A wireless federated learning method based on gradient hierarchical transmission
By separating gradient information into sign and absolute value and constructing a single-step convergence performance model for resource optimization, the problem of insufficient reliability of gradient information transmission in wireless federated learning is solved, thereby improving the convergence stability and efficiency of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-05
AI Technical Summary
Existing wireless federated learning suffers from insufficient reliability in gradient information transmission under conditions of limited wireless resources, leading to a decline in model convergence performance. Furthermore, traditional methods fail to effectively utilize the difference between gradient sign and absolute value in model convergence, resulting in inefficient resource allocation.
Gradient information is separated into symbolic information and absolute value information. A single-step convergence performance model is constructed. Wireless resources are allocated through joint optimization to ensure reliable transmission of key information. Gradient reconstruction and aggregation are performed at the receiving end.
It significantly improves the training stability and convergence efficiency of wireless federated learning, and enables efficient model updates in resource-constrained environments.
Smart Images
Figure CN122160834A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wireless communication technology, and more specifically to a wireless federated learning method based on gradient hierarchical transmission. Background Technology
[0002] Existing research on wireless federated learning often employs methods such as gradient compression, quantization, or device selection to reduce communication overhead. However, most of these methods treat gradient information as a whole, failing to fully consider the difference in importance between gradient sign and absolute value in model convergence. In fact, the gradient sign determines the direction of neural network parameter updates and has a dominant influence on convergence stability, while the absolute value of the gradient primarily affects convergence accuracy and speed. Under unreliable wireless link conditions, if the sign information is mistransmitted, it will lead to incorrect model update direction, severely impacting system convergence performance. Therefore, how to prioritize the reliable transmission of critical gradient information in resource-constrained wireless environments has become a bottleneck problem in improving the efficiency and stability of wireless federated learning.
[0003] Furthermore, due to differences in local data scale and statistical distribution, the contributions of gradients from different devices to global model convergence are significantly uneven. Traditional averaging aggregation mechanisms do not consider this difference in importance, leading to the inability to achieve optimal allocation of limited wireless resources. Therefore, it is necessary to construct a unified representation model oriented towards convergence performance, explicitly linking the importance of gradients from different devices to training performance, thereby achieving synergistic optimization of communication and learning performance. Summary of the Invention
[0004] Purpose of the invention: The purpose of this invention is to provide a wireless federated learning method based on gradient hierarchical transmission, so as to solve the problem of degraded model convergence performance caused by insufficient reliability of gradient information transmission and low efficiency of wireless resource allocation in wireless resource-constrained scenarios.
[0005] Technical Solution: The wireless federated learning method based on gradient hierarchical transmission described in this invention is applied to a wireless network including a parameter server and multiple distributed devices, and includes the following steps:
[0006] Step 1: In each training round, each distributed device calculates the local gradient and splits the local gradient into sign information to represent the direction of parameter update and absolute value information to represent the magnitude of parameter update.
[0007] Step 2: The parameter server performs joint optimization allocation of the first communication resource used to transmit symbolic information and the second communication resource used to transmit absolute value information based on the pre-built single-step convergence performance model. The single-step convergence performance model is used to represent the correlation between the transmission reliability of different types of information and the reduction in model training loss.
[0008] Step 3: Based on the results of the joint optimization allocation, each distributed device transmits its symbol information and absolute value information to the parameter server via a wireless link.
[0009] Step 4: The parameter server reconstructs the local gradients of each distributed device based on the received sign and absolute value information, and aggregates all the reconstructed gradients to update the global model parameters.
[0010] Furthermore, in step 1, the gradient is separated into sign information and absolute value information as follows: each device quantizes its local gradient and explicitly splits the quantized gradient into an independent sign data packet and an independent absolute value data packet, wherein the sign data packet contains only the sign bit of each element of the gradient, and the absolute value data packet contains the quantized magnitude of each element of the gradient.
