Wireless federated learning method assisted by intelligent reflecting surface for covert aerial computing

Abstract

The application relates to a wireless federated learning method for intelligent reflecting surface assisted covert aerial computing. The intelligent reflecting surface is used to realize the covert aerial computing based federated learning. The reflecting phase and amplitude can be designed to improve the signal at the access point and reduce the eavesdropping performance at the eavesdropper, so that the performance of the federated learning system based on the covert aerial computing is improved. More user equipment is selected under certain mean square error requirements and covert constraints, so that lower training loss and higher test precision are realized.

Description

Technical Field

[0001] The present application relates to the field of mobile communication technology, and in particular to a wireless federated learning method for intelligent reflective surface-assisted covert aerial computing. Background Art

[0002] Federated learning, an emerging distributed artificial intelligence (AI) method with privacy-preserving properties, holds particular promise for wireless communications and is considered a key solution for achieving ubiquitous AI in 6G. Federated learning can serve as a supporting technology in mobile edge networks, enabling collaborative training of machine learning (ML) models and deep learning (DL) for mobile edge network optimization.

[0003] Because edge devices typically connect to edge servers via wireless channels, the model parameters received by the edge servers are inevitably distorted by channel fading and additive noise. The required radio resources scale linearly with the number of edge devices participating in federated learning. When the number of edge devices is large, significant communication latency is introduced during model aggregation, becoming a performance-limiting factor in federated learning.

[0004] To address these challenges, federated learning, powered by over-the-air computing, has emerged to improve learning performance under limited communication bandwidth and strict latency requirements. Over-the-air computing leverages the waveform superposition characteristics of multi-access channels to combine concurrent data transmission from multiple devices with function computation. Furthermore, since edge servers in federated learning are only interested in aggregated models, over-the-air computing, as a non-orthogonal multiple access (NOMA) scheme, is considered a solution for achieving efficient spectrum utilization and low power consumption. However, unfavorable propagation environments can lead to security issues, which in turn can lead to aggregate privacy leaks in federated learning models based on over-the-air computing. Summary of the Invention

[0005] In order to overcome at least one shortcoming in the prior art, the present application provides a wireless federated learning method for covert aerial computing assisted by an intelligent reflective surface.

[0006] In a first aspect, a wireless federated learning method for intelligent reflective surface-assisted covert air computing is provided, comprising:

[0007] Establish a wireless federated learning model, which includes antenna base stations, multiple user devices, smart reflective surfaces, and eavesdroppers.

[0008] Each user device transmits a first signal to the smart reflecting surface and transmits a second signal to the base station. After receiving the first signal, the smart reflecting surface generates a reflected signal and transmits the reflected signal to the base station. An eavesdropper can receive the first concealed signal during the process of the user device transmitting signals to the smart reflecting surface and the base station, and can receive the second concealed signal during the process of the smart reflecting surface transmitting the transmitted signal to the base station. The reflected signal is in phase with the second signal, and the first concealed signal is in phase with the second concealed signal.

[0009] Determine, among multiple user devices, a selection set of user devices that participate in wireless federated learning in each round of communication;

[0010] Wireless federated learning is performed based on a determined selection set of user devices participating in wireless federated learning and a wireless federated learning model.

[0011] In one embodiment, determining a selection set of user devices that participate in wireless federated learning in each round of communication among multiple user devices includes:

[0012] Determine the priority of all user devices;

[0013] Determine the number k of user equipment in the user equipment selection set;

[0014] According to the priorities of the user equipment, k user equipments with higher priorities are selected to form a user equipment selection set. The selected k user equipments satisfy the following formula:

[0015]

[0016]

[0017] |Θ m,n |≤1,m=1,2,…,M,n=1,2,…,N

[0018]

