A robust federated learning method

Abstract

The present application relates to a kind of robust federated learning methods, belong to communication technical field, including the following steps: S1: the federated learning algorithm capable of reducing the model performance loss caused by the data heterogeneity between nodes is constructed, and the communication model between edge network with unreliable communication link and calculation model;S2: according to the analysis result of robust federated learning and channel condition, improve the performance of federated learning model after training, meet in the deployment in the edge network of limited resources, construct federated learning performance optimization problem;S3: under mathematical assumption, the theoretical convergence edge of federated learning algorithm under the influence of unreliable communication link is deduced;S4: the federated learning performance optimization problem is converted into the maximum matching problem of bipartite graph, and the best node selection and channel resource allocation strategy in each round of training are found using Hungarian algorithm.

Description

Technical Field

[0001] This invention belongs to the field of communication technology and relates to a robust federated learning method. Background Technology

[0002] In the era of the Internet of Things (IoT), the widespread adoption of IoT devices, the increase in high-bandwidth applications, and the performance and quantity growth of mobile devices have led to an exponential increase in the amount of data generated on edge networks. This phenomenon has prompted a migration of big data sources from massive centralized cloud data centers to IoT edge infrastructure devices, which also possess substantial computing and storage resources. To better handle large volumes of data, Artificial Intelligence (AI) has become one of the main methods for big data analysis in recent years. This is because AI's ability to analyze large datasets and obtain highly generalized patterns has proven far more effective than human methods in many fields.

[0003] However, the diverse data in edge networks often exists in the form of "data silos," meaning that within an organization, data is isolated and stored in different departments or systems, unable to be shared or interoperated across systems. Take the financial industry as an example: financial institutions can use AI models to learn about different users' information to provide customized products and services. However, given that financial data is linked to highly sensitive data such as personal financial information, indiscriminate sharing could lead to user information leaks; furthermore, it is usually stored in different banking institutions, preventing free interaction. This deployment method of traditional AI models, which requires direct collection of customer data, naturally raises numerous information security and other issues.

[0004] With the growth of data volume in the network, machine learning (ML) in the AI ​​field has also undergone a significant transformation. Compared to traditional centralized learning methods that rely on large-scale datasets on cloud servers, the current trend in machine learning emphasizes distributed learning using small-scale datasets from edge devices. This paradigm shift is driven by the increasing need to reduce latency and minimize transmission overhead, and the transition of data sources from cloud-centric systems to edge-centric IoT infrastructure. Furthermore, previously, AI models could not be trained on edge nodes due to the performance limitations of mobile devices, and were objectively deployed only on high-performance servers. However, with the continuous advancement of chip manufacturing technology in recent years, devices such as mobile phones, tablets, and wearable smart devices now possess considerable computing power. Therefore, to fully utilize the data and device computing power in edge networks while protecting user privacy and breaking down "data silos," federated learning (FL) has emerged as one of the most promising distributed learning frameworks, attracting widespread attention from academia and industry since its inception.

[0005] FL, a distributed learning paradigm, emphasizes collaboration between a large number of nodes and a central server, training a global model on their respective collected data. Within each global training epoch, the central server sends the global model to all node devices within the framework. The node devices then train this global model using their locally stored datasets. After a certain number of local training epochs, the node devices upload their local models to the central server for aggregation to obtain a new global model. Importantly, this paradigm only requires the exchange of model parameters between the node devices and the central server, thus avoiding the need to directly share raw data.

[0006] While the idea of ​​deploying federated learning (FL) to train machine learning (ML) models in edge networks holds considerable promise, its practical implementation faces several challenges: limited resources in edge networks restrict the number of communication rounds between nodes and base stations within the federated learning framework; unreliable communication links can lead to issues such as device outages, local model uploads, or global model downloads; and data heterogeneity between nodes results in non-independent and identically distributed (Non-IID) data across different nodes, significantly impacting FL model performance. Therefore, ensuring a robust FL method can be deployed despite these three major challenges is crucial for successfully completing various business operations within edge networks. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a robust federated learning method that addresses the need to deploy robust federated learning services in edge wireless networks by comprehensively considering factors such as the resource characteristics of the wireless network, data heterogeneity between nodes, and channel conditions, thereby optimizing the model performance of the federated learning algorithm.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A robust federated learning method includes the following steps:

[0010] S1: Construct federated learning algorithms that can reduce model performance loss caused by data heterogeneity between nodes, as well as communication and computation models between edge networks with unreliable communication links;

[0011] S2: Based on the analysis results of robust federated learning and channel conditions, improve the performance of the federated learning model after training to meet the requirements of deployment in resource-limited edge networks, and construct a federated learning performance optimization problem.

