A federated learning model aggregation method based on over-the-air computation and energy harvesting

Abstract

The application provides a federated learning model aggregation method based on aerial computing and energy collection, proposes a model aggregation design optimization algorithm using aerial computing technology and energy collection technology, considers the correlation of device energy in different training rounds, realizes reasonable scheduling of limited energy in the device by alternately optimizing the participating device, the transmission device transmission power and the beamforming vector of the receiving end, improves the model aggregation precision, and further improves the model test precision. The scheme provided by the application is an online optimization method, which is different from the existing offline optimization method. The design of the application in each communication round only depends on the current channel state information and the arrival energy, and is more suitable for actual application scenarios.

Description

TECHNICAL FIELD

[0001] The application relates to a model aggregation method in wireless federated learning and belongs to the technical field of information. BACKGROUND

[0002] With the popularization of various emerging intelligent applications, frequent data exchange between edge servers and devices will cause huge communication overhead and serious privacy leakage problems. Federated learning is a distributed machine learning scheme that only exchanges training models or gradient information, which can effectively reduce communication overhead and avoid privacy leakage. Federated learning includes the following application scenarios:

[0003] 1) Personalized recommendation on mobile devices: Federated learning can train personalized recommendation models on multiple mobile devices, and the data in these devices does not need to be uploaded to the cloud, thereby protecting the privacy of users.

[0004] 2) Medical health field: Federated learning can train medical image classification and recognition models among multiple hospitals and research institutions, improve the accuracy of the model, and protect the privacy information of patients.

[0005] 3) Internet of Things: Federated learning can perform model collaboration and optimization among multiple intelligent devices, sensors and monitors, thereby improving the reliability and efficiency of the system.

[0006] 4) Financial industry: Federated learning can jointly learn transaction behavior patterns among multiple banks, thereby reducing fraud and risk.

[0007] 5) Edge computing: Federated learning can be performed on edge devices, thereby saving network bandwidth and reducing delay, improving computing efficiency and privacy protection.

[0008] In the wireless communication networking scenario, when the number of edge devices is large, the information amount of high-dimensional models and gradient information is large, and the information transmission rate from the device to the edge server is limited by the traditional wireless orthogonal multiple access method, which is easy to cause large communication delay, thereby affecting the efficiency of federated learning. In order to reduce the communication delay of the model aggregation process, air computing technology can be introduced to utilize the superposition of radio waves to realize fast model aggregation. In addition, due to the limited energy of the device itself, periodic local training and model uploading will generate a large amount of device energy consumption, which may cause distortion of the aggregated signal, even stop training, and thus affect the training effect of federated learning. SUMMARY

[0009] The purpose of the present application is to solve the energy limited dilemma of the wireless federated learning system.

[0010] In order to achieve the above objectives, the technical solution of the present invention is to provide a federated learning model aggregation method based on air computing and energy harvesting, which is characterized by comprising the following steps:

[0011] Step 1: Based on the device energy update model, Transmit power Receiver beamforming vector m t Perform joint design and construct the specific optimization problem as shown below:

[0012]

[0013]

[0014]

[0015]

[0016] Where, is the battery capacity of device k in round t, and are the energy costs of device k for communication and local training in round t, is the energy collected by device k in round t, is the expectation of the transmitted signal and the received noise, F(ω T ) is the global loss function after T rounds of training, ω T is the global model after T rounds of training, is the maximum transmit power constraint of device k;

[0017] Step 2: Virtual Queue To record the energy level of device k, perform online optimization on the optimization problem constructed in step 1. During the optimization, the participating devices are fixed. The optimization objective function of each round is:

[0018]

[0019] in: is the standard deviation of the gradient of device k; d is the dimension of the model; is the channel state information between device k and the server; is the variance of Gaussian white noise; U t =(1-γμ) T-1-t ,γ is the learning rate of federated learning, μ>0 is Conditional constant; (·) H is the conjugate transpose of the vector; τ is the duration of each round of training;

[0020] Fix the device transmit power and solve the optimal receiving end beamforming vector (m t ) * :

[0021]

[0022] Where, I M is the identity matrix of dimension M.

