is a widely used simulation framework for modeling and evaluating resource management strategies in edge and fog computing environments. It supports latency-aware resource allocation, hierarchical topologies, and basic mobility. However, iFogSim lacks built-in support for serverless computing paradigms and machine learning workflows. To address some of these gaps, iFogSim2 [ 3 ] was introduced, which enhances mobility modeling and introduces microservice management and clustering capabilities. Despite these improvements, neither framework supports serverless computing for AI-driven edge applications. To fill this gap, Ghaseminya et al. proposed FogFaaS [ 7 ] , an extension of iFogSim that simulates serverless computing in fog environments. FogFaaS introduces components to model cold starts, dynamic resource scaling, and function invocation behavior [ 8 ] . However, it does not support AI workloads, especially FL, which is increasingly relevant in edge scenarios where privacy and data locality are essential. FL allows decentralized model training across multiple devices while preserving data privacy [ 9 ] . The foundational work by McMahan et al. [ 10 ] introduced the Federated Averaging (FedAvg) algorithm, which has become the basis for many FL implementations. Flower [ 4 ] is a flexible FL framework that supports different ML backends and training configurations. Despite its flexibility, Flower does not offer simulation of edge resource constraints or integration with FaaS execution models. II-A Gap and Motivation While several simulation tools support edge/fog environments and others focus on FL experimentation, none of the existing platforms provide a unified simulation of FL-aware serverless computing (See Table I ). There is a critical need for a framework that combines the dynamic scalability of FaaS with the intelligence and privacy-preserving nature of FL in resource-constrained edge environments. This motivates the development of FedFog , a federated-learning-aware serverless simulation framework that extends FogFaaS by integrating FL task scheduling, model aggregation, and resource-aware orchestration within iFogSim. TABLE I : Comparison of Existing Frameworks with the Proposed FedFog Simulator Feature / Framework iFogSim [ 2 ] FogFaaS [ 7 ] OpenFaaS [ 6 ] Flower [ 4 ] iFogSim2 [ 3 ] FedFog (Proposed) FaaS Support ✗ ✓ ✓ ✗ ✗ ✓ Simulation Capability ✓ ✓

Read the complete text