[0012] S3: Derive the theoretical convergence edge of the federated learning algorithm under the influence of unreliable communication links under mathematical assumptions;

[0013] S4: Transform the federated learning performance optimization problem into a bipartite graph maximum matching problem, and use the Hungarian algorithm to find the best node selection and channel resource allocation strategy in each round of training.

[0014] Furthermore, step S1 specifically includes the following steps:

[0015] S11: The federated learning framework is a global federated learning model trained collaboratively by a central server and multiple edge node devices. In this process, nodes only transmit local model parameters to the central server, and all nodes do not communicate directly with each other. At the beginning of each training round, the central server distributes the global model to each node, and each node uses its local dataset to perform a local update process on the global model with a certain step size, resulting in different local models. After the nodes participating in federated learning send their local models to the central server, the central server performs weighted aggregation to obtain a new round of global model, and repeats the above steps until the target accuracy or preset training time is reached.

[0016] S12: Defined within federated learning in edge networks, a base station has both a transceiver system and an aggregation system, can act as a central server, and has a set of nodes within its coverage area. The dataset collected at node i is Furthermore, the solver used by node i to find the optimal solution for the local model is:

[0017]

[0018] in Let be the model parameters of node i after the τth local update in generation t. Let w be the model parameters of node i after the (n-1)th local update in generation t. t Let be the parameters of the global federated learning model in generation t. The alternative loss function is defined as μ, where μ is the coefficient of the proximal term.

[0019] S13: The latency required for node i to train the local model is:

[0020] T i cmp =τSλ i / f i

[0021] Where S is the number of bits required for the transmission model of node i, δ is the energy consumption coefficient of the GPU of node i, and f i For its frequency, λ i The number of GPU cycles required to process one bit of data; the energy consumption required for node i to train the local model is...

[0022]

[0023] S14: The communication resources between the base station and the node are mainly the available spectrum resources in the network, with a total bandwidth of B, divided into R sub-channels of equal size. The bandwidth of each sub-channel is:

[0024] W = B / R (R ≤ U)

[0025] The available set of sub-channels is represented as After node i completes its local update process, the uplink rate at which it transmits its local federated learning model parameters to the base station is:

[0026]

[0027] Where χ represents the channel allocation decision, x i,l The corresponding decision indicator is x. i,l =1 indicates that the model parameters of node i will be uploaded to the central server through channel l; otherwise, x i,l =0; Additional conditions Used to restrict a single subchannel to be used only by a single node within the FL framework; P i Let H be the uplink transmission power of node i, and H be the corresponding uplink channel gain. l The interference caused by other nodes using uplink channel l is denoted as N0, where N0 is the power spectral density of the channel noise; the uplink transmission delay between node i and the base station is:

[0028] T i com (χ)=S / r i (χ)

[0029] S15: Based on OFDMA technology and assuming the channel is a quasi-static Rayleigh fading channel, the probability that each local federated learning model is affected by noise during uplink transmission is:

[0030]

[0031] in Let be the PER of node i on subchannel l, and l0 be the signal quality threshold of the quasi-static Rayleigh fading channel. This represents the average signal-to-noise ratio.

[0032] Furthermore, step S2 specifically includes:

[0033] S21: Taking into account q in step S15 i,l Then, the global aggregation process in federated learning was modified as follows:

[0034]

[0035] Where n i n is a flag indicating whether the base station allows node i to participate in the federated learning process. i =1 means the base station allows node i to upload local model parameters; e i e is a flag indicating whether to aggregate node i in the model. i =1 means that the local model parameters of node i are not affected by noise, and it is allowed to participate in aggregation;

[0036] S22: The optimization problem of minimizing empirical risk is:

[0037]

[0038] stC1:

[0039] C2:

[0040] C3:

[0041] C4:

[0042] C5:

[0043] Where n = [n1,...,n] i Select a vector for the node. Let γ be the average loss function across all node data; C1-C3 represent the requirement to ensure a one-to-one correspondence between each node i participating in aggregation and the allocable sub-channel l; C4 represents the requirement that nodes participating in training must meet latency requirements, where γ T C5 represents the maximum latency for completing each round of the federated learning algorithm; C5 indicates that the nodes participating in training must meet the energy consumption requirements, where γ E This represents the maximum energy consumption for completing the federated learning algorithm in each round.