[0019] Among them, w i is the transmit power of user equipment i, m is the focused beamforming vector at the base station, H represents the transposed conjugate, S is the user equipment selection set, P0 is the maximum transmit power, and h rb is the channel gain from the smart reflector to the base station, h ir The channel gain from user equipment i to the smart reflector, h ib is the channel gain from user equipment i to the base station, h rw is the channel gain from the smart reflective surface to the eavesdropper, h iw is the channel gain from user device i to the eavesdropper, σ w is the noise power of the eavesdropper, ε represents the constraint threshold of the concealment condition, L is the number of times the channel is used, Θm,n is the reflection coefficient of the reflection unit in the mth row and nth column of the smart reflection surface, M is the total number of rows of reflection units in the smart reflection surface, and N is the total number of columns of reflection units in the smart reflection surface.

[0020] In one embodiment, the smart reflecting surface generates a reflected signal after receiving the first signal, including:

[0021] Determining a diagonal reflection matrix of the smart reflective surface, where the diagonal reflection matrix includes a reflection coefficient of each reflective unit in the smart reflective surface;

[0022] A reflection signal is generated according to the diagonal reflection matrix of the smart reflection surface.

[0023] In a second aspect, a wireless federated learning device for intelligent reflective surface-assisted covert aerial computing is provided, comprising:

[0024] A wireless federated learning model building module is used to build a wireless federated learning model, which includes an antenna base station, multiple user devices, a smart reflective surface, and an eavesdropper.

[0025] Each user device transmits a first signal to the smart reflecting surface and a second signal to the base station. After receiving the first signal, the smart reflecting surface generates a reflected signal and transmits the reflected signal to the base station. An eavesdropper can receive the first concealed signal during the user device's transmission of signals to the smart reflecting surface and the base station, and can receive the second concealed signal during the smart reflecting surface's transmission of the transmitted signal to the base station. The reflected signal is in phase with the second signal, and the first concealed signal is in phase with the second concealed signal.

[0026] A user equipment selection set determination module is used to determine a user equipment selection set participating in wireless federated learning in each round of communication among multiple user equipments;

[0027] The wireless federated learning module is used to perform wireless federated learning based on a determined selection set of user devices participating in the wireless federated learning and a wireless federated learning model.

[0028] In one embodiment, the user equipment selection set determination module is further configured to:

[0029] Determine the priority of all user devices;

[0030] Determine the number k of user equipment in the user equipment selection set;

[0031] According to the priorities of the user equipment, k user equipments with higher priorities are selected to form a user equipment selection set. The selected k user equipments satisfy the following formula:

[0032]

[0033]

[0034] |Θ m,n |≤1,m=1,2,…,M,n=1,2,…,N

[0035]

[0036] Among them, w i is the transmit power of user equipment i, m is the focused beamforming vector at the base station, H represents the transposed conjugate, S is the user equipment selection set, P0 is the maximum transmit power, and h rb is the channel gain from the smart reflector to the base station, h ir The channel gain from user equipment i to the smart reflector, h ib is the channel gain from user equipment i to the base station, h rw is the channel gain from the smart reflective surface to the eavesdropper, h iw is the channel gain from user device i to the eavesdropper, σ w is the noise power of the eavesdropper, ε represents the constraint threshold of the concealment condition, L is the number of times the channel is used, Θ m,n is the reflection coefficient of the reflection unit in the mth row and nth column of the smart reflection surface, M is the total number of rows of reflection units in the smart reflection surface, and N is the total number of columns of reflection units in the smart reflection surface.

[0037] In one embodiment, the smart reflecting surface generates a reflected signal after receiving the first signal, including:

[0038] Determining a diagonal reflection matrix of the smart reflective surface, where the diagonal reflection matrix includes a reflection coefficient of each reflective unit in the smart reflective surface;

[0039] A reflection signal is generated according to the diagonal reflection matrix of the smart reflection surface.

[0040] Compared with the prior art, this application has the following beneficial effects:

[0041] This application uses smart reflective surfaces to implement covert federated learning based on air computing. The reflection phase can be designed to enhance the signal at the access point and reduce the eavesdropping performance at the Willie point, thereby improving the performance of the federated learning system based on covert air computing.