[0011] Furthermore, in step 2, the specific process of constructing the single-step convergence performance model is as follows: by analyzing the impact of different combinations of success or failure of symbolic information and absolute value information in wireless transmission on model updates, the model is constructed. The expected decrease of the loss function in a single round of training is expressed as a function of the probability of successful transmission of symbolic information, the probability of successful transmission of absolute value information, and the local gradient norm of each device.
[0012] Furthermore, in step 2, the joint optimization process is as follows: Based on the single-step convergence performance model, with the goal of maximizing the expected decrease in the training loss in a single round, a joint optimization problem of resource allocation is constructed. The decision variables of the problem include the ratio of transmit power allocation between symbol information and absolute value information within each device, and the ratio of system bandwidth allocation among multiple devices; the joint optimization problem is solved using an alternating iterative strategy.
[0013] Furthermore, the alternating iterative strategy for solving the joint optimization problem is as follows: under the condition of fixed bandwidth allocation among devices, the optimal power allocation ratio within each device is solved using the Newton-Raphson method; under the condition of fixed power allocation within devices, the optimal bandwidth allocation ratio among multiple devices is solved using the SCA method; by iterating the above two solution processes repeatedly until convergence, a joint resource allocation scheme that balances the reliability of symbol information transmission and overall convergence efficiency is obtained.
[0014] Furthermore, in step 4, the local gradient reconstruction for each device is performed as follows: The parameter server first determines whether the symbol data packet for each device has been correctly received, following the symbol priority principle; only when the symbol information is correctly received is the gradient direction of the device retained; if the absolute value information is also correctly received, provided that the symbol information is correct, the gradient magnitude is reconstructed; if the absolute value information transmission fails, the preset reference absolute value compensation is used to reconstruct the gradient magnitude; if the symbol information transmission fails, all gradient contributions of that device in this round are directly discarded.
[0015] The wireless federated learning system based on gradient hierarchical transmission described in this invention includes:
[0016] The separation module is used to calculate the local gradient of each distributed device in each training round and split the local gradient into sign information to represent the direction of parameter update and absolute value information to represent the magnitude of parameter update.
[0017] Joint optimization module: This module is used by the parameter server to jointly optimize the allocation of the first communication resource for transmitting symbolic information and the second communication resource for transmitting absolute value information based on the pre-built single-step convergence performance model. The single-step convergence performance model is used to represent the correlation between the transmission reliability of different types of information and the reduction in model training loss.
[0018] Transmission module: Used by each distributed device to transmit its own symbol information and absolute value information to the parameter server via a wireless link according to the results of joint optimization allocation;
[0019] Reconstruction module: The parameter server reconstructs the local gradients of each distributed device based on the received symbol and absolute value information, and aggregates all reconstructed gradients to update the global model parameters.
[0020] An electronic device according to the present invention includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program as described in any one of 1-6.
[0021] The present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, describes the method described in any one of 1-6.
[0022] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: The present invention achieves accurate characterization of the differences in gradient information of different categories and the contributions of different devices by constructing a hierarchical transmission mechanism with symbol-absolute value separation and a single-step convergence performance characterization model. On this basis, it performs joint allocation optimization of power and bandwidth, thereby prioritizing the reliable transmission of key gradient information in resource-constrained wireless environments, and significantly improving the stability, convergence efficiency and overall performance of the federated learning training process. Attached Figure Description
[0023] Figure 1 This is a flowchart of the present invention;
[0024] Figure 2 This is a comparison chart of the convergence performance of federated learning under different transmission schemes of the present invention. Detailed Implementation
[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0026] like Figure 1 As shown, this embodiment of the invention provides a wireless federated learning method based on gradient hierarchical transmission, including the following steps:
[0027] Step S1: At the start of each round of federated learning training, each distributed device calculates its local gradient vector based on the current global model parameters and local data, and uploads the gradient norm to the PS for subsequent wireless resource allocation optimization. Simultaneously, the device separates the sign and absolute value of the elements in the gradient vector, encapsulating the gradient sign vector into a sign data packet and quantizing the gradient absolute value vector before encapsulating it into an absolute value data packet. This constructs a system consisting of one PS and... A wireless federated learning system composed of distributed devices. In the... During the training rounds, each device Based on the current global model Calculate local gradient and the gradient norm Uploaded to PS for subsequent wireless resource allocation optimization. Meanwhile, the device uses... A random quantizer with 1 quantized bit right Quantization is performed to obtain the quantization gradient. To ensure the reliability of gradient symbol transmission, symbol vector Encapsulated as a symbolic data packet, The absolute values of each element in the vector form an absolute value vector. It is encapsulated as an absolute value data packet.