[0044] Furthermore, step S3 specifically includes:

[0045] S31: Let F i (x) satisfies nonconvexity and L-smoothness, and the gradient of the local loss function and the gradient between the local and global loss functions are both bounded, while the mini-batch gradient descent on the local substitution loss function is bounded.

[0046] S32: Based on the assumptions of step S31, derive... The theoretical convergence boundary.

[0047] Furthermore, step S4 specifically includes:

[0048] S41: Based on the theoretical convergence boundary in step S32, the relevant parameters affecting the convergence speed and performance of federated learning are obtained as follows:

[0049]

[0050] Therefore, the optimization problem in step S22 is transformed into:

[0051]

[0052] stC1-C5

[0053] Optimization Objective This ensures that the model can converge to the target accuracy more quickly during training;

[0054] S42: Based on the optimization problem in step S41, the solution process of this problem is regarded as: establishing a one-to-one relationship between each node participating in model uploading and the available sub-channels;

[0055] The optimization problem is transformed into a bipartite graph matching problem. First, a bipartite graph is constructed. Where θ is the set of edges between nodes and sub-channels, θ i,n ∈θ is an edge between node i and sub-channel l that satisfies condition θ i,l ∈{0,1},θ i,l =1 indicates that node i used sub-channel l to upload local model parameters; therefore, θ i,l The weights are assigned as follows:

[0056]

[0057] After determining the weight of each edge in the bipartite graph, the Hungarian algorithm is used to obtain the set of nodes that achieve the maximum matching and the matching relationship between the nodes and the sub-channels. At this point, the base station will select nodes in the node set in each round to upload the local federated learning model using the corresponding sub-channels and aggregate the parameters that are not affected by noise.

[0058] The beneficial effects of this invention are as follows: Firstly, addressing the performance degradation of federated learning models caused by unreliable communication and data heterogeneity among nodes in edge networks with limited resources, this invention provides a robust federated learning method. By analyzing the impact of data heterogeneity among nodes, as well as the communication model, computational model, and unreliable communication environment in the network, a federated learning algorithm is constructed that simultaneously considers incomplete aggregation and minimizing model performance loss caused by data heterogeneity. Secondly, the theoretical convergence boundary of the algorithm is derived, and the optimal node selection and channel resource allocation strategy in each training round is obtained through the Hungarian algorithm. The method provided by this invention can effectively improve the performance of federated learning algorithm models and has broad application prospects.

[0059] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0060] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0061] Figure 1 This is a diagram of the federated learning architecture of the present invention. Detailed Implementation

[0062] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0063] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0064] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0065] Please see Figure 1 This invention provides a robust federated learning method, which preferably includes the following steps:

[0066] Step 1: Constructing a Robust Federated Learning Algorithm and a Communication and Computation Model for Edge Networks with Unreliable Communication Links: Due to the heterogeneity of data among nodes participating in federated learning, and the fact that the uplink transmission of model parameters typically traverses unreliable communication links, while simultaneously meeting finite latency and energy consumption requirements, a robust federated learning algorithm was constructed. The preferred specific steps include:

[0067] Step 1.1: In the federated learning framework, a global federated learning model is trained collaboratively by a central server and multiple edge node devices (hereinafter referred to as nodes). During this process, nodes only transmit local model parameters to the central server and do not directly transmit local data; furthermore, nodes do not communicate directly with each other. At the start of each training round, the central server distributes the global model to each node. Each node then uses its local dataset to perform a local update process on the global model with a certain step size, resulting in a different local model. After the nodes participating in federated learning send their local models to the central server, the central server performs weighted aggregation to obtain a new global model, and repeats the above steps until the target accuracy or preset training time is reached.

