The architecture is composed of three fundamental layers: 

1. 1. Central Cloud Server
2. 2. Edge Servers
3. 3. Edge Devices

which together form the backbone of an edge computing-based federated learning system.

• 1. Edge Devices: these are portable, user-facing devices such as smartphones, smartwatches, and tablets located at the network’s edge. Due to their compact and mobile nature, they typically have limited processing and storage capabilities, making them unsuitable for heavy computational tasks. However, they continuously generate and store large amounts of user data—often sensitive and privacy-related.
• 2. Edge Servers: Positioned close to edge devices, these servers benefit from short communication distances and high bandwidth, ensuring fast and efficient data transfer. Compared to edge devices, edge servers have greater computational power, more storage, and stable energy supply. They are well-suited to assist edge devices by offloading and executing more demanding computing tasks.
• 3. Cloud Server: Located at remote data centers, cloud servers are geographically distant from edge components, resulting in longer communication times and limited bandwidth. Despite this, they offer extensive computing and storage resources, making them ideal for performing large-scale processing operations.

Unlike traditional federated learning, Edge Federated Learning (EFL) introduces an intermediate layer for distributed training. In EFL:

• – The cloud server first sends an initial global model to the edge servers. 
• – Edge devices then connect to their respective edge servers to download the model parameters. 
• – After local training, edge devices send their model updates to the edge server.
• – Edge servers aggregate the models. This intermediate aggregation reduces communication overhead.
• – Once local updates are collected from all edge servers, a final global aggregation is performed by the cloud server. When the model reaches a target accuracy, the updated global parameters are distributed back to edge devices through the edge servers.
• 
  

Scale of Federation

Just like traditional federated learning, edge federated learning can also be divided into two main categories based on the scale of participation: cross-silo and cross-device edge federated learning.

Cross-Device Federated Learning with Cloud-Edge Collaboration

Cross-device federated learning can be significantly enhanced when integrated with a cloud-edge collaborative framework.

Unlike traditional cross-device FL, this architecture leverages edge servers to offer additional protection for local network traffic. If an edge server is compromised, it can be removed from the training process without affecting the rest of the system, thus isolating attacks and preserving the integrity of other devices’ training processes.

Beyond improved security, edge servers can also safely handle computational tasks offloaded from client devices, providing low-latency processing capabilities. This is especially beneficial for mobile devices, which may have limited processing power but maintain strong communication links with nearby edge servers, reducing their computational load during model training.

In cross-device edge federated learning, the number of participating training nodes can reach into the millions, with each node typically possessing only a small amount of data and limited computational capability. These nodes are generally portable devices or sensors, such as smartphones, smartwatches, or IoT-enabled equipment.

Cross-Silo Federated Learning with Cloud-Edge Collaboration

In the cloud-edge collaborative model, cross-silo FL gains additional flexibility and efficiency. In this context, each silo’s dataset is better treated as an individual learning task rather than just part of a fragmented global dataset.

One of the key challenges in cross-silo FL is the Non-IID (non-identically distributed) nature of data across different clients. Traditionally, meta-learning and transfer learning techniques are used within centralized cloud environments to address this. However, cloud-edge collaboration introduces an alternative: Hierarchical Federated Learning through clustering.

This method involves grouping clients based on similarities in their data distribution, with each group managed by a dedicated edge server. Studies have validated that this cluster-based approach is effective in mitigating Non-IID challenges, leading to better model performance and more efficient federated training.

In cross-silo edge federated learning, the number of participating nodes is relatively small, but each node, typically an edge server, must be equipped with ample computational resources to handle large-scale datasets. For instance, major e-commerce platforms may use this setup to recommend products by training on tens of millions of shopping records distributed across geographically separated data centers.

Challenges of Edge Federated Learning

Security and privacy are two of the most critical challenges in implementing federated learning (FL) within edge computing environments.

Due to the inherently heterogeneous and decentralized nature of both technologies, predicting the behavior of participating nodes becomes difficult. During training, some edge nodes may act maliciously and attempt to compromise the learning process.

Moreover, curious or semi-honest nodes and servers may try to infer private information from the uploaded model updates, potentially compromising user privacy.

Security Challenges in Edge Federated Learning

Security issues in Edge Federated Learning (EFL) stem largely from the lack of trust among the various components of the system. 