[0042] Compared with a baseline solution without smart reflective surfaces, this application achieves lower training loss and higher test accuracy by selecting more devices under certain mean square error requirements and hidden constraints; in addition, this application enables federated learning to converge faster and achieve more accurate predictions than other baseline solutions in experiments training deep convolutional neural networks (CNNs) on the MNIST dataset. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The present application may be better understood by referring to the following description in conjunction with the accompanying drawings, which together with the following detailed description are incorporated into and form a part of this specification. In the drawings:

[0044] Figure 1 shows a schematic diagram of the structure of the wireless federated learning module;

[0045] Figure 2 A comparison chart showing the average number of selected user equipments of the present application method and the existing method when ε=0.1 changes with the mean square error threshold γ is shown;

[0046] Figure 3 A comparison chart showing the average number of selected user equipments of the present application method and the existing method when ε=0.01 changes with the mean square error threshold γ is shown;

[0047] Figure 4 A comparison of the training loss between the present invention and the existing methods is shown;

[0048] Figure 5 A comparison chart of the test accuracy of the present application method and the existing method is shown. DETAILED DESCRIPTION

[0049] Exemplary embodiments of the present application are described below with reference to the accompanying drawings. For the sake of clarity and conciseness, not all features of actual embodiments are described in this specification. However, it should be understood that in the process of developing any such actual embodiment, many implementation-specific decisions may be made to achieve the developer's specific goals, and these decisions may vary from one implementation to another.

[0050] It is also necessary to explain here that, in order to avoid obscuring the present application due to unnecessary details, the accompanying drawings only show the device structure closely related to the solution according to the present application, while other details that are not closely related to the present application are omitted.

[0051] It should be understood that the present application is not limited to the described embodiments due to the following description with reference to the accompanying drawings. In this document, where feasible, the embodiments may be combined with each other, features between different embodiments may be replaced or borrowed, and one or more features may be omitted in one embodiment.

[0052] The intelligent reflecting surface (IRS) is a revolutionary technology for enabling intelligent and programmable wireless environments. It is a software-controllable metasurface composed of a large number of passive scattering components. It manipulates the phase of impinging signals in a full-duplex manner, transforming a random fading channel into an intelligent environment conducive to communication. Unlike traditional active transmission technologies, intelligent reflecting surfaces do not require a transmission radio frequency (RF) chain. Compared to existing technologies based on active components, this reduces energy and hardware costs. Deploying intelligent reflecting surfaces in harsh wireless environments can yield benefits, making it feasible to use them to compensate for the amplitude reduction and misalignment of over-the-air computations in federated learning systems, achieving lower training loss and higher test accuracy with fewer communication rounds. Furthermore, intelligent reflecting surfaces can create a favorable propagation environment to improve the quality of legitimate communications and protect adversaries from detecting malicious attack signals.

[0053] In wireless federated learning with smart reflective surfaces assisting covert over-the-air computing, the smart reflective surface can, on the one hand, reflect the desired signal in phase with the signal at the access point, thereby strengthening the signal and improving the accuracy of model aggregation. This is known as wireless model aggregation enhancement. On the other hand, the covert signal received at the warden is weakened by the smart reflective surface. The smart reflective surface can reflect signals with opposite phases to the signal at the warden, which is known as wireless model aggregation leakage.

[0054] The present application provides a wireless federated learning method for intelligent reflective surface-assisted covert aerial computing, including:

[0055] Step S1: Establish a wireless federated learning model, which includes an antenna base station, multiple user devices, a smart reflecting surface, and an eavesdropper. Here, the smart reflecting surface includes multiple reflecting units.

