Deep learning assisted RIS physical layer secure transmission beamforming method

Abstract

The invention provides a deep learning assisted RIS physical layer secure transmission beam forming method, and belongs to the technical field of wireless communication. According to the method, the feature extraction capability of the deep learning algorithm and the constant modulus constraint solving capability of the Riemannian manifold optimization technology are combined, and the optimization direction and step length are dynamically regulated and controlled in the optimization process, so that the algorithm has an opportunity to explore a wider search space, and the greedy of the optimization process of the traditional optimization algorithm is effectively reduced. Compared with a traditional optimization algorithm, the method has more flexible and efficient optimization capacity and lower operation complexity, the system achieves a better safety rate, and meanwhile, compared with a traditional deep learning method, the optimization process of the method has higher interpretability.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a deep learning-assisted reconfigurable intelligent surface (RIS) physical layer secure transmission beamforming method. Background Art

[0002] The openness and broadcasting nature of wireless communications make transmitted signals vulnerable to illegal eavesdropping, especially in application scenarios involving sensitive data and critical infrastructure, where security requirements are more stringent. Traditional encryption technology and security protection methods can no longer effectively cope with complex electromagnetic environments and increasingly severe security threats, and innovative solutions are urgently needed to enhance the anti-eavesdropping capabilities and information confidentiality of communication systems.

[0003] The physical layer security transmission technology based on RIS dynamically adjusts the propagation characteristics of wireless signals, transforming passive channel adaptation into active channel adjustment, providing an innovative supplement to traditional security mechanisms and effectively enhancing the system's anti-eavesdropping capabilities. The physical layer security transmission system based on RIS mainly performs joint beamforming by the transmitter and RIS, giving full play to the accurate control capability of RIS on electromagnetic space. Through the virtual line-of-sight link and beamforming algorithm constructed by RIS, the legitimate signal is more focused on the legitimate receiving end and bypasses the eavesdropper, thereby effectively improving the system security performance. The physical layer security transmission based on RIS is usually modeled as a problem of maximizing the system security rate through the design of beamforming vectors and phase shift matrices, where the beamforming vector needs to meet the total power constraint of the base station and the phase shift matrix needs to meet the constraint of constant modulus 1. The optimization function and constraint conditions of this problem have non-convex characteristics, which bring difficulties to the beamforming design.

[0004] At present, there are two main solutions for the RIS-based secure beamforming problem, namely, convex optimization-based methods and artificial intelligence-based methods.

[0005] Methods based on convex optimization need to solve the non-convex characteristics of the problem first. For example, the original problem can be transformed into a convex problem using relaxation or approximation techniques. In addition, since each element of the RIS phase shift matrix satisfies the constraint of modulus 1, each element can be regarded as being on the Riemann complex circle manifold. This feature can be used to map the modulus 1 constraint of each element of the RIS phase shift matrix to the Riemann complex circle manifold using the Riemann manifold optimization technique, eliminating the influence of non-convex constraints and directly solving the unconstrained optimization problem on the Riemann complex circle manifold. Furthermore, for the highly coupled base station beamforming vector and RIS phase shift matrix, it is necessary to fix one variable as a constant, optimize the other variable, and then alternately fix and iterate the optimization until the final result converges, so as to achieve decoupling. The above-mentioned explicit solution to the beamforming problem based on the convex optimization method can effectively obtain a local suboptimal solution, but it has three defects. On the one hand, this method has undergone multiple iterative optimizations and involves a large amount of numerical calculations, resulting in high computational complexity, which grows polynomially or even exponentially, and is difficult to meet the application requirements of dynamic regulation of RIS in actual deployment. On the other hand, the above-mentioned alternating optimization algorithm will lead to performance loss and cannot fully obtain the security performance gain brought by RIS deployment. Finally, the alternating optimization algorithm cannot explore a broader search space due to the relatively fixed iterative update rules.

