A computationally hierarchical federated learning training method

Abstract

The present invention belongs to the field of Internet of Things technology and discloses an efficient computationally hierarchical federated learning training method, comprising: Step 1: constructing a computationally hierarchical federated learning system model; Step 2: allowing users to participate in federated training in the federated learning system model without the need for local gradient calculation or continuous data transmission, which is implemented through online and offline methods; and Step 3: providing solutions to challenges presented to the federated learning system model by competition, user mobility, new data tasks, and privacy protection. The present invention enables resource-constrained devices to perform federated training based on on-device training and lightweight message transmission. The online and offline methods allow devices to participate in federated training without the need for local gradient calculation or continuous data transmission, thereby reducing the training burden and communication overhead while ensuring performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet of Things, and specifically is a computationally hierarchical federated learning training method. Background Art

[0002] With the proliferation of mobile devices and edge devices such as smart vehicles and drones, user-generated data is experiencing a continuous explosion in growth. Against this backdrop, data-driven artificial intelligence technologies, such as machine learning, have made significant progress in a wide range of applications. However, traditional machine learning solutions require the collection of distributed user data to a central server for model training, raising concerns about data privacy. Federated learning has been proposed to address this issue by enabling decentralized nodes to collaboratively train global models without sending local data. Federated learning, based on privacy-preserving mechanisms, has become a research hotspot in distributed machine learning.

[0003] Unlike centralized machine learning, the model training process in federated learning occurs on edge devices, where the raw data is collected and stored. Furthermore, to integrate knowledge learned across devices, a central server collects these model parameters and aggregates them into a global model. Generally speaking, participating devices bear the primary computational and data transmission responsibilities: training the global model with local datasets and uploading and downloading model updates to and from the server. This requires edge devices to have sufficient energy supply, computing power, and a stable network connection. However, in applications such as the Internet of Things and edge computing, devices with limited battery, network, memory, and computing resources cannot provide efficient collaborative training.

[0004] Most related research suggests reducing the communication load by using gradient compression or sparsification. Some researchers have investigated using locally accumulated gradients instead of regular gradients, and only transmitting sufficiently large gradients. By applying these methods, the amount of data transmitted has been reduced by over 99% without degrading model performance. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a computationally hierarchical federated learning training method, which constructs a computationally hierarchical federated learning system, enabling resource-constrained devices to perform federated training based on device training and lightweight message transmission.

[0006] In order to achieve the above object, the present invention is achieved through the following technical solutions:

[0007] This paper presents a computationally layered federated learning training method. First, a three-tier federated learning system model architecture (CLFL) is constructed. In addition to the edge layer and central server in traditional federation, CLFL also includes a user layer consisting of resource-constrained device nodes. The edge layer nodes not only coordinate collaborative training between user nodes but also provide computational assistance for them. Furthermore, principles for implementing CLFL are presented, and two practical methods are designed that allow devices to participate in federated training without requiring local gradient calculation or continuous data transmission. Finally, this paper provides feasible solutions to several key related research challenges.

[0008] The present invention is a computational hierarchical federated learning training method, specifically:

[0009] Step 1: Build a computationally layered federated learning system model. The federated learning system model consists of three parts: service providers, users, and a coordination center. Service providers are located in the edge layer, users are located in the user layer, and the coordination center is located in the cloud server layer. In the system model, edge nodes coordinate collaborative training between user nodes and provide computational assistance to user nodes.

[0010] Step 2: Build the principles that implement the federated learning system model in Step 1, allowing users to participate in federated training without the need for local gradient calculations or continuous data transmission. This principle is implemented through both online and offline methods.

[0011] Step 3: Provide solutions to the challenges posed by competition, user mobility, new data tasks, and privacy protection in the federated learning system model.