[0068] Step 1.2: Define federated learning in the edge network. A base station has a transceiver system and an aggregation system, can act as a central server, and has a set of nodes within its coverage area. The dataset collected at node i is And the solver used by node i to find the optimal solution of the local model is in Let be the model parameters of node i after the τth local update in generation t. Let w be the model parameters of node i after the (n-1)th local update in generation t. t Let be the parameters of the global federated learning model in generation t. The alternative loss function is defined as μ, where μ is the coefficient of the proximal term.

[0069] Step 1.3: The latency required for node i to train the local model is T. i cmp =τSλ i / f i Where S is the number of bits required for the transmission model of node i, δ is the energy consumption coefficient of the GPU of node i, and f i For its frequency, λ i The number of GPU cycles required to process one bit of data. The energy required for node i to train the local model is...

[0070] Step 1.4: The communication resources between the base station and the nodes mainly consist of available spectrum resources in the network, with a total bandwidth of B, divided into R sub-channels of equal size. The bandwidth of each sub-channel is W = B / R (R ≤ U). The set of available sub-channels is represented as follows: After node i completes its local update process, the uplink rate at which it transmits its local federated learning model parameters to the base station is [value missing]. Where χ represents the channel allocation decision, x i,l The corresponding decision indicator is x. i,l =1 indicates that the model parameters of node i will be uploaded to the central server through channel l; otherwise, x i,l =0, in addition, additional conditions The constraint restricts each subchannel to be used only by a single node within the FL framework to ensure a significant reduction in co-channel interference during transmission; P i Let H be the uplink transmission power of node i, and H be the corresponding uplink channel gain. l The interference caused by other nodes using uplink channel l is denoted by N0, where N0 is the power spectral density of the channel noise. The uplink transmission delay between node i and the base station is: T i com (χ)=S / r i (χ);

[0071] Step 1.5: In edge networks with unreliable communication links, the focus is primarily on uplink transmission. Therefore, using OFDMA technology as a background and assuming the channel is a quasi-static Rayleigh fading channel, the probability that each local federated learning model is affected by noise during uplink transmission is: in Let be the PER of node i on subchannel l, and l0 be the signal quality threshold of the quasi-static Rayleigh fading channel. The average signal-to-noise ratio;

[0072] Step 2: Constructing the Federated Learning Performance Optimization Problem: Based on the analysis of robust federated learning and channel conditions in S1, a federated learning performance optimization problem is designed to improve the performance of the trained federated learning model and meet the requirements for deployment in resource-constrained edge networks. The optimization specifically includes the following steps:

[0073] Step 2.1: For performance analysis purposes, it is crucial to study the impact of PER on the global model aggregation process. Taking into account q in S15... i,l Subsequently, the global aggregation process in federated learning was modified to... Where n i n is a flag indicating whether the base station allows node i to participate in the federated learning process. i =1 means the base station allows node i to upload local model parameters; e i e is a flag indicating whether to aggregate node i in the model. i =1 means that the local model parameters of node i are not affected by noise, and it is allowed to participate in aggregation;

[0074] Step 2.2: Considering the need to train federated learning models with limited resources while ensuring their performance, the optimization problem of minimizing empirical risk is as follows:

[0075]

[0076] stC1:

[0077] C2:

[0078] C3:

[0079] C4:

[0080] C5:

[0081] Where n = [n1, ..., n i Select a vector for the node. Let γ be the average loss function across all node data. C1-C3 represent the requirement to ensure a one-to-one correspondence between each node i participating in aggregation and the allocable sub-channel l; C4 represents the requirement that nodes participating in training must meet latency requirements, where γ T C5 represents the maximum latency for completing each round of the federated learning algorithm; C5 indicates that the nodes participating in training must meet the energy consumption requirements, where γ E This represents the maximum energy consumption for completing the federated learning algorithm in each round.

[0082] Step 3: Theoretical Convergence Boundary of Robust Federated Learning: To quantitatively study the performance of the proposed robust federated learning model, under reasonable assumptions common in other literature, the expected decrease in the loss function of the federated learning model after each training round compared to the previous round is derived, thus clarifying which parameters will affect the deployment of a robust federated learning model. The preferred approach includes the following steps:

[0083] Step 3.1: For the purpose of theoretical convergence analysis, the following assumptions are made:

[0084] Assumption 1: Nonconvexity: For any x, F i (x) is a non-convex function.