[0006] Artificial Intelligence (AI) algorithms represented by deep learning have efficient feature extraction capabilities and are widely used in joint beamforming optimization problems. AI-based intelligent beamforming algorithms are usually applied in a mode of offline training and online deployment. This method uses a large amount of data to train the model in advance, which can greatly reduce the computational complexity of the beamforming design of the RIS-assisted wireless communication system in actual application. However, since the labels required for network training are usually suboptimal solutions obtained by traditional convex optimization methods, the optimization performance of AI-based algorithms is limited by traditional convex optimization methods, and it is difficult to obtain better performance; and since the optimization process is completely replaced by a "black box" neural network, the uninterpretability of the neural network parameters limits the application of this method in sensitive scenarios that require strong interpretability and security. Summary of the invention

[0007] In view of the problem of non-optimal optimization results of traditional algorithms and high time complexity of the optimization process in the RIS physical layer security transmission beamforming problem, the present invention proposes a deep learning-assisted RIS physical layer security transmission beamforming method. The present invention combines the feature extraction capability of the deep learning algorithm and the ability of Riemann manifold optimization to solve constant modulus constraints, jointly optimizes the beamforming vector and the phase shift matrix, and realizes the physical layer security transmission beamforming design of the RIS-assisted communication system.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] A deep learning-assisted RIS physical layer secure transmission beamforming method comprises the following steps:

[0010] Step 1: Randomly initialize the base station beamforming vector and RIS phase shift matrix N is a set of complex numbers; the parameters of the neural network BF-Net used to optimize the beamforming vector are randomly initialized And the parameters of the neural network Theta-Net used to optimize the phase shift matrix parameter and parameters Including the weights and biases of each neuron in each network, both neural networks are deep neural networks; let the iteration number statistical variable i = 0, and the total number of iterations is T;

[0011] Step 2: Calculate the channel rate R at the legitimate user B (·) relative to the initialization beamforming vector w (0) The gradient is:

[0012]

[0013] and the channel rate R at the eavesdropper E (·) Gradient with respect to the initialization beamforming vector:

[0014]

[0015] in, is the cascade channel matrix of base station-RIS-legal user, is the cascade channel matrix of base station-RIS-eavesdropper, diag(·) is the diagonal matrix, (·) H To take the conjugate transpose operation, is the base station-RIS link channel matrix, is the base station-legal user direct link channel matrix, is the base station-eavesdropper link channel matrix, is the RIS-legal user link channel matrix, is the RIS-eavesdropper link channel matrix, and are the thermal noise at the receivers of the legitimate user and the eavesdropper, respectively;

[0016] Step 3: Calculate the system safety rate R S (·) relative to the initialization beamforming vector w (0) The gradient is:

[0017] Step 4: Gradient Input the neural network BF-Net, and compare the output of the neural network BF-Net with w (0) Add together and get Among them, BFN (i) (·) is the implicit function corresponding to the neural network BF-Net in the i-th cycle;

[0018] Step 5: Perform power normalization:

[0019]

[0020] Where P is the total transmission power of the base station, Ω(·) is the power normalization function, and ||·|| is the modulus length operation;

[0021] Step 6: Calculate the phase shift matrix Q (0) Relative to the safety rate R S The gradient of (·):

[0022]

[0023] Among them, a 1 =H B w (0) , a 2 =H E w (0) , (·) * To take the conjugate operation;

[0024] Step 7: Based on the gradient obtained in step 6 Calculate the phase shift matrix Q (0) Relative to the safety rate R S The Riemann gradient of (·) is:

[0025]

[0026] Among them, Re{·} is the real part operation, For Hadamard;

[0027] Step 8: Input the Riemann gradient into the neural network Theta-Net and compare the output of the neural network Theta-Net with Q (0) Add together and get TN (i) (·) is the implicit function of the neural network Theta-Net in the i-th cycle;

[0028] Step 9: Contract each element of the phase shift matrix to obtain a phase shift matrix that satisfies the constant modulus constraint.

[0029] Step 10: Calculate the loss function loss = -R based on the results of step (5) and step (9) S (w (i) ,Q (i) );

[0030] Step 11: Calculate the loss function relative to the network parameters according to the back-propagation algorithm and The gradient of , and then use the adaptive moment estimation optimizer to update the parameters of the two neural networks BF-Net and Theta-Net respectively:

[0031]

[0032] in, is the parameter of BF-Net in the i-th cycle, is the parameter of BF-Net in the i+1th cycle, is the loss function relative to The gradient of is the parameter of Theta-Net in the i-th cycle, is the parameter of Theta-Net in the i+1th cycle, is the loss function relative to The gradient of; Adam (·) is the adaptive moment estimation optimizer, α w is the learning rate of the adaptive moment estimation optimizer used to update the BF-Net parameters, α Q The learning rate of the adaptive moment estimation optimizer used to update the Theta-Net parameters;

[0033] Step 12: Update i=i+1. If i=T, output the beamforming vector w obtained at this time. (i) and the phase shift matrix Q (i) , otherwise jump to step 2;

[0034] Step 13: Perform secure transmission beamforming of the communication system according to the beamforming vector and phase shift matrix obtained in step 12.