[0012] Preferably, in the federated learning system model of step (1), service providers and users are divided into two layers according to their computing capabilities and data volumes: service providers are located in the edge layer, and users are located in the user layer. Specifically:

[0013] User: As the main beneficiary of the service, user nodes selectively request global models from the service provider at the edge layer. User participation in collaborative training is achieved through online and offline methods;

[0014] Service providers: At the edge layer, each service provider is equipped with high-performance computing hardware to train the federated learning system model and provide high-quality intelligent services to users. Each service provider has at least one training task to perform. Before joining the federated system, the service provider registers its metadata with the coordination center. The coordination center then clusters these service providers according to their training tasks to form different task clusters, and executes the federated learning system within the task clusters.

[0015] Coordination Center: Located at the cloud layer of the federated learning system model, the coordination center is responsible for recording metadata of service providers and responding to queries from all users.

[0016] Preferred: The metadata of each service provider includes three types of information: the identity of the service provider, the training task, and the amount of task data stored locally. The metadata is used by the coordination center to cluster the connected service providers and implement federated learning among them. Before obtaining services from service providers, individual users need to query information from the coordination center and select the most suitable service provider.

[0017] Preferably, the online method in step 2 is to "migrate" the computational load of the training task from the user layer to the edge layer or to build a data enhancer on the edge side.

[0018] Optimally, migrating the computational load of training tasks from the user layer to the edge layer involves introducing a trainable loss function to enhance model training on the edge. The specific process includes the following steps:

[0019] Step 2-1-1: The service provider first trains the model on its local dataset with supervised loss and auxiliary loss;

[0020] Step 2-1-2: The service provider then randomly selects a subset of connected users and distributes the trained model to the subset;

[0021] Step 2-1-3: The user evaluates the quality of the updated model on the local dataset and sends the performance score back to the service provider;

[0022] Step 2-1-4: The service provider trains the value function using the auxiliary loss parameters and performance scores;

[0023] Step 2-1-5: Use the value function to optimize the auxiliary loss function.

[0024] Preferably, building a data enhancer on the edge side specifically includes: before the service provider trains the target model, the data enhancer provides perturbations to the local data.

[0025] Preferably: the offline method is implemented using a knowledge distillation method. Specifically, in the knowledge distillation method, all user-level nodes communicate with the service provider only at three time points. The first time point is when the service provider completes the training of the model without the participation of user nodes, that is, local training or collaborative training is completed at the edge, and the pre-trained model will be downloaded by each user. After that, the user performs fine-tuning to customize the pre-trained model. During this period, the service provider waits for the second communication with the user, that is, the user uploads the locally modified model to the service provider. After receiving the personalized model, the service provider performs knowledge distillation to refine the target model by aggregating the user's knowledge. This process will be performed for several rounds on the edge side until the model converges. Finally, the user downloads the target model from the service provider and can freely disconnect from the network connection with the service provider.

[0026] Preferably, the step 3 is specifically as follows:

[0027] To address the challenges of competition, the federated learning system model is to design incentive mechanisms to promote collaborative training and achieve mutual benefits;

[0028] User mobility can be addressed at both the network and application layers. At the network layer, this can be achieved by expanding network coverage or ensuring user connectivity through forwarding within the user's mobility range. At the application layer, this can be achieved by switching to other servers and resuming interrupted training.

[0029] As new data tasks are continuously presented to the federated learning system model, the federated learning system model uses automated machine learning to automatically formulate appropriate algorithms and training hyperparameters based on the meta-features of the training tasks and training sets.

[0030] Privacy and Protection: When gradient information is shared between nodes, some attacks may result in the reconstruction of private data or degrade global model performance. For example, edge nodes can construct data inference attacks to expose private data on user devices. This is especially true when the training set on the user device is small, as reconstruction requires less computation. In addition to privacy threats, attackers may also attempt to prevent the model from achieving performance gains during training. To intuitively reflect the possible attack threats in CLFL, a threat model and feasible solutions for each entity are provided here:

[0031] 1) Coordination Center: The coordination center can be replaced by an attacker, disrupting the operation of the entire system. In this case, the best defense against attacks from the central node is to disconnect and select a new coordination center. An honest but curious center tends to analyze all transmitted information and attempt to obtain private information. This threat has been extensively studied in traditional federated learning, and the most common solution is to adopt differential privacy or encryption methods.