[0056] Figure 1 The structure diagram of the wireless federated learning module is shown in Figure 1 , each user equipment i is based on the channel gain h from user equipment i to the smart reflective surface ir The first signal is transmitted to the smart reflective surface, i=1,…,I, i is the index of the user equipment, I is the total number of user equipment, and based on the channel gain h from the user equipment i to the base station ib The second signal is transmitted to the base station. The smart reflective surface generates a reflected signal after receiving the first signal, and based on the channel gain h from the smart reflective surface to the base station rb , transmits the transmission signal to the base station; the eavesdropper can detect the channel gain h from the user device i to the eavesdropper during the process of the user device transmitting the signal to the smart reflective surface and the base station. iwThe first concealed signal is received, and in the process of the smart reflective surface transmitting the transmission signal to the base station, the channel gain h from the smart reflective surface to the eavesdropper is calculated based on the following equation: rw A second concealed signal is received; the reflected signal is in phase with the second signal, and the first concealed signal is in phase opposition with the second concealed signal;

[0057] Step S2: determining a selection set of user devices that participate in wireless federated learning in each round of communication from among multiple user devices;

[0058] Step S3: Based on the determined selection set of user devices participating in wireless federated learning and the wireless federated learning model, wireless federated learning is performed. Once the selection set of user devices is determined, only the user devices in the selection set participate in wireless federated learning. The specific process of wireless federated learning is conventional and will not be described in detail here.

[0059] In one embodiment, in step S2, determining a selection set of user equipments that participate in wireless federated learning in each round of communication from among multiple user equipments includes:

[0060] First, the priorities of all user devices are determined; a sparse conversion method is used to determine the difference between the mean square error requirement and the achievable mean square error of the user device. The smaller the difference, the higher the priority of the user device.

[0061] The number k of user equipment in the user equipment selection set is determined; here, a binary search method may be used to determine the number k of user equipment.

[0062] Finally, k user devices with higher priorities are selected according to the priorities of the user devices to form a user device selection set. The selected k user devices satisfy the following formula:

[0063]

[0064]

[0065] |Θ m,n |≤1,m=1,2,…,M,n=1,2,…,N

[0066]

[0067] Among them, w i is the transmit power of user equipment i, m is the focused beamforming vector at the base station, H represents the transposed conjugate, S is the user equipment selection set, P0 is the maximum transmit power, and h rb is the channel gain from the smart reflector to the base station, h ir The channel gain from user equipment i to the smart reflector, h ib is the channel gain from user equipment i to the base station, h rwis the channel gain from the smart reflective surface to the eavesdropper, h iw is the channel gain from user device i to the eavesdropper, σ w is the noise power of the eavesdropper, ε represents the constraint threshold of the concealment condition, L is the number of times the channel is used, Θ m,n is the reflection coefficient of the reflection unit in the mth row and nth column of the smart reflection surface, M is the total number of rows of reflection units in the smart reflection surface, and N is the total number of columns of reflection units in the smart reflection surface.

[0068] In this embodiment, the first formula of the formulas satisfied by the selected k user equipment represents the optimization objective of minimum and maximum signal fusion mean square error, the second formula represents the hidden constraint condition, the third formula represents the reflection coefficient constraint of the smart reflection surface, and the fourth formula represents the beamforming constraint of the base station.

[0069] In one embodiment, in step S1, the smart reflective surface generates a reflection signal after receiving the first signal, including:

[0070] First, determine the diagonal reflection matrix of the smart reflection surface. The diagonal reflection matrix includes the reflection coefficient of each reflection unit in the smart reflection surface. Here, the reflection coefficient of each reflection unit includes the phase shift and amplitude of the combined incident signal of all user devices in the user device selection set incident on the reflection unit. The diagonal reflection matrix is ​​expressed as Among them, ρ t is the amplitude of the combined incident signal corresponding to the tth reflection unit, θ t is the phase shift of the combined incident signal corresponding to the t-th reflector unit, 1≤t≤K, K is the number of transmitting units in the smart reflector, θ t ∈[0,2),ρ t ∈[0,1].

[0071] Then, a reflection signal is generated according to the diagonal reflection matrix of the smart reflection surface.