[0035] Furthermore, the neural network Theta-Net includes an input layer, a hidden layer and an output layer, wherein the number of neurons in the input layer and the output layer is twice the number of RIS reflection units, and the activation function is a ReLu function.

[0036] Furthermore, the neural network BF-Net includes an input layer, a hidden layer and an output layer, wherein the number of neurons in the input layer and the output layer is twice the number of base station antennas, and the activation function is a ReLu function.

[0037] The present invention has the following advantages:

[0038] 1. The present invention uses the feature extraction capability of deep learning algorithms to implicitly and directly solve the non-convex optimization problem in beamforming design. The optimization direction and step size in the optimization process are dynamically controlled by the neural network, so that the algorithm can find a more flexible and efficient optimization direction.

[0039] 2. The present invention utilizes the ability of Riemannian manifold optimization technology to solve the constant modulus problem to solve the constant modulus constraint of each element in the RIS phase shift matrix, and effectively solves the optimization of the phase shift matrix while satisfying the constant modulus constraint.

[0040] 3. In the present invention, the neural network is dynamically updated according to the loss function during the iteration process. This design makes the optimization process more flexible and efficient, and the method can be plug-and-play, effectively avoiding the pre-training overhead of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of an application scenario of a secure transmission system to which the present invention is applicable.

[0042] Figure 2 It is a principle block diagram of the present invention.

[0043] Figure 3 It is a simulation curve diagram comparing the performance of the present invention with that of the benchmark algorithm. DETAILED DESCRIPTION

[0044] The present invention is described in detail below with reference to the accompanying drawings.

[0045] A deep learning-assisted RIS physical layer secure transmission beamforming method, which is applicable to secure transmission system application scenarios such as Figure 1 As shown in the figure, a multi-antenna base station attempts to transmit private information to a legitimate user with the assistance of RIS, and there are eavesdroppers in the environment who attempt to steal private information. The base station Alice is located at the origin of the coordinate system (0m, 0m), the legitimate user Bob is located at (200m, 30m), and the eavesdropper Eve is located at (165m, 20m). RIS is deployed at (200m, 0m) for secure transmission of private information. The number of Alice antennas at the base station is M=16, the number of reflection units of RIS is N, and both the legitimate user Bob and the eavesdropper Eve are single-antenna receivers. This method combines the feature extraction capability of the deep learning algorithm and the ability of Riemann manifold optimization to solve constant modulus constraints, jointly optimizes the beamforming vector and phase shift matrix, and realizes the physical layer secure transmission beamforming design of the RIS-assisted communication system.

[0046] like Figure 2 As shown, the specific steps of the method are as follows:

[0047] Step 1: Collect channel data from the application scenario of RIS-assisted secure transmission, specifically the channel matrix of the Alice-Bob direct link Channel matrix of Alice-Eve direct link Channel Matrix of Alice-RIS-Bob Auxiliary Link Channel matrix of Alice-RIS-Eve auxiliary link

[0048] Step 2: Randomly initialize the base station beamforming vector w (0) and RIS phase shift matrix Q (0) , randomly initialize the parameters of the neural network BF-Net that optimizes the beamforming vector And the parameters of the neural network Theta-Net that optimizes the phase shift matrix The learning rate of the Adaptive Moment Estimation (Adam) optimizer of BF-Net is 1e-5, and the learning rate of the Adam optimizer of Theta-Net is 5e-5. The input layer of BF-Net has 32 neurons, 3 hidden layers, 32 neurons in each hidden layer, the activation function is the Rectified Linear Unit (ReLu) function, and the output layer has 32 neurons. The input layer of Theta-Net has N neurons, 3 hidden layers, 32 neurons in each hidden layer, the activation function is the ReLu function, and the output layer has N neurons. Let the iteration count variable i = 0, and record the total number of iterations as T;

[0049] Step 3: Calculate the channel rate R at the legitimate user B (·) relative to the initialization beamforming vector w (0) The gradient is:

[0050]

[0051] and the channel rate R at the eavesdropper E (·) Gradient with respect to the initialization beamforming vector:

[0052]

[0053] in, is the cascade channel matrix of base station-RIS-legal user, is the cascade channel matrix of base station-RIS-eavesdropper, diag(·) is the diagonal matrix, (·) H To take the conjugate transpose operation, is the base station-RIS link channel matrix, is the base station-legal user direct link channel matrix, is the base station-eavesdropper link channel matrix, is the RIS-legal user link channel matrix, is the RIS-eavesdropper link channel matrix, and are the thermal noise at the receivers of the legitimate user and the eavesdropper, respectively;

