Method for maximizing security energy efficiency based on unsupervised attention residual network

Abstract

The application relates to a security energy efficiency maximization method based on an unsupervised attention residual network. The method comprises the following steps: acquiring a transmitting signal of an integrated RIS, modeling an RIS-to-legal user channel, and respectively calculating a signal-to-interference-and-noise ratio and an achievable rate at a legal user and an eavesdropper. A multicast security capacity is designed according to a minimum value of the achievable rate of the legal user and a rate of an eavesdropping link, and total power consumption is obtained by combining RIS transmitting power and hardware dissipation, and energy utilization efficiency is calculated according to the total power consumption. In order to take the efficiency as a target, a legal user rate, RIS power and CRB constraints are added, and an optimization model is constructed. Then, an unsupervised attention residual network is used, a cosine annealing LR scheduling and a preset loss function are trained to obtain a supervised model, and finally, RIS phase shift and power allocation parameters are obtained by solving the model. The method can maximize the SEE in a dynamic vehicle-mounted environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to a safety energy efficiency maximization method based on an unsupervised attention residual network. BACKGROUND

[0002] In recent years, as a key candidate technology for 6G networks, integrated sensing and communication (ISAC) technology realizes the fusion of communication and sensing functions by jointly optimizing hardware and spectrum resources, meeting the demand of intelligent transportation systems (ITS) for ultra-reliable low-latency communication (URLLC) and high-precision environmental perception. However, traditional ISAC systems face severe challenges in dynamic vehicular environments: on the one hand, the high-dimensional optimization problem introduced by reconfigurable intelligent surfaces (RIS) leads to a sharp increase in the computational complexity of traditional mathematical optimization methods, making it difficult to meet the millisecond-level delay requirement; on the other hand, data-driven methods such as reinforcement learning (RL) have slow convergence speed and low exploration efficiency in continuous space, which cannot adapt to the rapid changes in channels in high-speed mobile scenarios.

[0003] Although supervised deep learning (DL) methods achieve real-time inference through offline label learning, they rely on large-scale codebook generation for label learning, which faces high labeling costs in large-scale RIS deployment. In addition, the frequent channel changes in dynamic vehicular environments require ISAC systems to have unsupervised adaptive capabilities, and existing solutions cannot simultaneously meet the safety energy efficiency (SEE) maximization and the constraint of perception accuracy (CRB) without prior labels. SUMMARY

[0004] Therefore, it is necessary to provide a safety energy efficiency maximization method based on an unsupervised attention residual network, which can maximize SEE in a dynamic vehicular environment and solve the technical problems of traditional methods, such as high computational complexity and dependence on labels.

[0005] A safety energy efficiency maximization method based on an unsupervised attention residual network, the method is applied to an RIS-assisted ISAC system including a transmitting integrated RIS configured with reflecting elements, a legitimate user, a sensing target, and an eavesdropper; the method comprises:

[0006] acquiring a transmitting signal of the transmitting integrated RIS and modeling a channel from the RIS to the legitimate user; calculating a received signal-to-interference-and-noise ratio (SINR) at the legitimate user and an achievable rate of the legitimate user based on the transmitting signal of the transmitting integrated RIS and the channel from the RIS to the legitimate user; calculating a received SINR at the eavesdropper and an achievable rate of the eavesdropping link based on a received communication signal at the eavesdropper;

[0007] The security capacity of the multicast system is designed according to the minimum value of the reachable rate of the legal user and the reachable rate of the eavesdropping link, and the total power consumption of the ISAC system is composed of the transmission power of the RIS and the dissipation of the related hardware components;

[0008] The utilization efficiency of the energy resource of the ISAC system is calculated by using the security capacity and the total power consumption, the optimization model of the RIS-aided ISAC system is constructed by taking the utilization efficiency of the energy resource of the ISAC system as an objective function and by using the pre-set reachable rate constraint of the legal user, the pre-set amplification power constraint of the RIS and the pre-set CRB constraint estimated by the reconnaissance target;

[0009] The unsupervised attention residual network is constructed, the unsupervised attention residual network is trained and verified by using the cosine annealing LR scheduling algorithm and the pre-set loss function, and the trained supervised deep neural network is obtained;

[0010] The RIS phase shift and the power allocation parameter are obtained by solving the RIS-aided ISAC system optimization model according to the trained supervised deep neural network.

