Edge Cloud Computing and Federated–Split Learning in Internet of Things

The wide deployment of the Internet of Things (IoT) necessitates new machine learning (ML) methods and distributed computing paradigms to enable various ML-based IoT applications to effectively process huge amounts of data. Federated Learning (FL) is a new collaborative learning method that allows multiple data owners to cooperate in ML model training without exposing private data. Split Learning (SL) is an emerging collaborative learning method that splits an ML model into multiple portions that are then trained collaboratively by different entities. FL and SL each have unique advantages and limitations and may complement each other in facilitating effective collaborative learning in an IoT environment. On the other hand, the rapid development of edge cloud computing technologies enables a distributed computing platform in IoT upon which FL and SL frameworks can be deployed. Therefore, the deployment of FL and SL in an edge cloud platform in an IoT environment has become an active area of research, attracting interest from both academia and industry. This Special Issue aims to present the latest research advances in this interdisciplinary field of edge cloud computing and federated–split learning.

This Special Issue includes twelve research articles addressing various aspects of edge cloud computing and federated–split learning, including technologies for improving the performance and efficiency of both FL and SL in edge cloud computing environments, mechanisms for protecting data privacy and enhancing system security in FL and SL frameworks, and ways to exploit FL-/SL-based ML methods and edge cloud computing technologies in order to support various IoT applications.

The constrained computing and communication resources available in IoT constitute the main challenge to high-performance federated and split learning. Therefore, technologies for computation and communication efficiency play a crucial role in the effective deployment of FL/SL frameworks in IoT. In [1], Nikolaidis et al. investigate the resource allocation problem in virtualization-based edge cloud computing systems in order to maximize the efficiency of the FL process. The authors consider factors such as computational and network capabilities, the complexity of datasets, and the specific characteristics of the FL workflow. They explore two scenarios: (i) running FL over a finite number of nodes and (ii) hosting multiple parallel FL workflows on the same set of nodes. The research findings indicate that the default configurations of state-of-the-art cloud orchestrators are sub-optimal when orchestrating FL workflows, demonstrating that different libraries and ML models exhibit diverse computational footprints. Building upon these insights, the authors discuss methods to mitigate computational interference and enhance the overall performance of the FL pipeline’s execution.

Task scheduling is another key technology that has been explored to enhance FL efficiency and performance. In [2], Cai et al. study two problems of task scheduling for FL in edge computing: (1) transmission power allocation (PA) and (2) the dual decision-making problems of joint request offloading and computational resource scheduling (JRORS). The authors propose an adaptive greedy dingo optimization algorithm (AGDOA) based on greedy policies and parameter adaptation in order to solve the PA problem. They also construct a binary salp swarm algorithm (BSSA) that introduces binary coding to solve the discrete JRORS problem. Their simulation results verify that the proposed algorithms improve FL convergence speed, shorten the system response time, and reduce energy consumption.

Future networking technologies such as the 6G network may greatly enhance edge-based IoT; however, they also introduce diverse and heterogeneous devices that present new challenges for FL and SL. In [3], Ridolfi and co-authors analyze FL processes tailored for 6G standards, implementing a practical FL platform that employs Raspberry Pi devices and virtual machines as client nodes and hosts the FL server on a Windows PC. Their analysis delves into the impact of computational resources, data availability, and heating issues across heterogeneous device sets.

The limited bandwidth for data transmission in edge-based IoT makes communication efficiency a critical aspect of deploying FL and SL in this domain. Optimal client selection and model aggregation offer promising approaches to achieving communication-efficient FL. In [4], the authors propose a Federated Learning via Clustering Optimization (FedCO) scheme to optimize model aggregation and reduce communication costs. In FedCo, participating clients are clustered based on the similarity of their model parameters; the best-performing client is selected from each group as a representative to communicate with the central server. The proposed FedCO approach updates clusters by repeatedly evaluating and splitting them, improving worker partitioning. The experimental results indicate that the FedCO approach is effective in reducing communication costs and improving model accuracy.