[0054] Step 4: Calculate the system safety rate R S (·) relative to the initialization beamforming vector w (0) The gradient is:

[0055] Step 5: Gradient Input neural network BF-Net, network output and w (0) Add together and get Among them, BFN (i) (·) is the implicit function of the neural network (BF-Net) used to optimize the beamforming vector in the i-th cycle;

[0056] Step 6: Perform power normalization:

[0057] Step 7: Calculate the phase shift matrix Q (0) Relative to the safety rate R S The gradient of (·):

[0058]

[0059] Among them, a 1 =H B w (0) , a 2 =H E w (0) , (·) * To take the conjugate operation;

[0060] Step 8: Based on the gradient obtained in step 7 Calculate the phase shift matrix Q (0) Relative to the safety rate R S The Riemann gradient of (·) is:

[0061]

[0062] Among them, Re{·} is the real part operation, and ° is the Hadamard product. ;

[0063] Step 9: Input the Riemann gradient into the neural network Theta-Net, and the network output is consistent with Q (0) Add together and get TN (i)(·) is the implicit function of the DNN neural network (Theta-Net) used to optimize the phase shift matrix in the i-th cycle;

[0064] Step 10: Contract each element of the phase shift matrix to obtain a phase shift matrix that satisfies the constant modulus constraint.

[0065] Step 11: Calculate the loss function loss = -R based on the results of step 6 and step 10 S (w (i) ,Q (i) );

[0066] Step 12: Calculate the loss function with respect to the two network parameters according to the back-propagation algorithm and The gradient of , and then use the Adaptive Moment Estimation (Adam) optimizer to update the parameters of the two neural networks respectively:

[0067]

[0068] Step 13: The iteration number statistical variable is incremented, that is, i=i+1. If i=T, the beamforming vector w obtained at this moment is output. (i) and Q (i) , otherwise jump to step 2.

[0069] Step 14: Design a secure transmission beamforming for the communication system based on the beamforming vector and phase shift matrix obtained in step 13.

[0070] Principle description:

[0071] This method combines the feature extraction capability of deep learning algorithms with the ability of Riemannian manifold optimization to solve constant modulus constraints, and dynamically controls the optimization direction and step size during the optimization process. This method has the opportunity to explore a wider search space during the optimization process, effectively reducing the greediness of the optimization process of traditional optimization algorithms, and obtaining better local suboptimal solutions for beamforming vectors and phase shift matrices, effectively improving the system safety rate.

[0072] The performance of this method under different numbers of RIS reflection units is analyzed and compared with the performance of existing benchmark methods. Three representative benchmark algorithms are selected. Benchmark method 1 is an alternating optimization algorithm based on semi-definite relaxation and Gaussian randomization, denoted as AO-SDR algorithm; benchmark method 2 is an alternating optimization algorithm based on Rayleigh quotient function and binary search, denoted as AO-OBO algorithm; benchmark algorithm 3 is an alternating optimization algorithm based on block coordinate descent, denoted as AO-CD algorithm. Figure 3It can be seen that this method is superior to the three existing benchmark algorithms under different RIS reflection unit numbers, and can enable the system to obtain a higher safety rate. When the number of RIS reflection units is 32, the safety rate obtained by the three benchmark methods is about 4.8bps / Hz, and the safety rate obtained by this method is 5.5bps / Hz. When the number of RIS reflection units is 144, the safety rate obtained by the three benchmark methods is about 6.7bps / Hz, and the safety rate obtained by this method is 6.9bps / Hz. The simulation results show that this method has superior performance and can obtain better local suboptimal solutions in complex non-convex problems.

[0073] The present invention combines the feature extraction capability of deep learning algorithms and the ability of Riemann manifold optimization technology to solve constant modulus constraints, dynamically controls the optimization direction and step size during the optimization process, so that the algorithm has the opportunity to explore a wider search space, effectively reducing the greed of the optimization process of traditional optimization algorithms. Compared with traditional optimization algorithms, the present invention has more flexible and efficient optimization capabilities and lower computational complexity, enabling the system to achieve a better safety rate. At the same time, compared with traditional deep learning methods, the optimization process of the present invention has stronger interpretability.