Edge nodes and servers often come from different sources, and not all of them can be trusted. While federated learning enhances user privacy by keeping data local, this same feature makes it more vulnerable to malicious behaviors.

Byzantine Attacks

These are especially problematic in distributed learning. In the edge computing context, their impact is amplified due to:

• – Device Heterogeneity: Edge devices and servers differ in origin and trust level.
• – Non-i.i.d. Data: The private data on each node varies significantly, making anomalies harder to detect.
• – Partial Participation: Not all nodes contribute to every training round, so assuming an honest majority is unreliable malicious nodes may dominate in any given round.

Common Byzantine attack methods include:

• – Gaussian Attack: Injecting random noise (e.g., Gaussian noise) into model gradients or weights.
• – Omniscient Attack: Malicious nodes align their updates with the direction of the aggregate of honest nodes, misleading the global model.
• – Flip Bit Attack: Altering bits in gradient or weight values to introduce subtle yet damaging changes.

Poisoning Attacks

Another major security threat in EFL is poisoning attacks, which aim to disrupt or compromise the model by manipulating inputs or local models. They are classified into:

• – Data Poisoning: The attacker manipulates the training data to influence the model’s behavior.
• – Model Poisoning: Instead of altering the data, the attacker tampers with the local model, embedding backdoors or malicious patterns into the global model.

Defence

Despite the different attack vectors, both Byzantine and poisoning attacks highlight the need for the central server to distinguish between honest and malicious updates during model aggregation.

To defend against such attacks, researchers have proposed Byzantine-resilient algorithms, which aim to filter or correct malicious updates. These generally fall into three categories:

• 1. Score-Based Methods: Assign a trust or credibility score to each update, selecting only the most reliable ones for aggregation.
• 2. Median-Based Methods: Use robust statistical techniques like the geometric median to reduce the influence of outliers.
• 3. Distance-Based Methods: Rely on Euclidean distances to detect and eliminate anomalous updates that deviate significantly from the norm.

Together, these defenses aim to enhance the security and robustness of federated learning in edge environments, enabling privacy-preserving intelligence even in untrusted, heterogeneous networks.

Privacy Challenges in Edge Federated Learning

Although federated learning is designed to protect individual nodes by keeping their training data local and avoiding direct data transmission, privacy breaches can still occur during the sharing of model information, such as model weights, between nodes and servers. 

In particular, curious or semi-honest participants, whether they are edge nodes or central servers, may attempt to infer sensitive data through the training process. In edge federated learning, the data exchanged typically includes locally computed model updates and aggregated weights from the server. As a result, attackers can potentially extract private information by analyzing these weight updates, either directly from the data uploaded by nodes (in the case of a curious server) or from the aggregated model parameters (in the case of a curious node).

Attacks

This risk essentially boils down to reconstructing private training data from model updates. Two common types of privacy attacks in federated learning include:

1. 1. Membership Inference Attack: This attack attempts to determine whether a particular data point was used in a node’s training set. If the training data contains sensitive information, such as medical or financial records, the attack can lead to significant privacy violations.
2. 2. Data Inference Attack: In this case, the adversary tries to recover the actual training data or a specific class of data used by a node, based on the model information shared during training.

Defences

To address these risks, privacy-preserving techniques in edge federated learning typically fall into two main categories:

1. 1. Algorithm-Based Methods: The most common approach is differential privacy, which involves adding carefully calibrated noise to the model updates before sharing them. This technique helps obscure individual data points while still allowing the model to learn effectively.
2. 2. Encryption-Based Methods: These operate at the communication layer, ensuring that information exchanged between parties is secure. Techniques such as secure multi-party computation (SMPC) are often used to encrypt data during transmission, so that it can be processed collaboratively without being exposed.

Many advanced privacy-preserving strategies today fall into these two main categories. However, some solutions combine multiple techniques from both categories.

EndNote

This article provided an overview of edge federated learning, emphasizing its role in enabling decentralized, privacy-preserving machine learning across resource-constrained edge environments. Information presented here integrates insights from recent academic research to offer accurate, current perspectives.

We examined the architectural layers, classifications (cross-device and cross-silo), and the integration of federated learning into edge computing systems.

We also explored key challenges related to communication efficiency, system scalability, and privacy threats, alongside mitigation techniques such as differential privacy, secure multi-party computation, and Byzantine-resilient aggregation.

As research continues to advance, edge federated learning stands out as a promising paradigm for intelligent, secure, and efficient AI at the edge.