Wireless communication method based on hybrid relay-intelligent reflection surface and federated learning

Abstract

The present invention relates to a wireless communication method based on a hybrid relay-intelligent reflector and federated learning. As a type of distributed machine learning, the core idea of ​​federated learning is to obtain an optimal global model by collaboratively training a learning model on an edge device, while avoiding the transmission of original data. Considering the inherent superposition characteristics of a wireless channel, air computing is introduced to support a large number of local devices uploading data. At the same time, some unsafe devices are screened out according to the confidentiality rate in physical layer security to obtain higher security. Since the hybrid relay-intelligent reflector can not only reflect but also amplify the incident signal, it can be used to enhance the gain of the uplink in the federated learning process to obtain better accuracy. The present invention develops a design method to alternately optimize device selection, hybrid relay-intelligent reflector amplification factor, phase shift, and receive beamforming vector to improve the accuracy and security of federated learning.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a wireless communication method based on hybrid relay-intelligent reflection surface and federated learning. Background Art

[0002] The traditional centralized learning method will bring huge communication, storage and computing burdens to the base station, and its security is questionable. In the face of the increasingly large amount of data, federated learning stands out with its distributed characteristics. First of all, as a type of distributed learning, federated learning enables multiple devices in the Internet of Things to collaboratively build a global model, and since it does not need to upload original data, the data is not easily stolen by eavesdroppers, so it has higher privacy and security. Secondly, federated learning can train each local model parameter in local model training, and then send the local model parameters to the base station. In this communication process, its overhead is significantly reduced compared to centralized learning, thereby reducing the burden of storage and computing.

[0003] In federated learning, the protection of the communication process is extremely important. In the communication process, the physical layer communication occupies an important position, but the physical layer is easy to be eavesdropped by eavesdroppers. Eavesdroppers can restore the complete information through the partial information they eavesdrop on. Therefore, sufficient attention must be paid to the security of the physical layer. Physical layer security technology is considered to be a promising tool that can prevent the transmission signal from being eavesdropped and obtain sufficient security by ensuring a certain confidentiality rate.

[0004] As a key enabler of the next generation network, the Intelligent Reconfigurable Surface can effectively reduce energy consumption and improve spectrum efficiency during signal transmission. The Intelligent Reconfigurable Surface can change the phase shift of the incident signal by controlling the reflection unit, thereby overcoming the unfavorable channel conditions during the propagation process. The hybrid relay-intelligent reflective surface combines the relay and the intelligent reflective surface. The relay can amplify the incident signal, while the intelligent reflective surface has a certain ability to cause the signal to produce a phase shift. Therefore, the hybrid relay-intelligent reflective surface can obtain better reflection performance, which is of great significance for the research on high efficiency and energy saving in wireless transmission systems. Summary of the invention

[0005] The object of the present invention is to provide a wireless communication method based on a hybrid relay-intelligent reflector surface and federated learning, so that the federated learning system can ensure sufficient accuracy and security of the device in the process of participating in the federated learning by jointly optimizing the device selection, the phase shift of the hybrid relay-intelligent reflector surface, the receiving beamforming vector and the amplification factor of the hybrid relay-intelligent reflector surface under the constraint of the transmission power.

[0006] To achieve the above-mentioned purpose, the technical scheme of the present invention is: a wireless communication method based on hybrid relay-intelligent reflection surface and federated learning, designing a physical layer security communication system based on hybrid relay-intelligent reflection surface and federated learning, including local equipment, base station, hybrid relay-intelligent reflection surface and eavesdropper; the local equipment serves as the transmitting end of the system, and the base station serves as the receiving end of the system. The transmitting end sends the signal to the base station in two ways, one way is for the local equipment to directly reach the base station, and the other way is for the local equipment to pass through the hybrid relay-intelligent reflection surface and then reach the base station; the eavesdropper eavesdrops on the signal sent by the local equipment in this process; by optimizing equipment selection, amplification factor of the hybrid relay-intelligent reflection surface, phase shift and receiving beamforming vector, the error of the transmission process is minimized while ensuring security, thereby improving the accuracy of federated learning;

[0007] The method comprises the following steps:

[0008] Step S1, constructing a communication model according to the mathematical expressions of hybrid relay-intelligent reflection surface and federated learning;

[0009] Step S2: Consider the model aggregation error generated in the process of federated learning, derive the upper limit of the learning loss in the iterative process, and thus construct it as an optimization problem;

[0010] Step S3, minimizing the upper limit of learning loss by alternately optimizing device selection, amplification factor of hybrid relay-intelligent reflection surface, phase shift and receiving beamforming vector: first, simplifying the upper limit of learning loss into a function expression that simultaneously contains device selection, amplification factor of hybrid relay-intelligent reflection surface, phase shift and receiving beamforming vector; then alternately optimizing the variables in this function expression to obtain the corresponding optimal value; in the communication process, by determining the optimal value, it will be possible to ensure that the system obtains sufficient accuracy and security.