[0074] In summary, the present invention jointly designs the base station beamforming vector and the RIS phase shift matrix, and utilizes the RIS-based information transmission system, the joint beamforming technology in the physical layer security and the channel control capability of RIS to intelligently control the signal propagation environment, thereby enhancing the channel quality of the legitimate receiver and deteriorating the channel quality of the eavesdropper, thereby achieving the purpose of improving the security of information transmission and ensuring the security of system information transmission.

Claims

1. A deep learning-assisted RIS physical layer secure transmission beamforming method, characterized in that: The steps include: Step 1: Randomly initialize the base station beamforming vector and RIS phase shift matrix N is a set of complex numbers; the parameters of the neural network BF-Net used to optimize the beamforming vector are randomly initialized And the parameters of the neural network Theta-Net used to optimize the phase shift matrix parameter and parameters Including the weights and biases of each neuron in their respective networks. Both neural networks are deep neural networks. Let the iteration count variable i = 0, and record the total number of iterations as T; Step 2: Calculate the channel rate R at the legitimate user B (·) relative to the initialization beamforming vector w (0) The gradient is: and the channel rate R at the eavesdropper E (·) Gradient with respect to the initialization beamforming vector: in, is the cascade channel matrix of base station-RIS-legal user, is the cascade channel matrix of base station-RIS-eavesdropper, diag(·) is the diagonal matrix, (·) H To take the conjugate transpose operation, is the base station-RIS link channel matrix, is the base station-legal user direct link channel matrix, is the base station-eavesdropper link channel matrix, is the RIS-legal user link channel matrix, is the RIS-eavesdropper link channel matrix, and are the thermal noise at the receivers of the legitimate user and the eavesdropper, respectively; Step 3: Calculate the system safety rate R S (·) relative to the initialization beamforming vector w (0) The gradient is: Step 4: Gradient Input the neural network BF-Net, and compare the output of the neural network BF-Net with w (0) Add together and get Among them, BFN (i) (·) is the implicit function corresponding to the neural network BF-Net in the i-th cycle; Step 5: Perform power normalization: Where P is the total transmission power of the base station, Ω(·) is the power normalization function, and ||·|| is the modulus length operation; Step 6: Calculate the phase shift matrix Q (0) Relative to the safety rate R S The gradient of (·): Where a1 = H B w (0) , a2=H E w (0) , (·) * To take the conjugate operation; Step 7: Based on the gradient obtained in step 6 Calculate the phase shift matrix Q (0) Relative to the safety rate R S The Riemann gradient of (·) is: Among them, Re{·} is the real part operation, For Hadamard; Step 8: Input the Riemann gradient into the neural network Theta-Net and compare the output of the neural network Theta-Net with Q (0) Add together and get TN (i) (·) is the implicit function of the neural network Theta-Net in the i-th cycle; Step 9: Contract each element of the phase shift matrix to obtain a phase shift matrix that satisfies the constant modulus constraint. Step 10: Calculate the loss function loss = -R based on the results of step (5) and step (9) S (w (i) ,Q (i) ); Step 11: Calculate the loss function relative to the network parameters according to the back-propagation algorithm and The gradient of , and then use the adaptive moment estimation optimizer to update the parameters of the two neural networks BF-Net and Theta-Net respectively: in, is the parameter of BF-Net in the i-th cycle, is the parameter of BF-Net in the i+1th cycle, is the loss function relative to The gradient of is the parameter of Theta-Net in the i-th cycle, is the parameter of Theta-Net in the i+1th cycle, is the loss function relative to The gradient of; Adam (·) is the adaptive moment estimation optimizer, α w is the learning rate of the adaptive moment estimation optimizer used to update the BF-Net parameters, α Q The learning rate of the adaptive moment estimation optimizer used to update the Theta-Net parameters; Step 12: Update i=i+1. If i=T, output the beamforming vector w obtained at this time. (i) and the phase shift matrix Q (i) , otherwise jump to step 2; Step 13: Perform secure transmission beamforming of the communication system according to the beamforming vector and phase shift matrix obtained in step 12.

2. According to a deep learning-assisted RIS physical layer secure transmission beamforming method according to claim 1, it is characterized in that: The neural network Theta-Net includes an input layer, a hidden layer and an output layer, wherein the number of neurons in the input layer and the output layer is twice the number of RIS reflection units, and the activation function is a ReLu function.

3. According to a deep learning-assisted RIS physical layer secure transmission beamforming method according to claim 1, it is characterized in that: The neural network BF-Net includes an input layer, a hidden layer and an output layer, wherein the number of neurons in the input layer and the output layer is twice the number of base station antennas, and the activation function is a ReLu function.

Images