[0028] Specifically, for any element in The quantification results are as follows:
[0029]
[0030] in Represents absolute value. exist Uniform distribution within the interval, represented as
[0031]
[0032] and and They represent Find the minimum and maximum absolute values of all elements in the set.
[0033] Step S2: Based on the mechanism of separating the sign and absolute values of local gradients, a federated learning single-step convergence performance representation model is established. This model establishes a functional relationship between the expected decrease of the training loss function and the success probability of symbolic data packet transmission, the success probability of absolute data packet transmission, and the device gradient norm. A mathematical mapping relationship is constructed to the importance of different types of data packets and different devices to the global model convergence. To characterize the differences in the importance of different types of data packets and different devices to the model convergence performance, a single-step convergence performance representation model is constructed, expressing the decrease in expected loss per round as a functional relationship between the success probability of symbolic data packet transmission, the success probability of absolute data packet transmission, and the gradient norm.
[0034] Specifically, in - Under smooth conditions, the loss function decreases in a single round of training while satisfying the following conditions:
[0035]
[0036] in For learning rate, This is the aggregated gradient estimate obtained by PS based on the received gradient information. (Right-hand side of the inequality) and The values of the symbol data packets, the probability of successful transmission of absolute value data packets, and the device's local gradient norm are related. Related. Let. Indicates device In the The probability of successful transmission of round symbol data packets. Let represent the probability of successful transmission of an absolute data packet. Then, the expected upper bound of the decrease in the single-round loss function can be expressed as:
[0037]
[0038]
[0039] in , For Hadama accumulation. The global gradient norm of the previous round, For the predefined absolute value vector norm, This is the upper bound of the norm difference between the local and global gradients. This represents the upper bound of the quantization error.
[0040] Step S3: Based on the single-step convergence performance characterization model, construct an importance-driven joint optimization problem for wireless resources. With the goal of maximizing the expected decrease in single-round training loss, the transmit power allocation between symbolic data packets and absolute data packets within the device and the wireless bandwidth allocation among multiple devices are used as joint optimization variables. Under fixed bandwidth allocation conditions, the optimal power allocation ratio within the device is solved using the Newton-Raphson method, and the optimal bandwidth allocation ratio among multiple devices is solved using the SCA method. An alternating iterative strategy is used to solve the power allocation and bandwidth allocation problems separately.
[0041] To maximize single-step convergence performance, an importance-driven joint optimization problem for wireless resources is constructed based on the expected upper bound of the single-round loss function descent. Power allocation between symbolic data packets and absolute data packets within a device, as well as bandwidth allocation among multiple devices, are used as joint optimization variables.
[0042] Specifically, the equipment In the Probability of successful transmission of symbolic data packets of the wheel Probability of successful transmission of absolute value data packets This is related to the corresponding power and bandwidth allocation. Let... and Representing the equipment The transmission power allocated to symbolic data packets and absolute value data packets, Indicates allocation to device The bandwidth resources then satisfy
[0043]
[0044]
[0045]
[0046] Based on the expected upper bound expression of the single-round loss function descent... Construct the following optimization problem:
[0047]
[0048]
[0049]
[0050]
[0051] in This represents the set of device serial numbers. During the optimization process, an alternating iterative method is used to decompose and solve the power allocation strategy and the bandwidth allocation strategy. In fixed bandwidth allocation... Under the given conditions, the optimal power allocation ratio between symbolic data packets and absolute data packets within the device is solved using the Newton-Raphson method. In fixed power distribution Under the given conditions, the optimal bandwidth allocation ratio among multiple devices is solved using the SCA method. The above alternating optimization process is iterated repeatedly until convergence, yielding a joint optimization solution.