[0011] In one embodiment of the present invention, in step S1, the constructed communication model can be expressed as:

[0012]

[0013] In the formula, y d represents the received signal at the base station, d represents the time slot, s d,m represents the transmission signal of the mth device in the dth time slot, M is the total number of devices, h UB,m is the channel gain vector of the mth device directly reaching the base station, and h H,m is the channel gain vector between the mth device and the hybrid relay-intelligent reflector, G is the channel gain vector between the hybrid relay-intelligent reflector and the base station, z H,mis the additive Gaussian white noise generated by the mth device when it passes through the hybrid relay-intelligent reflection surface, which has a mean of 0 and a variance of The complex Gaussian distribution of z BS is the additive Gaussian white noise generated at the base station, which has a mean of 0 and a variance of The complex Gaussian distribution of And meet σ 2 is the noise during device communication, I N represents the unit matrix of length N, Υ represents the reflection coefficient of the hybrid relay-intelligent reflector, Ψ represents the amplification factor of the hybrid relay-intelligent reflector, p m Represents the transmit power of the mth device;

[0014] The base station converts the received signal into:

[0015]

[0016] In one embodiment of the present invention, in step S2, the optimization problem constructed is as follows:

[0017]

[0018] subject to x m ∈{0, 1}, 1≤m≤M

[0019] |α l |=1,

[0020]

[0021] In the formula, x m is the device selection vector, K is the total data set of all devices, K m is the local data set of the mth device, f is the receive beamforming vector, the function variable α is taken from Υ, represents the specific value of the reflection coefficient of the hybrid relay-intelligent reflector, Λ represents the set of active elements in the hybrid relay-intelligent reflector, P0 is the upper limit of the device's transmit power, σ 2 is the noise in the device communication process, M is the total number of devices, and L is the size of the hybrid relay-intelligent reflective surface; in the above formula, the first constraint is to determine whether the device is selected to participate in federated learning, the second constraint represents the amplification factor limit for the passive components in the hybrid relay-intelligent reflective surface, and the third constraint represents the power limit of the hybrid relay-intelligent reflective surface; in the above formula, the first half represents the error caused by the selection of the device, and the second half represents the error generated in the communication process. The two parts together constitute the error generated in the model aggregation process of federated learning.

[0022] In one embodiment of the present invention, in step S3, the device selection method can be expressed as follows:

[0023] First, among local devices, some devices have poor channel conditions, which is manifested as small channel gain. Such devices will affect the overall learning process in federated learning, so they need to be excluded in device selection; secondly, some devices have poor confidentiality performance and are easily eavesdropped by eavesdroppers during signal transmission. Therefore, it is necessary to quantify the confidentiality performance of the device, that is, use the confidentiality rate as the judgment standard, retain the device when it is greater than or equal to a certain threshold, and exclude it when it is less than a certain threshold; the confidentiality rate is expressed as follows:

[0024]

[0025] In the formula, B represents the bandwidth allocated during signal transmission, h m (α) represents the channel gain from the device to the base station, h e (α) represents the channel gain from the device to the eavesdropper.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] (1) The present invention uses a hybrid relay-intelligent reflection surface to assist federated learning, and uses air computing technology to ensure the aggregation process of federated learning. The compatibility between air computing and the hybrid relay-intelligent reflection surface gives federated learning better efficiency and faster speed.

[0028] (2) The present invention considers the upper limit of learning loss, i.e., error, during the model aggregation process and improves accuracy by minimizing the error.

[0029] (3) The present invention takes into account the situation where there are eavesdroppers in the system, and uses the confidentiality rate as a criterion for judging security at the transmitting end to ensure secure transmission and improve the security of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a system structure diagram of a physical layer secure communication method based on hybrid relay-intelligent reflection surface and federated learning according to an example of the present invention.