[0032] 2) Edge Nodes: Malicious edge nodes can attack other edge nodes and users by injecting false data into transmitted messages. Filtering out-of-gradient information or identifying and blocking malicious nodes are possible solutions. As with the coordination center, privacy-preserving schemes are needed to prevent honest but curious edge nodes from leaking private data. 3) Users: Because in our system, users can connect to multiple edge nodes to obtain multiple services, malicious users can accelerate the damage of the system by anonymously spreading harmful messages to multiple edge nodes. A possible solution is to implement a secure system with abnormal user identification and strict user registration.

[0033] The beneficial effects of the present invention are:

[0034] (1) This paper constructs a computational hierarchical federated learning system model that enables resource-constrained devices to perform federated training based on device training and lightweight message transmission;

[0035] (2) This paper presents the implementation principles of CLFL and designs two example methods that allow devices to participate in federated training without the need for local gradient calculation or continuous data transmission;

[0036] (3) This paper proposes some key related research challenges, including incentive mechanisms, user mobility, automated machine learning, and privacy protection in CLFL, and provides feasible solutions to some of these challenges; (4) The simulation results of this paper show that compared with existing methods, the method proposed in this paper is feasible, reducing the training burden and communication overhead while ensuring performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the computational hierarchical federated learning system model of the present invention.

[0038] Figure 2 It is a flow chart of the online method based on the auxiliary loss function of the present invention.

[0039] Figure 3 It is a flow chart of the online method based on the auxiliary data enhancer of the present invention.

[0040] Figure 4 This is a flow chart of the offline method based on knowledge distillation of the present invention.

[0041] Figure 5 1 is a performance curve diagram of the five methods of the present invention. DETAILED DESCRIPTION

[0042] The following diagrams illustrate embodiments of the present invention. For clarity, many practical details are included in the following description. However, it should be understood that these practical details are not intended to limit the present invention. In other words, in some embodiments of the present invention, these practical details are not essential.

[0043] The present invention is a computational hierarchical federated learning training method, comprising the following steps:

[0044] S1. Build a computationally hierarchical federated learning system model consisting of three components: service providers (SPs), individual users, and a coordination center. Both service providers and users are participants in federated learning. Unlike traditional federations, these participating nodes are divided into two tiers based on their computing power and data volume: SPs are located in the edge tier, and users are located in the user tier. The following sections describe the main functions of these entities in detail.

[0045] User: As the main beneficiary of the service, user nodes can selectively request global models from the service provider at the edge layer. However, due to the heterogeneity of the data, external models may perform poorly on local datasets. To solve this problem, traditional federated learning regards users as training nodes, ignoring the fact that these individual users are mainly connected to the system through wireless networks with personal mobile devices. In this way, computing power and battery capacity are usually limited on the user side; training the model alone can easily lead to large energy consumption and CPU computing load. Under these limitations, users need to undertake heavy computing and transmission tasks in order to participate in collaborative training. The present invention provides two strategies to solve this type of problem, and names them as online method and offline method according to the communication modes of these two effective schemes.

[0046] Service Providers: At the edge layer, each SP is equipped with high-performance computing hardware, such as graphics processing units (GPUs), to perform training of machine learning models, and usually acts as a CPU / GPU intensive server. In addition, such nodes also have sufficient training data and storage capacity. The main purpose of the SP is to collaboratively train a well-performing model, aiming to provide high-quality intelligent services to individual users. In contrast to traditional federation, the SP is the main training node of the federation, and each SP contains at least one training task to be performed. Before joining the federation system, these SPs should register their metadata with the coordination center, i.e., the cloud server. The coordination center then clusters these SPs according to their training tasks to form different task clusters, and executes the federated learning scheme within a specific task cluster. In addition, multi-task federated learning (MFL) is also feasible in this architecture, which aims to improve the performance of a specific task by leveraging information from other related tasks.

