1
Fundamentals of Edge AI and Federated Learning

Atefeh Hemmati1, Hanieh Mohammadi Arzanagh2, and Amir Masoud Rahmani3

1Department of Computer Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran

2Department of Computer Science and Engineering, Shahid Beheshti University, Tehran, Iran

3Future Technology Research Center, National Yunlin University of Science and Technology, Douliou, Taiwan

1.1 Introduction

Machine Learning (ML) and Artificial Intelligence (AI) have advanced unparalleled in recent years. These developments have changed many industries, and interactions between people and technology have experienced significant changes. Data processing in cloud environments was a primary focus of traditional AI and ML techniques. The need for real-time processing has expanded due to the growth of the Internet of Things (IoT). Because of this, a new era of technology known as Edge AI and Federated Learning (FL) has appeared [1, 2].

Edge AI gives edge devices and servers the ability to assess data and carry out AI operations close to the data generated. Autonomous vehicles, remote monitoring, and customized user experiences are some of the applications of this technology that stand out. Because data is not centralized and is handled on local devices, edge AI can help with security and privacy issues [3–5].

At the same time, FL has become a crucial strategy for protecting data privacy and distributing the work of developing ML models among several devices. Devices train ML models using local data using this technique, sharing only the model’s parameters with a centralized server. This ensures that consumers’ privacy is protected and that sensitive data remains on the devices [2, 4].

With the assistance of edge AI and FL, we can create applications and systems that are smarter and safer. This is a new claim in the field of AI. With this novel approach, we can utilize the sophisticated capabilities of AI in a connected world while enhancing privacy, security, and efficiency [2, 6, 7].

The concepts and foundations of Edge AI and FL are entirely clarified in this chapter. This chapter also examines the advantages and challenges of Edge AI and FL. We explore the benefits, which include decreased latency, better bandwidth consumption, improved privacy, and increased robustness. We also present the challenges, such as resource limitations on edge devices, communication costs, and the requirement to manage uniformly distributed data distributions in FL. We also discuss real-world applications where Edge AI and FL have proven effective, opening the door for revolutionary solutions across numerous industries. We also discuss FL and Edge AI’s challenges, future research directions, and open issues.

The remaining part of the chapter is organized as follows: In Section 1.2, the fundamental concepts of Edge AI are discussed, including the advantages of Edge AI and the challenges faced by Edge AI. Section 1.3 introduces the concepts and fundamentals of FL, including the advantages of FL and the challenges associated with FL. In Section 1.4, the chapter emphasizes the combined power of FL and Edge AI, highlighting the benefits of their integration. Section 1.5 offers background insights into the technological landscape and motivations behind exploring Edge AI and FL. Applications of Edge AI and FL are discussed in Section 1.6. Section 1.7 addresses challenges, future research directions, and potential solutions for integrating Edge AI and FL. Finally, Section 1.8 concludes the chapter.

1.2 Concepts and Fundamentals of Edge AI

In an era of data generation and IoT, Edge AI is emerging as a revolutionary paradigm that brings AI computing closer to the edge of the network, where data is generated. Unlike traditional AI architectures that rely on cloud servers for processing, Edge AI distributes computing tasks to local edge devices such as smartphones, smart sensors, and edge servers. This change in the deployment of AI brings significant benefits and exciting possibilities in various fields [3, 8].

1.2.1 Defining Edge AI

Figure 1.1 shows that Edge AI directly integrates AI and ML capabilities on edge devices. This allows devices to perform data processing and inference locally, reducing the need to transfer data to centralized cloud servers constantly. This decentralized operation reduces latency in terms of time and avoids dependence on a stable Internet connection [9, 10].

Figure 1.1 Edge AI architecture.

Edge devices play a vital role in the edge AI ecosystem. These devices are typically equipped with limited resources compared to powerful cloud servers but with sufficient computing capabilities to run lightweight AI models. In other words, these devices can perform AI operations with their limited resources [11]. There are many examples of edge devices, including smartphones, tablets, smart cameras, wearables, unmanned aerial vehicles (UAVs), and IoT sensors [12, 13]. These devices can collect, process, and interpret data without frequent communication with central servers and are ideally suited for situations where network bandwidth is limited, or data privacy is required. With these edge devices, the possibility of analyzing and using data locally increases to a certain extent, which leads to reduced response delays and improved user experience [5, 14, 15].

Edge AI has improved the process of directly deploying AI models to edge devices. This type of deployment results in significant latency reduction and increased privacy due to its local processing capabilities. Several different architectures have been used to implement edge AI, each of which covers its own specific needs and applications [16]:

• Local Processing in IoT Devices: IoT devices mostly have limited computing resources, which makes direct deployment of complex AI models challenging. This type of architecture is suitable for applications that require fast processing, such as sensor data analysis, anomaly detection, and simple classification tasks [1, 17, 18]. In this instance, a unique local processing technique suggested by Bebortta et al. [19] supports an enhanced IoT platform structure for smart buildings. The proposed approach helps decrease bandwidth at the nodes’ data collection level from a green computing standpoint. The researchers also discussed the most effective usage of sensors in wireless sensor networks (WSNs), who used the well-known queue model to calculate costs associated with non-Poisson and Poisson arrival of data packets at local processors. The experiments show that the suggested model successfully utilizes green computing standards. Therefore, this study offers a thing-centric, data-centric, and service-oriented IoT architecture within the framework.
• Edge Servers and Gateways: Edge servers and gateways are more powerful computing devices closer to the edge devices. They act as intermediaries between edge devices and cloud servers, performing initial AI processing before sending data to cloud servers for further analysis. These devices can accommodate more resource-intensive models and are suitable for applications such as video analytics, natural language processing (NLP), and data preprocessing [2, 20]. As a real-world example, Rahmani et al. [21] utilized the ideal position of these gateways at the network’s edge to provide several more advanced solutions, including storage locally, real-time local data processing, and integrated data mining, for giving an intelligent electronic health care gateway in the procedure. They then suggested creating a geo-distributed intermediate layer of awareness between sensor nodes and the cloud to use fog computing for medical IoT devices. Their fog-assisted design might handle obstacles in omnipresent medical facilities, such as mobility, energy consumption, flexibility, and dependability issues, by focusing on part of the duties of the sensor network and a distant medical center. Also, we can mention that Li et al. [22] suggested edge content-centric networking (ECCN). This enabling strategy combined Software-Defined Networking (SDN) with content-centric networking into a structured framework. To separate the data and control planes of ECC and CCN, SDN technology was included in the hierarchical framework. To manage data transmission, an SDN framework was created. To assess the effectiveness of the ECCN system, two apps were also deployed in the testbed. Thorough computations and results of experiments from the experimentation applications show that ECCN surpassed the original structures.
• Edge Computing: The concept of edge computing expands the conceptual scope of edge computing by introducing a hierarchical architecture that spans from the edge to the cloud. In this setup, middle fog nodes are responsible for AI processing and decision-making, minimizing latency and network traffic. These nodes may be located in cellular base stations, access points, and local data centers. Edge computing is particularly suitable for applications that require real-time analytics in distributed environments such as smart cities and industrial IoT [23]. Chen et al. [24] addressed a distributed computing model called Mobile Edge Computing (MEC), which...