[0031] Figure 2 This is an illustration of the federated learning part when the total number of devices in the example of the present invention is M.

[0032] Figure 3 This is a simulation diagram of the test accuracy of an example of the present invention when the number of devices M=40, the number of receiving antennas N=5, and the size of the hybrid relay-intelligent reflector L=40 under the Rice channel. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0034] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0035] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0036] like Figure 1 As shown, this example provides a physical layer secure communication method based on hybrid relay-intelligent reflection surface and federated learning. The wireless communication system under consideration specifically includes a device, a hybrid relay-intelligent reflection surface, a base station, and an eavesdropping device. When the signal is transmitted through the hybrid relay-intelligent reflection surface, it will cause phase shift and amplification of the signal. The receiving beamforming technology is used at the base station to enhance the received signal.

[0037] like Figure 2 In federated learning, first the device needs to perform model training locally based on its own data set to obtain the corresponding model parameters, such as gradients, weights in the neural network, etc. After that, each local device uploads its own model parameters to the central server, i.e., the base station. The base station then assigns a certain weight based on the amount of data of each local device, and weights it with the model parameters to finally obtain the global parameters. The base station broadcasts the global parameters to each device, and each device takes the global parameters into account in the loss function for the next round of training. Through continuous rounds of iteration, an optimal global model will be obtained. The update of the global model can be expressed as:

[0038]

[0039] Among them, w represents the global model parameters, t represents the number of iterations, λ is the learning rate, m is the number of devices participating in federated learning, and K m is the data set of the mth device, Indicates the parameters received at the base station.

[0040] In the iterative training task of the local dataset, the loss function can be defined as follows:

[0041]

[0042] In the formula, u i With v i are the sample vectors and corresponding labels in the device dataset respectively. The total loss function can be obtained by weighting the loss function of each device.

[0043]

[0044] For ease of representation, the channel coefficient vector from the mth device to the base station is defined as:

[0045]

[0046] Among them, h UB,m represents the channel coefficient from the device to the base station, G represents the channel coefficient from the hybrid relay-intelligent reflector to the base station, and h H,m represents the channel coefficient from the device to the hybrid relay-intelligent reflector surface, where α represents the reflection coefficient of the hybrid relay-intelligent reflection surface. Through the channel coefficient vector, it can be known that the channel coefficient from the device to the base station is the superposition of two signals from the device directly to the base station and the device to the base station through the hybrid relay-intelligent reflection surface. In the present invention, the number of antennas of the device is 1, the size of the hybrid relay-intelligent reflection surface is L, and the number of receiving antennas of the base station is N.

[0047] In this scenario, the present invention considers dividing the time of the Tth round of iteration into different time slots, using s d,m represents the transmission signal of the device in time slot d, y d represents the received signal of the corresponding time slot, z H,m is the additive Gaussian white noise generated by the mth device at the hybrid relay-intelligent reflection surface, which has a mean of 0 and a variance of The complex Gaussian distribution of z BS is the additive Gaussian white noise generated at the base station, which has a mean of 0 and a variance of The complex Gaussian distribution of And meet Therefore, the model formula from transmitting signal to receiving signal can be obtained:

[0048]

[0049] Since the transmitted signal will experience noise in the channel, it is necessary to specifically represent the signal emitted by the device. In the present invention, it is made into local gradient statistical information, that is:

[0050]

[0051] Among them, t is the iteration round, m is the number of devices, and w t is the global model of the tth round. Then g m,t Ignore the number of rounds in and decompose it into D terms.

[0052]

[0053]

[0054]

[0055] In the formula, g m,d is the local gradient information in the dth time slot. In the above formula, the gradient information obtained by local training is converted into a transmission signal, and then through the wireless channel, y is obtained. d , that is, the received signal of the base station. In the base station, according to the above expression, the receive beamforming processing is performed, and the processed expression is:

[0056]

[0057]

[0058] In the formula, is the received signal after beamforming. f is the receive beamforming vector. η is a normalized scalar, and satisfies η>0. By adding Enables the receiving end to obtain the original signal.

[0059] By The total gradient information in the tth round can be obtained by superimposing the signals of each time slot in the tth round. The local model parameters can be obtained through the gradient information, so that the average calculation of the model can be performed at the base station.