[0052] Step S4: PS determines whether to retain the gradient information of the corresponding device based on the received symbol data packet. When both the symbol data packet and the absolute value data packet are received correctly, gradient reconstruction is performed based on the absolute value data packet. When the symbol data packet is received correctly but the absolute value data packet is received incorrectly, a preset reference absolute value is introduced for compensation reconstruction, and the global aggregated gradient estimate is obtained through normalization correction to complete the model parameter update.
[0053] At the receiving end, PS performs joint decision-making and gradient reconstruction on the received symbol data packets and absolute value data packets according to the gradient symbol priority principle. It decides whether to retain the gradient of the corresponding device based on the correctness of the symbol data packets, and uses the reference absolute value for compensation estimation when the absolute value data packets cannot be reliably recovered, thereby obtaining the aggregated gradient estimate for global model update.
[0054] Specifically, in the first During the training rounds, regarding the equipment When gradient sign data Upon successful reception, the PS first retains the gradient direction information of the device. If the absolute value data... If both are successfully received, then on the PS end, the device... The gradient is reconstructed as
[0055]
[0056] If symbolic data Successfully received absolute value data If reception fails, a preset reference absolute value vector will be used. To receive compensation
[0057]
[0058] in This is the absolute value vector of the previous round's global model, initially set to 0. When symbolic data... When a reception fails, the gradient contribution of that device in the current round is discarded because an incorrect gradient descent direction has a significant impact on model convergence. After completing gradient reconstruction, PS aggregates all reconstructed gradients to obtain the global gradient estimate, i.e.
[0059]
[0060] in This is used to mitigate the impact of gradient discarding on the unbiasedness of global gradient estimation. After completing global gradient estimation, based on the learning rate... Update global model parameters:
[0061] .
[0062] The technical terms involved in this invention are explained as follows:
[0063] SCA: Round-by-Round Convex Approximation
[0064] PS: Parameter server
[0065] Non-IID: Non-independent and identically distributed
[0066] DDS: Direct Disposal Solution
[0067] In the context of limited wireless resources making it difficult to balance convergence efficiency and transmission reliability, this invention constructs a parameter update mechanism that separates gradient signs and absolute values. The quantized gradient is split into symbolic and absolute value data packets for independent transmission, with the gradient norm used as a reference for device importance. Next, a single-step convergence performance model is established, linking the expected loss reduction with the probability of symbolic success, the probability of absolute value success, and device importance. Then, aiming to maximize the expected loss reduction in a single round, a power and bandwidth joint optimization model is constructed, employing alternating solutions of the Newton-Raphson and successive convex approximation methods to achieve importance-driven hierarchical resource allocation. Finally, at the receiving end, gradient reconstruction and aggregation updates are performed based on the sign-first principle, thereby improving the convergence stability and resource utilization efficiency of wireless federated learning.
[0068] To verify the effectiveness of the present invention, a simulation experiment was conducted. The parameters involved in the simulation experiment are shown in the table below:
[0069] Table 1 Simulation Experiment Parameter Table
[0070] like Figure 2As shown, this invention introduces a gradient sign-absolute value separation transmission mechanism and combines it with an importance-driven resource allocation strategy, achieving faster convergence and better performance under conditions of limited wireless resources. The baselines for comparison include: an ideal scenario with error-free gradient transmission; a scheduling scheme that alleviates resource pressure solely through user selection; a DDS scheme that directly discards transmitted erroneous gradients; and a 1-bit transmission scheme that only transmits gradient sign information. In contrast, this invention achieves refined resource allocation while preserving critical gradient information, resulting in faster convergence and performance closer to the ideal scenario under communication-constrained conditions.