[0023] Fix the receiving end beamforming vector and solve for the optimal transmit power

[0024]

[0025] Where,

[0026] Update the virtual queue using the following formula:

[0027]

[0028] Step 3: Determine the participating equipment in each round through Gibbs sampling method

[0029] Step 4: Repeat steps 2 and 3 until the global loss function value of federated learning converges.

[0030] Preferably, in step 1, the established device energy update model is expressed as:

[0031] Preferably, in step 2, the optimal transmission power is not satisfied. When the power is corrected to

[0032] Preferably, in step 2, the transmission scalar is expressed as Where, and m t is the optimal transmit power and the optimal receiver beamforming vector on the server side.

[0033] Preferably, the step 3 comprises the following steps:

[0034] vector For collection The indicator variable, is a binary indicator variable for the device K, The indicator variable s t Perform J max times sampling and obtain a series of samples {s t,j} until convergence, where j∈{1,…,J max}; In each sampling, sample s t,j From the collection Distribution in China and Israel Sampling is performed, where Indicates that t,j-1 The indicator variable for the different i-th element in J(s t,j ) is the optimization problem constructed in step 2 in s t,j The objective function under , parameter β>0 is used to accelerate convergence.

[0035] This paper proposes a model aggregation design optimization algorithm that utilizes air computing technology and energy harvesting technology. It considers the correlation of device energy within different training rounds and realizes the reasonable scheduling of limited energy in the device by alternately optimizing the transmission power of participating devices, transmission equipment, and the beamforming vector of the receiving end, thereby improving the accuracy of model aggregation and thus enhancing the accuracy of model testing.

[0036] The solution proposed in the present invention is an online optimization method. Unlike existing offline optimization methods, the design of the present invention in each communication round only relies on the current channel state information and arrival energy, which is more in line with actual application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of the present invention;

[0038] Figure 2 The performance comparison between the present invention and the existing solutions is illustrated. DETAILED DESCRIPTION

[0039] Below in conjunction with specific embodiment, further set forth the present invention.Should be understood that these embodiments are only used to illustrate the present invention and are not used in limiting the scope of the present invention.In addition, should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall equally within the scope limited by the appended claims of the application.

[0040] like Figure 1 As shown, the present invention discloses a federated learning model aggregation method based on air computing and energy harvesting, which specifically includes the following steps:

[0041] Step 1: Establish the device energy update model as shown below:

[0042]

[0043] Where, is the battery capacity of device k in round t, and are the energy costs of device k for communication and local training in round t, is the energy collected by device k in round t. To minimize the expectation of the global loss function after T rounds of training, the present invention performs Transmit power Receiver beamforming vector m t Perform joint design and construct the specific optimization problem as shown below:

[0044]

[0045]

[0046]

[0047]

[0048] Where, is the expectation of the transmitted signal and the received noise, F(ω T ) is the global loss function after T rounds of training, ω T is the global model after T rounds of training, is the maximum transmit power constraint of device k.

[0049] Step 2: By introducing a virtual queue To record the energy level of device k, perform online optimization on the optimization problem constructed in step 1.

[0050] When fixed participating equipment The optimization objective function for each round is:

[0051]

[0052] in: is the standard deviation of the gradient of device k; d is the dimension of the model; is the channel state information between device k and the server; is the variance of Gaussian white noise; U t =(1-γμ) T-1-t ,γ is the learning rate of federated learning, μ>0 is Conditional constant; (·) H is the conjugate transpose of the vector; τ is the duration of each round of training. Fixed device transmit power, the optimal receiving end beamforming vector (m t ) · Expressed as:

[0053]

[0054] Where, I M is the identity matrix of dimension M.

[0055] Fixed receive beamforming vector, optimal transmit power is denoted as:

[0056]

[0057] where,

[0058] When the transmit power does not satisfy , the power can be corrected as

[0059] The transmission scalar can be denoted as

[0060] So far, the virtual queue is updated as follows

[0061] Step 3, when the fixed transmit power and receive beamforming vector, the participating devices in each round can be determined by the Gibbs sampling method Specifically, the vector is an indicator variable of the set , and is a binary indicator variable of device K, and s t is sampled J max times and a series of samples {s t,j} is obtained until convergence, where j∈{1,…,J max}. In each sampling, the sample s t,j is sampled from the set with distribution , where denotes an indicator variable different from the i-th element in s t,j-1 , J(s t,j ) is the objective function of the optimization problem constructed in step 2 under s t,j , and the parameter β>0 is used to accelerate convergence.