[0085] Assume 2L-smoothing: for any x and y, F i (x)≤F i (y)+(xy) T ▽F i (y)+L / 2||xy|| 2 .

[0086] Assumption 3: The gradient of the local loss function is bounded: For During the local FL model update process, the expected boundary of the descent gradient of the local loss function is:

[0087] Assumption 4: The gradients of the local and global loss functions are bounded: For Given D1, D2 ≥ 0, the descent gradients of the local loss function and the global loss function satisfy the following relationship:

[0088] Assume that mini-batch gradient descent on the local substitution function is bounded: for From B k ∈D i Any data ξ∈B drawn uniformly and randomly k Its expected stochastic gradient is unbiased and satisfies the following condition: And the variances between stochastic gradients satisfy

[0089] Assumption 1 holds true for loss functions in neural networks such as MLPs or CNNs. Assumption 2 characterizes the rate of change of the function. Assumptions 3 and 4 constrain the local gradient updates of each node in each round, measuring the sample distribution across the local dataset. Assumption 5 illustrates the performance gap between mini-batch SGD and GD.

[0090] Step 3.2: Based on the assumptions of Step 3.1, the following was derived: The theoretical convergence boundary is:

[0091]

[0092] in K i Let λ be the amount of local data at node i, λ be the local learning rate of federated learning, and τ be the local update step size of federated learning. It is for e i The expected value of g(G,τ,λμ,σ) is given as:

[0093]

[0094] Step 4: Node Selection and Channel Resource Allocation Scheme for Federated Learning under Unreliable Links: The performance optimization problem in Step 2.2 is transformed into a bipartite graph matching problem, and a node selection and channel resource allocation scheme is provided for the base station in the federated learning paradigm before the start of each round. The preferred scheme includes the following steps:

[0095] Step 4.1: Based on the theoretical convergence boundary in Step 3.2, the relevant parameters affecting the convergence speed and performance of federated learning are obtained as follows: Therefore, the optimization problem in step 2.2 is transformed into:

[0096]

[0097] stC1-C5

[0098] Optimization Objective This ensures that the model can converge to the target accuracy more quickly during training;

[0099] Step 4.2: Based on the optimization problem in Step 4.1, the solution process is considered as establishing a one-to-one relationship between each node participating in model uploading and the available sub-channels. Therefore, the optimization problem is transformed into a bipartite graph matching problem. First, a bipartite graph is constructed. Where θ is the set of edges between nodes and sub-channels, θ i,n ∈θ is an edge between node i and sub-channel l that satisfies condition θ i,l ∈{0,1},θ i,l=1 indicates that node i used sub-channel l to upload local model parameters. Therefore, θ i,l The weights are assigned as follows:

[0100]

[0101] After determining the weight of each edge in the bipartite graph, the Hungarian algorithm can be used to obtain the set of nodes that achieve the maximum matching and the matching relationship between nodes and sub-channels. At this point, the base station will select nodes from the node set in each round, upload the local federated learning model using the corresponding sub-channel, and aggregate the parameters that are not affected by noise.