[0072] Based on the same inventive concept as the wireless federated learning method for intelligent reflective surface-assisted covert over-the-air computing provided in the aforementioned embodiment, the present application further provides a wireless federated learning device for intelligent reflective surface-assisted covert over-the-air computing, comprising:

[0073] A wireless federated learning model building module is used to build a wireless federated learning model, which includes an antenna base station, multiple user devices, a smart reflective surface, and an eavesdropper.

[0074] Each user device transmits a first signal to the smart reflecting surface and a second signal to the base station. After receiving the first signal, the smart reflecting surface generates a reflected signal and transmits the transmitted signal to the base station. An eavesdropper can receive the first concealed signal during the user device's transmission of signals to the smart reflecting surface and the base station, and can receive the second concealed signal during the smart reflecting surface's transmission of the transmitted signal to the base station. The reflected signal is in phase with the second signal, and the first concealed signal is in phase with the second concealed signal.

[0075] A user equipment selection set determination module is used to determine a user equipment selection set participating in wireless federated learning in each round of communication among multiple user equipments;

[0076] The wireless federated learning module is configured to perform wireless federated learning based on a determined selection set of user devices participating in the wireless federated learning and a wireless federated learning model. Here, after the user device selection set is determined, only user devices in the user device selection set participate in the wireless federated learning.

[0077] The specific implementation functions of each module in the above embodiment are consistent with the specific implementation functions in the aforementioned method embodiment, and will not be repeated here.

[0078] In order to further demonstrate the effectiveness of the method of the present application, computer numerical simulation methods are used to verify the comparison results between the method of the present application and other existing methods in terms of the average number of selected user devices, training loss and test accuracy.

[0079] Figure 2 A comparison chart showing the average number of selected user equipments of the present invention and the existing method when ε=0.1 changes with the mean square error threshold γ is shown. Figure 3 A comparison chart showing the average number of selected user equipments of the present invention and the existing method when ε=0.01 changes with the mean square error threshold γ is shown. Figure 4 The following figure shows the comparison of the training loss between the present invention and the existing method. Figure 5 The comparison chart of the test accuracy of the present invention and the existing method is shown. Figure 2-5 It can be seen that this application achieves lower training loss and higher test accuracy by selecting more devices under certain mean square error requirements and hidden constraints. The method of this application is superior to existing methods in terms of the average number of selected devices, training loss and test accuracy.

[0080] In summary, the wireless federated learning method for covert aerial computing assisted by intelligent reflective surfaces in this application has the following technical effects:

[0081] By using smart reflecting surfaces to implement covert federated learning based on air computing, the reflection phase can be designed to enhance the signal at the access point and reduce the eavesdropping performance at Willie, thereby improving the performance of the federated learning system based on covert air computing; compared with the baseline scheme without smart reflecting surfaces, the present application achieves lower training loss and higher test accuracy by selecting more devices under certain mean square error requirements and concealment constraints; in addition, the present application enables federated learning to converge faster than other baseline schemes and achieve more accurate predictions in experiments on training deep convolutional neural networks (CNN) on the MNIST dataset.