[0062] Step 4: repeat steps 2 and 3 until the federal learning global loss function value converges.

[0063] The embodiment is implemented on the premise of the technical solution of the application, and gives a detailed implementation manner and specific operation process. Specifically, Figure 2 is the performance comparison of the method proposed in the application and the existing method on the MNIST data set with the same simulation parameter settings when N=4, K=5, and V=1. Among them, we assume that the environmental energy arrives in the form of energy blocks, and the number of arrivals in each round of training obeys the Poisson process with parameter 2, and each energy block size obeys the uniform distribution of [0, 1]J. The transmission time of each round is τ=0.01s, and the maximum capacity of the battery E maxPmax(k) = 10J, maximum transmit power of device k In the figure, the comparative scheme 1 performs local training and transmission with all energy in each round of training, and the comparative scheme 2 divides the existing energy by the expected training number to ensure that each subsequent round of training can be normally performed. Under the same simulation parameter setting, the higher the test accuracy, the better the performance. As can be seen from the figure, to achieve the same test accuracy, the method disclosed by the present application will adopt fewer training rounds, and thus has higher energy efficiency.

Claims

1. A federated learning model aggregation method based on over-the-air computation and energy harvesting, characterized in that, Comprising the steps of: Step 1, updating the model based on the energy of the device transmit power receive beamforming vector m t Joint design is performed to construct a specific optimization problem as shown in the following formula: where, is the battery capacity of device k at round t, and are the energy cost of device k for communication and local training at round t, respectively, is the energy harvested by device k at round t, is the expectation of the transmitted signal and the received noise, F(ω T ) is the global loss function after T rounds of training, ω T is the global model after T rounds of training, is the maximum transmit power constraint of device k; Step 2, Virtual Queue The energy level of the recording device k is recorded, and the optimization problem constructed in step 1 is optimized on-line, and the participating devices are fixed during optimization The optimization objective function of each round is: where: is the standard deviation of the gradient for device k; d is the dimension of the model; is the channel state information between device k and the server; is the variance of the Gaussian white noise; U t = (1 - γμ) T-1-t γ is the learning rate of federated learning, μ > 0 is is a constant of the condition; (·) H is the conjugate transpose of a vector; τ is the length of each round of training; The fixed device transmit power is solved to obtain the optimal server-side receiving end beamforming vector (m t ) * : wherein I is a unit matrix of dimension M M is a unit matrix of dimension M; Fixed receive beamforming vector, solve the optimal transmit power In the formulae, Updating the virtual queue using the following formula: Step 3, determining participating devices per round by Gibbs sampling Step 4, repeat step 2 and step 3 until the federated learning global loss function value converges.

2. The federated learning model aggregation method based on over-the-air computation and energy harvesting of claim 1, wherein, In step 1, the established device energy update model is represented as:

3. The federated learning model aggregation method based on over-the-air computation and energy harvesting of claim 1, wherein, In Step 2, the optimal transmit power is solved to not satisfy the power is corrected to 4. The federated learning model aggregation method based on over-the-air computation and energy harvesting of claim 1, wherein, In step 2, the transmission scalar is denoted as 5. The federated learning model aggregation method based on over-the-air computation and energy harvesting of claim 1, wherein, The step 3 comprises the steps of: The step 3 comprises the steps of: Vector is an indicator variable for the set , is a binary indicator variable for the device K, and s is an indicator variable for the set t J max samples are taken and a sequence of samples is obtained until convergence, where j ∈ {1, …, J max}; in each sampling, a sample s t,j is sampled from the set with distribution , where denotes an indicator variable that is different from the i-th element in s t,j-1 , J(s t,j ) is the objective function of the optimization problem constructed in step 2 at s t,j , and the parameter β > 0 is used to accelerate convergence.

Images