[0060] Next, by comparing the received signal at the receiving end under ideal conditions with the received signal at the receiving end after actually passing through the channel, the error of model aggregation during the communication process can be obtained.

[0061] Ideally, the received signal at the receiving end is:

[0062]

[0063] The actual received signal at the receiving end after passing through the channel is:

[0064]

[0065] Through the above two formulas, we can get the model aggregation error in the learning process:

[0066]

[0067] After substituting the formula into the formula, we can get the specific error value:

[0068]

[0069] In the above formula, D is the total number of time slots. At the same time, if the error value needs to be achieved, the transmit power equalization factor p m And the normalization factor η needs to satisfy the following conditions:

[0070]

[0071]

[0072] At the same time, if the device is selected before federated learning, errors caused by device selection will occur:

[0073]

[0074] In the above formula, β is used to ensure the boundedness of the local gradient vector.

[0075] Therefore, through the two error formulas, we can analyze the upper limit of the average difference between the global model parameters of the current round and the optimal global model parameters, that is:

[0076]

[0077] Where ω is the Lipschitz constant, which is used to ensure that the F function meets the Lipschitz continuity condition. * is the optimal global model parameter, w0 is the initial global model parameter, for the sake of simplicity, the δ function and χ function are expressed as follows:

[0078]

[0079]

[0080] In the formula, μ is a parameter that ensures the strong convexity of the F function. In the average difference upper limit, when the rounds are continuously iterated, the smaller its value is, the faster the convergence speed is and the closer it is to the optimal global model. Therefore, the expression after the global model is continuously iterated can be analyzed:

[0081]

[0082] It is not difficult to determine from the above formula that after continuous iterations, the upper limit of the average difference after the final convergence is determined by the δ function. Therefore, the δ function needs to be used as the optimization function.

[0083] Therefore, the optimized objective function can be obtained:

[0084]

[0085] subject to x m ∈{0, 1}, 1≤m≤M

[0086] |α l |=1,

[0087]

[0088] In the optimization problem above, And α l It can be divided into amplitude part and phase part, that is θ l ∈[π,2π]. It should be noted that in the hybrid relay-intelligent reflector, Λ represents the set of active components in the hybrid relay-intelligent reflector, where |α l |>1, L represents the total number of components of the hybrid relay-intelligent reflector; if it is a passive component part, then |α l |=1. In the above formula, represents the reflection coefficient of the total hybrid relay-intelligent reflector, and Υ=Φ+Ψ, which represents the reflection coefficients of passive components and active components respectively. The passive components are specifically represented by Φ=diag(φ1,...φ L ), where φ l =α l , among which The active element is specifically expressed as Ψ=diag(ψ1,...ψ L ), where ψ l =α l , among which α l ∈Λ. P in the constraint a (α) represents the transmission power of the active element of the hybrid relay-smart reflector and satisfies the following conditions:

[0089]

[0090] In the optimization problem of the above formula, the optimization function is obtained mainly by alternately optimizing f and α. The specific process is as follows:

[0091] First, when selecting fixed equipment, the following optimization problem can be obtained:

[0092]

[0093] subject to|α l |=1,

[0094]

[0095] At this point, the optimization problem can be expressed as a minimum-maximum problem. First, the maximum problem of the internal term can be calculated, that is:

[0096]

[0097] By traversing all devices, a maximum value can be found, and the device with the maximum value is named m0. Then the problem is transformed into the following:

[0098]

[0099] 1. Optimize f when α is given

[0100] During optimization, the optimal receive beamforming vector can be obtained by simplifying the optimization problem:

[0101]

[0102] In the above formula, K represents the total data volume of device m0.

[0103] 2. Optimize α when given f

[0104] In the hybrid relay-smart reflector, the |α of active and passive components l | is different, but its θ l All obey |θ l | 2 = 1, so it can be divided into two parts when optimizing α, as shown below:

[0105] 1) The first part is to optimize the phase coefficient of the hybrid relay-smart reflector

[0106] First, the objective function is optimized, and the optimized expression is as follows:

[0107]

[0108] In the above formula, for The lth term of l is the lth column of G. When optimizing the phase coefficient, the entire hybrid relay-intelligent reflector is assumed to be a passive intelligent reflector. According to this, the optimal phase of the hybrid relay-smart reflector can be obtained:

[0109]

[0110] In the above formula, arg{·} represents the phase of the complex-valued vector.