[0047] Coordination Center: Located in the cloud layer of the federated learning system, the coordination center is responsible for recording metadata of service providers and responding to queries from other entities (usually acting as an I / O-intensive server). Specifically, the metadata of each SP includes three types of information: the identity of the SP such as its IP address, the training task, the classification or regression task it performs, and the amount of task data stored locally. These metadata are used by the coordination center to cluster connected SPs and implement federated learning among them. For example, applying federated averaging requires confirming the network addresses of SPs with the same training task and their data volume in order to assign aggregate weights to each training model. On the other hand, individual users need to query information from the coordination center before obtaining services from the SP and select the most suitable SP.

[0048] S2. We present the principles for implementing CLFL and design two practical approaches: an online approach and an offline approach. These approaches allow devices to participate in federated training without requiring local gradient computation or continuous data transmission.

[0049] In a layered computing system architecture, participating nodes are divided into two categories based on computing power and data volume. Within this architecture, inter-layer training collaboration also includes two components: cloud-to-edge collaboration and edge-to-user collaboration. This paper provides two types of solutions: online and offline approaches. The principles of edge-to-user collaboration can be summarized as: reducing user computing and communication burdens, protecting the privacy of both users and service providers, and ensuring the quality of service received by users. This section introduces these methods using a scenario where a single service provider connects to multiple users.

[0050] (1) Online Methods: The main idea of online methods is to allow users to participate in the SP training process without excessive computational effort. Specifically, users act as training instructors to judge the SP's training model. The SP then needs to modify model parameters based on user feedback to improve its service quality. For demonstration purposes, this paper provides two examples to explain the main execution process of the online method.

[0051] Auxiliary loss function: such as Figure 2As shown in the figure, a trainable loss function (a two-layer fully connected neural network) is introduced to enhance model training on the edge. The purpose of designing an auxiliary loss function is to enable the trained model to achieve performance improvements on the user side without directly accessing user data. This mechanism is often implemented in meta-learning schemes for domain generalization problems, where multiple datasets are collected in a central server and the auxiliary loss function is optimized using the validation dataset. In this scenario, due to privacy regulations, edge nodes cannot independently train the loss function; furthermore, offloading this training task to user devices would significantly increase the user's computational cost. To address this issue, a value function with the same structure as the auxiliary loss is used to optimize the loss function parameters. The key idea is to collect user feedback and train a value function to map the loss function parameters to feedback scores. Using the trained value function, the SP is able to optimize the auxiliary loss function to maximize the value function. This optimization strategy is similar to optimization strategies based on reinforcement learning (RL), where the user-side service quality can be viewed as the environment and the auxiliary loss function acts as the actor in RL. From another perspective, the value function can be viewed as a hypernetwork that learns to produce a parameterized loss function to guide the training of the target model.

[0052] Using the above components, the training process for each round of communication can be summarized into five steps: 1) The SP first trains the model on its local dataset with a supervised loss and an auxiliary loss; 2) The SP then randomly selects a subset of connected users and distributes the trained model to them; 3) These users then evaluate the quality of the updated model on their local dataset and send performance scores back to the SP; 4) The SP trains a value function using the parameters of the auxiliary loss and the performance scores; 5) The value function is used to optimize the auxiliary loss function. Unlike traditional federation, the computational cost involved on the user side only includes the forward propagation of input data, skipping the backward propagation of gradient calculations. Furthermore, the user device no longer needs to cache the outputs of intermediate neural network layers (which is required for backpropagation); the uploaded messages only contain feedback on the model's performance. Through this online training approach, users can significantly reduce computation, communication, and memory usage. In return, the SP at the edge layer takes on more training tasks, thus incurring higher computational costs. In short, this method "shifts" the computational load of training tasks from the user layer to the edge layer without exposing personal data.