[0102] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When the program is executed, it can implement the steps of the method. The storage medium may be, for example, ROM / RAM, magnetic disk, optical disk, etc.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A robust federated learning method, characterized in that: Includes the following steps: S1: Construct federated learning algorithms that can reduce model performance loss caused by data heterogeneity between nodes, as well as communication and computation models between edge networks with unreliable communication links; S2: Based on the analysis results of robust federated learning and channel conditions, improve the performance of the federated learning model after training to meet the requirements of deployment in resource-limited edge networks, and construct a federated learning performance optimization problem. S3: Derive the theoretical convergence edge of the federated learning algorithm under the influence of unreliable communication links under mathematical assumptions; S4: Transform the federated learning performance optimization problem into a bipartite graph maximum matching problem, and use the Hungarian algorithm to find the best node selection and channel resource allocation strategy in each round of training. Step S1 specifically includes the following steps: S11: The federated learning framework is a global federated learning model trained collaboratively by a central server and multiple edge node devices. In this process, nodes only transmit local model parameters to the central server, and all nodes do not communicate directly with each other. At the beginning of each training round, the central server distributes the global model to each node, and each node uses its local dataset to perform a local update process on the global model with a certain step size, resulting in different local models. After the nodes participating in federated learning send their local models to the central server, the central server performs weighted aggregation to obtain a new round of global model, and repeats the above steps until the target accuracy or preset training time is reached. S12: Defined within federated learning in edge networks, a base station has both a transceiver system and an aggregation system, can act as a central server, and has a set of nodes within its coverage area. , where nodes The dataset collected above is And nodes The solver used to find the optimal solution for the local model is: in For nodes In the The first generation The model parameters after the next local update. For nodes In the The first generation The model parameters after the next local update. For the first The parameters of the global federated learning model of the generation, To replace the loss function, where These are the coefficients of the proximal terms; S13: Node The latency required to train a local model is: in For nodes The number of bits required for the transmission model. For nodes GPU energy efficiency For nodes The calculation frequency, For nodes The number of GPU cycles required to process one bit of data; nodes The energy required to train a local model is S14: The communication resources between base stations and nodes mainly consist of available spectrum resources in the network, with a total bandwidth of [missing information]. and divided into There are 3 sub-channels of the same size, and the bandwidth of each sub-channel is: The available set of sub-channels is represented as ;node After the local update process is completed, the uplink rate at which the parameters of the local federated learning model are transmitted to the base station is: in, For channel allocation decisions, These are the corresponding decision-making indicators. Represents a node The model parameters will be transmitted through the channel. Upload to the central server, otherwise Additional conditions This is used to restrict a single subchannel to be used only by a single node within the FL framework; For nodes Uplink transmission power, For the corresponding uplink channel gain, Use uplink channels for nodes in other regions The resulting interference Power spectral density of channel noise; node The uplink transmission delay between the base station and the base station is: S15: Based on OFDMA technology and assuming the channel is a quasi-static Rayleigh fading channel, the probability that each local federated learning model is affected by noise during uplink transmission is: in For nodes In sub-channel PER on The signal quality threshold for a quasi-static Rayleigh fading channel. This represents the average signal-to-noise ratio.

2. The robust federated learning method according to claim 1, characterized in that: Step S2 specifically includes: S21: Taking into account the process in step S15 Then, the global aggregation process in federated learning was modified as follows: in Whether the base station allows the node The hallmark of participation in the federal learning process. Represents the base station's allowed nodes Upload local model parameters; Whether to aggregate nodes The model's logo Representative node The local model parameters are not affected by noise, allowing them to participate in aggregation; S22: The optimization problem of minimizing empirical risk is: in Choose a vector for the node. Let be the average loss function across all node data; This means that every node participating in the aggregation must be ensured. With assignable sub-channels The relationships between them are one-to-one; This indicates that nodes participating in training must meet latency requirements, where The maximum latency for completing the federated learning algorithm in each round; This indicates that nodes participating in training must meet energy consumption requirements, among which This represents the maximum energy consumption for completing the federated learning algorithm in each round.

3. The robust federated learning method according to claim 2, characterized in that: Step S3 specifically includes: S31: Let Satisfying nonconvexity and - Smooth, and the gradients of the local loss function and the gradients between the local and global loss functions are bounded, while the mini-batch gradient descent on the local substitution loss function is bounded. S32: Based on the assumptions of step S31, derive... The theoretical convergence boundary.

4. The robust federated learning method according to claim 3, characterized in that: Step S4 specifically includes: S41: Based on the theoretical convergence boundary in step S32, the relevant parameters affecting the convergence speed and performance of federated learning are obtained as follows: Therefore, the optimization problem in step S22 is transformed into: Optimization Objective This ensures that the model can converge to the target accuracy more quickly during training; S42: Based on the optimization problem in step S41, the solution process of this problem is regarded as: establishing a one-to-one relationship between each node participating in model uploading and the available sub-channels; The optimization problem is transformed into a bipartite graph matching problem. First, a bipartite graph is constructed. ,in Let be the set of edges between nodes and sub-channels. For nodes and sub-channel The edges satisfy the condition , Indicate node Sub-channels were used Upload local model parameters; therefore, for The weights are assigned as follows: After determining the weight of each edge in the bipartite graph, the Hungarian algorithm is used to obtain the set of nodes that achieve the maximum matching and the matching relationship between nodes and sub-channels. At this point, the base station will select nodes in the node set in each round to upload the local federated learning model using the corresponding sub-channels and aggregate the parameters that are not affected by noise.

Images