[0011] The above-mentioned security energy efficiency maximization method based on the unsupervised attention residual network firstly lays a foundation through accurate channel modeling and signal analysis. The RIS transmission signal is obtained and the channel from the RIS to the legal user is modeled, and the signal-to-interference-and-noise ratio and the reachable rate of the legal user and the eavesdropper are calculated, respectively, to provide accurate data support for the security capacity design. The multicast security capacity designed in this way, combined with the total power consumption calculation of the RIS transmission power and the hardware dissipation, can accurately quantify the energy utilization efficiency and avoid the error accumulation of the traditional evaluation method. Taking the energy utilization efficiency as the target, the legal user rate, the RIS power and the CRB constraint are included, which not only guarantees the communication security and the sensing accuracy, but also avoids the resource waste caused by unconstrained optimization. This solves the complexity problem caused by high-dimensional optimization in traditional mathematical methods, and the constraint condition also reduces the dimension of the solution space. By introducing the unsupervised attention residual network framework, the SEE maximization of the ISAC system in the dynamic environment can be realized without pre-labeled data. The double-channel attention mechanism effectively extracts key channel features, the residual connection guarantees the gradient flow of the deep network, and the dynamic penalty mechanism adaptively balances the SEE optimization and the constraint satisfaction. Compared with the traditional zero forcing (ZF) method, the method can significantly reduce the computational complexity while guaranteeing similar SEE performance, and is suitable for large-scale RIS deployment in dynamic vehicle scenarios, achieving a balance between computational efficiency and optimization performance. The SEE maximization in the dynamic vehicle environment can be realized, and the technical problems of traditional methods such as high computational complexity and dependence on labels are solved. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 It is a schematic diagram of the RIS-aided ISAC system in one embodiment;

[0013] Figure 2 Flowchart of a security energy efficiency maximization method based on unsupervised attention residual network in an embodiment;

[0014] Figure 3 Module diagram of an unsupervised attention residual network in an embodiment;

[0015] Figure 4 Module diagram of a multi-task output in another embodiment. DETAILED DESCRIPTION

[0016] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0017] The security energy efficiency maximization method based on unsupervised attention residual network provided by the present application can be applied to an RIS-aided ISAC system as shown in Figure 1 The RIS-aided ISAC system contains a TX-RIS with M reflecting elements, K legitimate users, L sensing targets and a single-antenna eavesdropper. The TX-RIS dynamically optimizes the reflection phase shift and power allocation through unsupervised learning, maximizes the SEE under the premise of meeting the CRB constraint and communication rate requirement.

[0018] In an embodiment, as shown in Figure 2 A security energy efficiency maximization method based on unsupervised attention residual network is provided, comprising the following steps:

[0019] In step 202, the transmission signal of the transmission integrated RIS is obtained, and the channel from the RIS to the legitimate user is modeled. The signal-to-interference-and-noise ratio received at the legitimate user and the achievable rate of the legitimate user are calculated based on the transmission signal of the transmission integrated RIS and the channel from the RIS to the legitimate user. The signal-to-interference-and-noise ratio received at the eavesdropping legitimate user and the achievable rate of the eavesdropping link are calculated based on the communication signal received at the eavesdropper.

[0020] As shown in Figure 1 The RIS-aided ISAC system contains a TX-RIS with M reflecting elements, which integrates a digital phased array and a radio frequency power distribution layer using a forced feeding architecture, and can transmit communication signals and sensing signals simultaneously.

[0021] Suppose the strong feeding type RIS has M arrays, then the phase shift vector of the RIS can be represented as:

[0022] (1)

[0023] wherein, denotes the phase shift coefficient of the mth element of the RIS.

[0024] The transmit signal of the TX-RIS is denoted as:

[0025]

[0026] wherein, , denote the communication signal transmitted by the RIS to the user and the sensing signal to the target, respectively, , denote the transmit power allocated by the RF-fed antenna to the user and the target, respectively, , denote the phase shift vectors used by the RIS for communication and sensing, respectively, and both satisfy equation (1).

[0027] In real scenarios, the user is usually located beyond the Rayleigh distance of the transmitter. Due to the existence of LoS component and NLoS component in far-field plane wave propagation, the channel from the RIS to the legitimate user k can be modeled as a Rician fading channel, expressed as follows:

[0028]

[0029] wherein, denotes the path loss at a reference distance of 1 meter, denotes the corresponding path loss exponent, denotes the distance between the RIS and the user k, denotes the Rician factor of the RIS-user link, is the deterministic LoS component, is the NLoS component modeled as Rayleigh fading. The Array Response Vector (ARV) can be denoted as:

[0030]

[0031] wherein, and denote the elevation angle and the azimuth angle, respectively, is the inter-element spacing, is the carrier wavelength, and denotes the Kronecker product. The expression derivation is the same as .