Client selection and model aggregation technologies have also been leveraged in [5] for reducing model training time while improving model accuracy in FL. The authors of this paper introduce the Latency-awarE Semi-synchronous client Selection and mOdel aggregation for federated learNing (LESSON) method, which allows clients to participate at different frequencies, thus mitigating straggler issues and expediting model convergence. Their simulation results show that LESSON outperforms two baseline methods, FedAvg and FedCS, in terms of convergence speed and maintains a higher model accuracy than FedCS.

Although FL is seen as a privacy-preserving distributed machine learning method, recent research has shown it to be vulnerable to some privacy attacks. Homomorphic Encryption (HE) and Differential Privacy (DP) are two promising techniques for privacy protection in FL. In [6], Aziz et al. first present consistent attacks on privacy in FL and then provide an overview of HE and DP techniques for securing FL in next-generation internet applications. This paper discusses the strengths and weaknesses of these techniques in different settings, as described in the literature, with a particular focus on the trade-off between privacy and convergence, as well as the computation overheads involved.

Blockchain technologies have been employed as an effective approach to improving FL performance in a variety of ways, including protecting data privacy and system security. In [7], Liu et al. examine the EIFFeL framework, a protocol for decentralized real-time messaging in continuous integration and delivery pipelines. The authors introduce an enhanced scheme that leverages the trustworthy nature of blockchain technology. The proposed scheme eliminates the need for a central server and any other third party, thereby mitigating the risks associated with any potential breach.

Combining federated and split learning may aid in fully exploiting the advantages of both while mitigating their respective shortcomings. Thus, this matter has recently become an active research topic attracting extensive interest. In [8], the authors propose a multi-level split–federated learning (multi-level SFL) framework that merges the benefits of both SL and FL. This framework leverages the Message Queuing Telemetry Transport (MQTT) protocol to geographically cluster IoT devices, employing edge and fog computing layers for the initial model parameter aggregation. Their simulation experiments verify that the multi-level SFL framework outperforms traditional SFL by improving the model accuracy and convergence speed in large-scale IoT environments.

The aforementioned research primarily focuses on addressing the challenges of deploying FL/SL frameworks in edge cloud computing-based IoT environments. Another active theme of research is employing FL-/SL-based machine learning techniques together with edge cloud computing technologies to solve various problems in a broad range of IoT applications.

In [9], Zhou and co-authors design an image pre-processing method and propose a lightweight neural network model called LINGE (Lightweight Neural Network Models for the Edge). This paper proposes an FL-based distributed intelligent edge computing technology for disease risk prediction. The proposed scheme performs prediction model training and inference at the edge without increasing storage space, reduces communication load on the network, and releases computing pressure on the server.

The authors of [10] present a edge cloud collaborative banking data open application scenario, focusing on the critical need for an accurate and automated sensitive data classification and categorization method. In this paper, the authors propose a scheme, UP-SDCG, for automatically classifying and grading financial data and develop a financial data hierarchical classification library. The results of their experimental analysis indicate that UP-SDCG achieves a precision of over 95%, outperforming the baseline models.

In [11], the authors propose a dynamic watermarking service framework, E-SAWM, designed for edge cloud scenarios. This framework incorporates dynamic watermark information at the edge, allowing for the precise tracking of leakage throughout the data-sharing process. E-SAWM utilizes semantic analysis to generate highly realistic pseudo statements that ensure resistance to removal or destruction. Their experimental results demonstrate the effectiveness and efficiency of the proposed scheme.

Various FL-based methods have been proposed for security protection in edge-based IoTs, including intrusion detection systems (IDSs) in Internet of Vehicles (IoVs). In [12], Alsamir et al. present a comprehensive review of FL-based IDSs in an IoV environment. Their paper introduces a general taxonomy to describe FL systems in order to ensure a coherent structure and guide future research. Then, the authors identify the relevant state of the art in FL-based IDS technologies in the IoV domain, covering the years from FL’s inception in 2016 through to 2023, discussing challenges and future research directions based on the existing literature.

As Guest Editors, we would like to take this opportunity to thank all the authors who submitted their manuscripts to this Special Issue. Furthermore, we would like to acknowledge all the reviewers whose thorough reviews have helped improve the quality of the manuscripts in this Special Issue. Last but not least, we would like to express our appreciation to the MDPI Editorial Team, who have provided unwavering support throughout this project.