[0111] 2) The second part is to optimize the amplitude coefficient of the hybrid relay-intelligent reflector

[0112] In the setting of the hybrid relay-intelligent reflective surface, the present invention adopts a fixed hybrid relay-intelligent reflective surface, that is, the position and number of active elements are fixed.

[0113] Substituting the optimal phase of the hybrid relay-smart reflector, we can get:

[0114]

[0115] According to the above formula, the objective function can be transformed into:

[0116]

[0117] In the above formula, for n∈Λ, we have:

[0118]

[0119]

[0120]

[0121] r l =|f H γ l | 2

[0122]

[0123] There are certain power limitations on the active component portion of the hybrid relay-smart reflector.

[0124]

[0125] The above formula is equivalent to:

[0126]

[0127] In the above formula, make l∈Λ. Therefore, the above formula can be rewritten as:

[0128] After obtaining the optimization problem and constraints, the optimization problem can be solved. First, the optimization problem is expressed as:

[0129]

[0130]

[0131] In the above formula, by taking the derivative of the optimization problem, the closed solution of the optimization problem can be obtained under the condition of considering the constraints:

[0132]

[0133] in,

[0134] Therefore, through the above, the optimal α and f can be calculated when the equipment set is known. In the process of alternating optimization, the content that still needs to be optimized is: optimizing the equipment set when α and fi are known.

[0135] In the objective function, the first half, namely:

[0136]

[0137] This part is of course the minimum value when all devices are involved, that is, 0. However, the objective function still has the second half, namely:

[0138]

[0139] When gradually eliminating equipment, the first half of the formula gradually decreases, but the second half gradually increases, so the two formulas need to reach a trade-off under which the optimal objective function can be obtained. This principle needs to be respected when selecting equipment.

[0140] Next, we select the device. First, we calculate the channel gain of all devices to the base station, that is, h m (α) is sorted from large to small. Then the top q devices are selected to participate in federated learning and the corresponding objective function value is calculated, q∈[1,M]. The optimal q is selected by traversing all values ​​of q. * The optimal equipment set.

[0141] In getting q * After that, traverse q * The devices participating in learning calculate their corresponding confidentiality rates:

[0142]

[0143] In the formula, B represents the bandwidth allocated during signal transmission, h m (α) represents the channel gain from the device to the base station, h e (α) represents the channel gain from the device to the eavesdropper. When the confidentiality rate is less than 0, it means that the device does not have sufficient security and is excluded from the q * When the confidentiality rate is greater than or equal to 0, it means that the signal transmission of the device can guarantee sufficient security and it is retained in q *The confidentiality rate can be considered as the second round of equipment screening. Under two rounds of screening, accuracy and security can be guaranteed at the same time.

[0144] Therefore, in the example of the present invention, alternating optimization can be performed. First, the device set, the receiving beamforming vector, the phase of the hybrid relay-intelligent reflector, and the amplitude coefficient of the hybrid relay-intelligent reflector are initialized, and then the four parameters are optimized in sequence, and after multiple alternating optimizations, the optimal value can be obtained.

[0145] Figure 3 A performance indicator diagram for verifying the test accuracy of the algorithm proposed in the present invention is given. The diagram contains five cases: noise-free channel, hybrid relay-intelligent reflector surface, passive intelligent reflector surface, device-base station direct channel, and no device selection. It can be seen from the figure that the test accuracy of the algorithm proposed in the present invention has stronger stability and accuracy than the use of passive intelligent reflector surface in federated learning. In the simulation setting, random Rice channels are regenerated every 100 iterations, and the corresponding accuracy will fluctuate. When using passive intelligent reflector surface in federated learning, the precise value is smaller than that using hybrid relay-intelligent reflector surface. It can be easily seen from the figure that the use of hybrid relay-intelligent reflector surface in federated learning is better than the use of intelligent reflector surface.

[0146] The present invention proposes a physical layer secure communication method based on hybrid relay-intelligent reflection surface and federated learning. In the example of the present invention, the purpose of optimizing accuracy is achieved by alternately optimizing device selection, receiving beamforming vector, phase shift of hybrid relay-intelligent reflection surface, and magnification factor of hybrid relay-beamforming vector. In this way, a federated learning process with high accuracy, high convergence speed and high security can be obtained.