Claims
1. A wireless federated learning method based on gradient hierarchical transmission, applied to a wireless network including a parameter server and multiple distributed devices, characterized in that, Includes the following steps: Step 1: In each training round, each distributed device calculates the local gradient and splits the local gradient into sign information to represent the direction of parameter update and absolute value information to represent the magnitude of parameter update. Step 2: The parameter server performs joint optimization allocation of the first communication resource used to transmit symbolic information and the second communication resource used to transmit absolute value information based on the pre-built single-step convergence performance model. The single-step convergence performance model is used to represent the correlation between the transmission reliability of different types of information and the reduction in model training loss. Step 3: Based on the results of the joint optimization allocation, each distributed device transmits its symbol information and absolute value information to the parameter server via a wireless link. Step 4: The parameter server reconstructs the local gradients of each distributed device based on the received sign and absolute value information, and aggregates all the reconstructed gradients to update the global model parameters.
2. The wireless federated learning method based on gradient hierarchical transmission according to claim 1, characterized in that, In step 1, the gradient is separated into sign information and absolute value information as follows: Each device quantizes its local gradient and explicitly splits the quantized gradient into an independent sign data packet and an independent absolute value data packet. The sign data packet contains only the sign bit of each element of the gradient, and the absolute value data packet contains the quantized magnitude of each element of the gradient.
3. The wireless federated learning method based on gradient hierarchical transmission according to claim 1, characterized in that, In step 2, the specific process of constructing the single-step convergence performance model is as follows: By analyzing the impact of different combinations of success or failure of symbolic information and absolute value information in wireless transmission on the model update, the model is constructed. The expected decrease of the loss function in a single round of training is expressed as a function of the probability of successful transmission of symbolic information, the probability of successful transmission of absolute value information, and the local gradient norm of each device.
4. The wireless federated learning method based on gradient hierarchical transmission according to claim 1, characterized in that, In step 2, the joint optimization process is as follows: Based on the single-step convergence performance model, with the goal of maximizing the expected decrease in the training loss in a single round, a joint optimization problem of resource allocation is constructed. The decision variables of the problem include the ratio of transmit power allocation between symbol information and absolute value information within each device, and the ratio of system bandwidth allocation among multiple devices; the joint optimization problem is solved using an alternating iterative strategy.
5. The wireless federated learning method based on gradient hierarchical transmission according to claim 4, characterized in that, The alternating iterative strategy for solving the joint optimization problem is as follows: under the condition of fixed bandwidth allocation among devices, the optimal power allocation ratio within each device is solved using the Newton-Raphson method; under the condition of fixed power allocation within devices, the optimal bandwidth allocation ratio among multiple devices is solved using the SCA method; by iterating the above two solution processes repeatedly until convergence, a joint resource allocation scheme that balances the reliability of symbol information transmission and the overall convergence efficiency is obtained.
6. The wireless federated learning method based on gradient hierarchical transmission according to claim 1, characterized in that, In step 4, the local gradient reconstruction for each device is performed as follows: The parameter server first determines whether the symbol data packet for each device has been correctly received, following the symbol priority principle; only when the symbol information is correctly received is the gradient direction of the device retained; if the absolute value information is also correctly received, the gradient magnitude is reconstructed, provided that the symbol information is correct; if the absolute value information transmission fails, the preset reference absolute value compensation is used to reconstruct the gradient magnitude; if the symbol information transmission fails, all gradient contributions of that device in this round are directly discarded.
7. A wireless federated learning system based on gradient hierarchical transmission, characterized in that, include: Separate modules; In each training epoch, each distributed device calculates its local gradient and splits the local gradient into sign information representing the direction of parameter update and absolute value information representing the magnitude of parameter update. Joint optimization module: This module is used by the parameter server to jointly optimize the allocation of the first communication resource for transmitting symbolic information and the second communication resource for transmitting absolute value information based on the pre-built single-step convergence performance model. The single-step convergence performance model is used to represent the correlation between the transmission reliability of different types of information and the reduction in model training loss. Transmission module: Used by each distributed device to transmit its own symbol information and absolute value information to the parameter server via a wireless link according to the results of joint optimization allocation; Reconstruction module: The parameter server reconstructs the local gradients of each distributed device based on the received symbol and absolute value information, and aggregates all reconstructed gradients to update the global model parameters.
8. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method according to any one of claims 1-6.
9. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the method described in any one of claims 1-6.