[0053] Auxiliary Data Enhancer: Another feasible approach for edge-to-user collaboration is to build a data enhancer on the edge side. Figure 3As shown in the figure, before the SP trains the target model, the enhancer provides perturbations to the local data, which is used to train a more robust model. Similar to the auxiliary loss, the effectiveness of the enhancer is evaluated by the user and optimized based on the feedback score given by the user. The training of the enhancer relies on a value function that learns the mapping from the parameters of the auxiliary enhancer to the performance score. This user-guided data augmentation scheme aims to enhance the model to learn more discriminative or domain-invariant features from limited data. The efficiency of this method is the same as that of the auxiliary loss-based method, which reduces the user's computational load, communication overhead, and memory usage by introducing additional computational cost to the SP. In this system, this trade-off is reasonable because it is acceptable for the SP equipped with high-performance computing hardware.

[0054] (2) Offline method: Although the online method is sufficient to build a CLFL system, this method still requires users to frequently download models from the server and participate in guiding edge-side model training. The transmission delay and quality of the wireless network may become a bottleneck for system efficiency and deteriorate the performance fairness of the target model. For example, SP tends to obtain more feedback scores from users with superior network environment, thereby reducing the service quality of users with lower network bandwidth. Due to this challenge, the present invention balances the communication complexity and computational cost of the user side and proposes an offline method to reduce network bandwidth requirements. As an embodiment, the present invention provides a solution based on knowledge distillation (KD) to present the key operations of the offline method.

[0055] Knowledge Distillation: In this approach, all user-level nodes communicate with the SP at only three points in time. The first point is when the SP completes model training without the participation of user nodes. That is, local or collaborative training is completed at the edge, and the pre-trained model is downloaded by each user. Afterward, the user performs fine-tuning or model personalization to customize the pre-trained model. During this time, the SP awaits a second communication with the user, during which the user uploads the locally modified model to the SP. After receiving the personalized model, the SP performs knowledge distillation, refining the target model by aggregating the user's knowledge. This process continues for several rounds on the edge until the model converges. Finally, the user downloads the target model from the SP and is free to disconnect from the SP's network. Figure 4 This paper summarizes the complete process of instantiating this offline method. Federated learning based on knowledge distillation has been widely studied as an effective method for extracting knowledge from existing user models. Unlike previous work, this scheme does not require repeated distribution of the global model and further optimization on the user side. The personalized model is uploaded only once and continuously guides the SP to improve the target model. In this way, network performance no longer affects the system's operational efficiency and service quality, but is only affected by the choice of personalized learning algorithm and extraction method.

[0056] S3. Some key related research challenges are proposed, including incentive mechanisms, user mobility, automated machine learning, and privacy protection in CLFL, and feasible solutions are provided for some of these challenges.

[0057] Competition and Incentives: In traditional federated learning, all participating nodes are direct beneficiaries of collaborative training, thus participating in training brings practical benefits to nodes. In contrast, edge nodes (SPs) in the CLFL system earn rewards from users by providing them with high-quality and reliable services. To compete, each SP tends to attract as many users as possible to earn more rewards. Collaborating with more users can further improve the quality of their services. Furthermore, these SPs may refuse to collaborate with each other or intentionally provide malicious information during collaborative training. In this case, the main research challenge of CLFL is to design a reasonable incentive mechanism to promote collaborative training and achieve mutual benefits. One possible solution is to embed a transaction system into a consensus platform, such as a blockchain, and distribute the revenue earned by a single SP to its collaborating SPs. This mechanism allows SPs to earn revenue by participating in joint training with other SPs, thereby strengthening cooperation and weakening competition among SPs.