[0147] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A wireless communication method based on hybrid relay-intelligent reflection surface and federated learning, characterized in that: Design a physical layer secure communication system based on hybrid relay-intelligent reflection surface and federated learning, including local devices, base stations, hybrid relay-intelligent reflection surfaces and eavesdroppers; the local device is the transmitter of the system, and the base station is the receiver of the system. The transmitter sends the signal to the base station in two ways, one is the local device directly to the base station, and the other is the local device through the hybrid relay-intelligent reflection surface and then to the base station; the eavesdropper eavesdrops on the signal sent by the local device in this process; by optimizing device selection, hybrid relay-intelligent reflection surface amplification factor, phase shift and receiving beamforming vector, the error of the transmission process is minimized while ensuring security, thereby improving the accuracy of federated learning; The method comprises the following steps: Step S1, constructing a communication model according to the mathematical expressions of hybrid relay-intelligent reflection surface and federated learning; Step S2: Consider the model aggregation error generated in the process of federated learning, derive the upper limit of the learning loss in the iterative process, and thus construct it as an optimization problem; Step S3, minimizing the upper limit of the learning loss by alternately optimizing device selection, amplification factor of the hybrid relay-intelligent reflector, phase shift, and receive beamforming vector: first, simplifying the upper limit of the learning loss to a function expression that simultaneously contains device selection, amplification factor of the hybrid relay-intelligent reflector, phase shift, and receive beamforming vector; then alternately optimizing the variables in the function expression to obtain the corresponding optimal value; in the communication process, by determining the optimal value, it will be possible to ensure that the system obtains sufficient accuracy and security; In step S1, the constructed communication model can be expressed as: In the formula, y d represents the received signal at the base station, d represents the time slot, s d,m represents the transmission signal of the mth device in the dth time slot, M is the total number of devices, h UB,m is the channel gain vector of the mth device directly reaching the base station, and h H,m is the channel gain vector between the mth device and the hybrid relay-intelligent reflector, G is the channel gain vector between the hybrid relay-intelligent reflector and the base station, z H,m is the additive Gaussian white noise generated by the mth device when it passes through the hybrid relay-intelligent reflection surface, which has a mean of 0 and a variance of The complex Gaussian distribution of z BS is the additive Gaussian white noise generated at the base station, which has a mean of 0 and a variance of The complex Gaussian distribution of And meet σ 2 is the noise during device communication, I N represents the unit matrix of length N, Υ represents the reflection coefficient of the hybrid relay-intelligent reflector, Ψ represents the amplification factor of the hybrid relay-intelligent reflector, p m Represents the transmit power of the mth device; The base station converts the received signal into: In the formula, is the received signal after beamforming processing, f is the receive beamforming vector, and H is the conjugate transpose; In step S2, the optimization problem is constructed as follows: subject to x m ∈{0,1},1≤m≤M P a (a)≤P a max In the formula, x m is the device selection vector, K is the total data set of all devices, K m is the local data set of the mth device, f is the receive beamforming vector, the function variable α is taken from Υ, represents the specific value of the reflection coefficient of the hybrid relay-intelligent reflector, Λ represents the set of active elements in the hybrid relay-intelligent reflector, P0 is the upper limit of the device's transmit power, σ 2 is the noise in the device communication process, and M is the total number of devices; in the above formula, the first constraint is to determine whether the device is selected to participate in federated learning, the second constraint represents the amplification factor limit for the passive components in the hybrid relay-intelligent reflector, and the third constraint represents the power limit of the hybrid relay-intelligent reflector; in the above formula, the first half represents the error caused by the selection of the device, and the second half represents the error generated in the communication process. The two parts together constitute the error generated in the model aggregation process of federated learning; In step S3, the device selection method can be expressed as follows: First, among local devices, some devices have poor channel conditions, which is manifested as small channel gain. Such devices will affect the overall learning process in federated learning, so they need to be excluded in device selection; secondly, some devices have poor confidentiality performance and are easily eavesdropped by eavesdroppers during signal transmission. Therefore, it is necessary to quantify the confidentiality performance of the device, that is, use the confidentiality rate as the judgment standard, retain the device when it is greater than or equal to a certain threshold, and exclude it when it is less than a certain threshold; the confidentiality rate is expressed as follows: In the formula, B represents the bandwidth allocated during signal transmission, h m (α) represents the channel gain from the device to the base station, h e (α) represents the channel gain from the device to the eavesdropper.

Images