[0082] The above descriptions are merely examples of various embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any modifications or substitutions that can be readily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included within the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A wireless federated learning method for intelligent reflective surface-assisted covert air computing, characterized in that: include: Establish a wireless federated learning model, which includes antenna base stations, multiple user devices, smart reflective surfaces, and eavesdroppers. Each user equipment transmits a first signal to a smart reflecting surface and a second signal to a base station. The smart reflecting surface generates a reflected signal after receiving the first signal and transmits the reflected signal to the base station. The eavesdropper can receive the first concealed signal during the process of the user equipment transmitting signals to the smart reflecting surface and the base station, and can receive the second concealed signal during the process of the smart reflecting surface transmitting the transmitted signal to the base station. The reflected signal is in phase with the second signal, and the first concealed signal is in phase with the second concealed signal. Determining, among the plurality of user equipments, a selection set of user equipments that participate in wireless federated learning in each round of communication; Performing wireless federated learning based on the determined selection set of user devices participating in wireless federated learning and the wireless federated learning model; Among the multiple user equipments, determining a selection set of user equipments that participate in wireless federated learning in each round of communication includes: Determine the priority of all user devices; Determine the number k of user equipment in the user equipment selection set; According to the priorities of the user equipment, k user equipments with higher priorities are selected to form a user equipment selection set. The selected k user equipments satisfy the following formula: in, is the transmit power of user equipment i, m is the focused beamforming vector at the base station, represents the transposed conjugate, S is the user equipment selection set, is the maximum transmit power, is the channel gain from the smart reflector to the base station, The channel gain from user equipment i to the smart reflector surface, is the channel gain from user equipment i to the base station, is the channel gain from the smart reflector to the eavesdropper, is the channel gain from user device i to the eavesdropper, is the noise power of the eavesdropper, represents the constraint threshold of the hidden condition, is the number of times the channel is used, is the reflection coefficient of the reflective unit in the mth row and nth column of the smart reflective surface, M is the total number of rows of reflective units in the smart reflective surface, and N is the total number of columns of reflective units in the smart reflective surface; Represents the diagonal reflection matrix of the smart reflective surface.

2. The method according to claim 1, wherein The smart reflecting surface generates a reflected signal after receiving the first signal, including: Determining a diagonal reflection matrix of the smart reflective surface, wherein the diagonal reflection matrix includes a reflection coefficient of each reflective unit in the smart reflective surface; The reflection signal is generated according to the diagonal reflection matrix of the smart reflection surface.

3. A wireless federated learning device with intelligent reflective surface-assisted covert air computing, characterized in that: include: A wireless federated learning model building module is used to build a wireless federated learning model, which includes an antenna base station, multiple user devices, a smart reflective surface, and an eavesdropper. Each user device transmits a first signal to a smart reflecting surface and a second signal to a base station. After receiving the first signal, the smart reflecting surface generates a reflected signal and transmits the reflected signal to the base station. The eavesdropper can receive the first concealed signal during the process of the user device transmitting signals to the smart reflecting surface and the base station, and can receive the second concealed signal during the process of the smart reflecting surface transmitting the transmitted signal to the base station. The reflected signal is in phase with the second signal, and the first concealed signal is in phase with the second concealed signal. A user equipment selection set determination module is used to determine a user equipment selection set that participates in wireless federated learning in each round of communication among the multiple user equipments; A wireless federated learning module, configured to perform wireless federated learning based on a determined selection set of user devices participating in the wireless federated learning and the wireless federated learning model; The user equipment selection set determination module is further configured to: Determine the priority of all user devices; Determine the number k of user equipments in the user equipment selection set; According to the priorities of the user equipment, k user equipments with higher priorities are selected to form a user equipment selection set. The selected k user equipments satisfy the following formula: in, is the transmit power of user equipment i, m is the focused beamforming vector at the base station, represents the transposed conjugate, S is the user equipment selection set, is the maximum transmit power, is the channel gain from the smart reflector to the base station, The channel gain from user equipment i to the smart reflector surface, is the channel gain from user equipment i to the base station, is the channel gain from the smart reflector to the eavesdropper, is the channel gain from user device i to the eavesdropper, is the noise power of the eavesdropper, represents the constraint threshold of the hidden condition, is the number of times the channel is used, is the reflection coefficient of the reflective unit in the mth row and nth column of the smart reflective surface, M is the total number of rows of reflective units in the smart reflective surface, and N is the total number of columns of reflective units in the smart reflective surface; Represents the diagonal reflection matrix of the smart reflective surface.

4. The device according to claim 3, characterized in that The smart reflecting surface generates a reflected signal after receiving the first signal, including: Determining a diagonal reflection matrix of the smart reflective surface, wherein the diagonal reflection matrix includes a reflection coefficient of each reflective unit in the smart reflective surface; The reflection signal is generated according to the diagonal reflection matrix of the smart reflection surface.

Images