[0058] User mobility: To update smart services or participate in model training in a timely manner, users with mobile devices and wireless networks need to maintain connectivity to their selected edge servers. However, due to user mobility, the maintainability of continuous communication links may change. In extreme cases, collaborative training between users and edge servers may be damaged or interrupted due to network changes; achieving training robustness in such situations is an open problem in CLFL. Solutions can be designed at both the network and application layers. From the network layer, user connectivity can be ensured by expanding network coverage or by forwarding messages to each other within the user's mobility range. From the application layer, switching to a connection to another server and resuming interrupted training is also a viable solution.

[0059] AutoML for CLFL: As new edge servers and users join the CLFL, new data and training tasks are constantly presented to the system. This behavior necessitates the design of new learning algorithms and training strategies to adapt to new datasets and emerging tasks. Manually designing optimal algorithms and their hyperparameters for specific scenarios often requires specialized knowledge and significant labor costs. For example, scenarios such as smart healthcare require the combined expertise of AI and medical experts to develop high-performance algorithms. Furthermore, training strategies in CLFL should also consider user engagement, including acceptable communication and training loads. Automated machine learning (AutoML) can streamline this process by automatically determining appropriate algorithms and training hyperparameters based on meta-features of the training task and training set. The specific process of AutoML is to evaluate the performance of any algorithm on a set of tasks and then learn a performance model that outputs the optimal algorithm for new tasks. In CLFL, AutoML tasks can be offloaded to a coordination center in the cloud, which has access to metadata about training tasks and performance information for all edge nodes.

[0060] Privacy and Protection: When gradient information is shared between nodes, some attacks may have the effect of reconstructing private data or deteriorating global model performance. For example, edge nodes can construct data inference attacks to expose private data on user devices. This is especially true when the training set on the user device is small (reconstruction requires less computation). In addition to privacy threats, attackers may also attempt to prevent the model from achieving performance gains during training. To intuitively reflect the possible attack threats in CLFL, the threat model and feasible solutions for each entity are given here:

[0061] 1) Coordination Center: The coordination center could be replaced by an attacker, disrupting the entire system. In this case, the best defense against attacks from the central node is to disconnect and select a new coordination center. An honest but curious center tends to analyze all transmitted information and attempt to obtain private information. This threat has been extensively studied in traditional federated learning, and the most common solutions are differential privacy or encryption methods.

[0062] 2) Edge Nodes: Malicious edge nodes can attack other edge nodes and users by injecting false data into transmitted messages. Filtering out-of-gradient information or identifying and blocking malicious nodes are possible solutions. As with the coordination center, privacy protection schemes are needed to prevent honest but curious edge nodes from leaking private data. 3) Users: Because users can connect to multiple edge nodes to access multiple services in the federated learning system of the present invention, malicious users can accelerate system damage by anonymously spreading harmful messages to multiple edge nodes. One possible solution is to implement a secure system with abnormal user identification and strict user registration.

[0063] In order to make the performance and efficiency of online and offline methods clearer, the present invention provides an embodiment of auxiliary loss (AL) and knowledge distillation (KD) methods.

[0064] Set up a single edge node (SP) and 20 users. Two cross-domain datasets are distributed to the SP and users; the users and SP hold the SVHN dataset and the MNIST dataset, respectively. In addition, the user side uses a non-normal distribution setting based on the Dirichlet distribution, and the distribution parameter α is set equal to 0.5. The evaluation of these algorithms will be performed on the test set of SVHN to reflect the performance of these algorithms on the user side. In comparison, the traditional federated learning method (FedAvg, i.e., the federated averaging algorithm) is also implemented here, in which the above SP is both a central node and a training node. The number of training rounds in FedAvg and AL is set to 100. In each training round, each node performs a local training and interacts with the SP. Since the offline method requires that the local node and the central node communicate only at a specific point in time, the present invention sets the total number of training rounds for the user side and the edge side to 100 times. To visualize the impact of different user-side local training times on the final performance, the KD method is performed here using four different settings: KD-1 (50 rounds of pre-training, 40 rounds of fine-tuning training, and 10 rounds of distillation training), KD-2 (50 rounds of pre-training, 30 rounds of fine-tuning training, and 20 rounds of distillation training), KD-3 (50 rounds of pre-training, 20 rounds of fine-tuning training, and 30 rounds of distillation training), and KD-4 (50 rounds of pre-training, 10 rounds of fine-tuning training, and 40 rounds of distillation training). The neural network used in the experiment is ResNet-9; the loss function and value function use the same structure as in, and all neural networks are optimized using stochastic gradient descent (SGD), and the learning rate is set to 1e-3.

[0065] Figure 5The performance curves of the five methods are shown, where the dashed line represents the local training time on the user node. It can be observed that the performance of the offline methods (KD-1 to KD-4) continues to decline and oscillate until the local fine-tuning stage. This is due to the data disparity between the edge and user nodes; in other words, the model trained on the SP cannot be directly used to serve users. Another phenomenon with offline methods is that more or less local fine-tuning does not necessarily lead to better performance. The results indicate that the impact of local training needs to be considered when applying offline methods to a system. For the online method (AL), the performance curve exhibits the same characteristics as the offline method in the first 20 rounds and achieves faster convergence than FedAvg in subsequent training. This means that sufficient local user evaluation is necessary to demonstrate the importance of the auxiliary loss. Furthermore, the superior performance of AL reflects its advantages in extracting robust features and adapting to non-IID data.

[0066] Table I summarizes the final test accuracy (Acc.), communication cost (CMC), and computation cost (CPC) involved in these methods, where CMC represents the total amount of data sent and received during the entire training process. All methods use FedAvg as the comparison baseline.

[0067]

[0068] The results show that offline methods require the least data transfer and training computation in terms of communication costs, but perform poorly. In contrast, online methods achieve the highest accuracy while halving the training load and reducing communication overhead. This comes at the cost of the SP incurring the additional training load of auxiliary loss and value functions. While the above results demonstrate the respective advantages and disadvantages of online and offline methods, further research is needed to determine their applicability in practical applications for three reasons:

[0069] (1) The training algorithm used on the edge determines the quality of the pre-trained model. This quality refers to whether it conflicts with the model personalization on the user side. In the case of a low-quality model, the time required to personalize the pre-trained model will be greatly increased, thereby reducing the local computational efficiency of the user.

[0070] (2) The effectiveness of fine-tuning training on the user side is the second reason. Through effective algorithms, users can obtain models with better performance in less training time. This type of algorithm is also called passive domain adaptation technology.

[0071] (3) Online methods require more frequent interactions between edge nodes and user nodes, which can significantly improve the unbiasedness of edge-side training. In contrast, achieving the same effect with offline methods depends more on the choice of algorithm.

[0072] Since both online and offline methods have their own advantages and disadvantages, there is no theoretically optimal strategy that is suitable for all CLFL-based applications. For specific situations, designing an improved or hybrid method of these two methods may produce better results.

[0073] The computational hierarchical federated learning architecture constructed by the present invention enables resource-constrained devices to perform federated training based on device training and lightweight message transmission.

[0074] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A computationally hierarchical federated learning training method, characterized by: The federated learning training method includes the following steps: Step 1: Build a computationally layered federated learning system model. The federated learning system model consists of three parts: service providers, users, and a coordination center. Service providers are located in the edge layer, users are located in the user layer, and the coordination center is located in the cloud server layer. In the system model, edge nodes coordinate collaborative training between user nodes and provide computational assistance to user nodes. Step 2: Build the principles that implement the federated learning system model in Step 1, allowing users to participate in federated training without the need for local gradient calculations or continuous data transmission. This principle is implemented through both online and offline methods. Step 3: Provide solutions to the challenges posed by competition, user mobility, new data tasks, and privacy protection in federated learning system models. in: Migrating the computational load of training tasks from the user layer to the edge layer involves introducing a trainable loss function to enhance model training on the edge. The specific process includes the following steps: Step 2-2-1: The service provider first trains the model on its local dataset with supervised loss and data augmenter; Step 2-2-2: The service provider then randomly selects a subset of connected users and distributes the trained model to the subset; Step 2-2-3: The user evaluates the quality of the updated model on the local dataset and sends the performance score back to the service provider; Step 2-2-4: The service provider trains the value function using the auxiliary loss parameters and performance scores; Step 2-2-5: Use the value function to optimize the auxiliary loss function; Before the service provider trains the target model, the data enhancer provides perturbations to the local data. The specific steps are as follows: Step 2-1-1: The service provider first trains the model on its local dataset with supervised loss and auxiliary loss; Step 2-1-2: The service provider then randomly selects a subset of connected users and distributes the trained model to the subset; Step 2-1-3: The user evaluates the quality of the updated model on the local dataset and sends the performance score back to the service provider; Step 2-1-4: The service provider trains the value function using the parameters and performance scores of the data enhancer. Step 2-1-5: Optimize the data augmenter using the value function; The offline method is implemented using a knowledge distillation method. Specifically, in the knowledge distillation method, all user-level nodes communicate with the service provider at only three time points. The first time point is when the service provider completes the training of the model without the participation of user nodes, that is, local training or collaborative training is completed at the edge, and the pre-trained model will be downloaded by each user. After that, the user performs fine-tuning to customize the pre-trained model. During this period, the service provider waits for the second communication with the user, that is, the user uploads the locally modified model to the service provider. After receiving the personalized model, the service provider performs knowledge distillation to refine the target model by aggregating the user's knowledge. This process will be performed for several rounds on the edge side until the model converges. Finally, the user downloads the target model from the service provider and freely disconnects from the network connection with the service provider.

2. The computationally hierarchical federated learning training method according to claim 1, characterized in that: In the federated learning system model of step 1, service providers and users are divided into two layers based on their computing capabilities and data volumes: service providers are located in the edge layer, and users are located in the user layer. Specifically: User: As the main beneficiary of the service, user nodes selectively request global models from the service provider at the edge layer. User participation in collaborative training is achieved through online and offline methods; Service providers: At the edge layer, each service provider is equipped with high-performance computing hardware to train the federated learning system model and provide high-quality intelligent services to users. Each service provider has at least one training task to perform. Before joining the federated system, the service provider registers its metadata with the coordination center. The coordination center then clusters these service providers according to their training tasks to form different task clusters, and executes the federated learning system within the task clusters. Coordination Center: Located at the cloud layer of the federated learning system model, the coordination center is responsible for recording metadata of service providers and responding to queries from other entities.

3. The computationally hierarchical federated learning training method according to claim 2, characterized in that: The metadata of each service provider includes three types of information: the identity of the service provider, the training task, and the amount of locally stored task data. The metadata is used by the coordination center to cluster connected service providers and implement federated learning among them. Before obtaining services from a service provider, individual users need to query information from the coordination center and select the most suitable service provider.

4. The computationally hierarchical federated learning training method according to claim 1, characterized in that: The online method in step 2 is to "migrate" the computational load of the training task from the user layer to the edge layer or build a data enhancer on the edge side.

5. The computationally hierarchical federated learning training method according to claim 1, characterized in that: The step 3 is specifically as follows: To address the challenges of competition, the federated learning system model is to design incentive mechanisms to promote collaborative training and achieve mutual benefits; User mobility is addressed at the network and application layers. At the network layer, user connectivity is ensured by expanding network coverage or forwarding traffic within the user's mobility range. At the application layer, connections are switched to other servers to resume interrupted training. As new data tasks are continuously presented to the federated learning system model, the federated learning system model uses automated machine learning to automatically formulate appropriate algorithms and training hyperparameters based on the meta-features of the training tasks and training sets.

Images