[0032] The signal received by the user k can be denoted as:

[0033]

[0034] wherein, represents a user at a noise power of an additive white Gaussian noise.

[0035] The received signal-to-interference-and-noise ratio (SINR) at the kth legitimate multicast user terminal can be expressed as:

[0036]

[0037] The achievable rate for the kth user can be expressed as:

[0038]

[0039] In the multicast communication scenario, the minimum communication rate is defined as the achievable rate of the worst user in the set of communicating users . Since multicast transmission requires all legitimate users to correctly receive the same common information, the performance of the ISAC system is limited by the worst user in terms of link quality. Therefore, the minimum communication rate directly reflects the minimum communication performance of the multicast system and is a key indicator to guarantee the fairness and reliability of multi-user communication. The minimum communication rate is the minimum value of the achievable rates of the users , which can be expressed as:

[0040]

[0041] The received communication signal at the eavesdropper can be expressed as:

[0042]

[0043] wherein, represents an AWGN with a noise power of and both satisfy , wherein represents the Boltzmann constant, represents the noise temperature, represents the noise bandwidth.

[0044] The received signal-to-interference-and-noise ratio (SINR) at the eavesdropper terminal can be expressed as:

[0045]

[0046] The achievable rate of the eavesdropping link can be expressed as:

[0047]

[0048] ​​Step 204, design the security capacity of the multicast system according to the minimum value of the reachable rate of the legitimate user and the reachable rate of the eavesdropping link, and the total power consumption of the ISAC system is composed of two parts of the transmission power of the RIS and the dissipation of the related hardware components.

[0049] The security capacity of the multicast system is:

[0050]

[0051] Each RIS element can be regulated by the controller through the phase, so there is dynamic power consumption for each element when the RIS works, the more the number of elements, the higher the phase adjustment degree of freedom (DoF), and the disadvantage is that the energy consumption increases. The total power consumption of the ISAC system can be represented by the transmission power of the RIS and the dissipation of the related hardware components, which can be represented as:

[0052] wherein, represents the dynamic power consumption of each RIS unit, represents the static power required for the RIS to maintain basic circuit operation, represents the total consumption power of the user's hardware.

[0053] Step 206, calculate the energy resource utilization efficiency of the ISAC system by using the security capacity and the total power consumption, take the energy resource utilization efficiency of the ISAC system as the objective function, and use the pre-set reachable rate constraint of the legitimate user, the RIS amplification power constraint and the CRB constraint estimated by the reconnaissance target to build an optimization model of the RIS-assisted ISAC system.

[0054] SEE is a key indicator for measuring the security performance of a wireless communication system under limited energy resources, and is defined as the security capacity per unit energy consumption, that is, the ratio of the security capacity of the ISAC system to the total power consumption, which represents the energy resource utilization efficiency of the ISAC system. The specific expression is as follows:

[0055]

[0056] Suppose there are L targets, which are located at different angles wherein, represents the azimuth angle, represents the elevation angle, and the target echo signal received by the RIS can be represented as:

[0057]

[0058] wherein, is the Radar Cross Section (RCS) of each target, is the array response vector, represents the AWGN with zero mean and covariance matrix The RCS of all targets follows the Swerling-II model, which is constant within each pulse duration but varies between pulses.

[0059] In radar systems, the theoretical lower bound of the accuracy of target azimuth parameter estimation is determined by the CRB, which is defined as the inverse matrix of the Fisher Information Matrix (FIM). The meaning of CRB is the theoretical minimum of the MSE of the parameter estimation method that is unbiased or asymptotically unbiased. In this scenario, both azimuth and elevation angle parameter estimation are considered, so the FIM of the i-th target can be expressed as:

[0060]

[0061] where The partial derivative term is:

[0062] ;

[0063] ;

[0064] where is the array element position matrix , the position of the i-th array element is , , , is the Hadamard product, and similarly .

[0065] The CRB is the inverse matrix of the FIM. Specifically, the CRB for the i-th reconnaissance target DOA estimation can be expressed as:

[0066] The diagonal elements of the CRB matrix are the lower bounds of the estimation variances of

[0067] and , i.e.:

[0068]

[0069] To ensure the performance of radar target positioning, the CRB of the azimuth and elevation angle of each target should not exceed the threshold , i.e.:

[0070] ​​

[0071] In the proposed reconnaissance communication integrated system, in order to balance the system performance and power consumption, the SEE is set as the optimization target. Under the premise of meeting the achievable rate requirements of the legal multicast users, the RIS amplification power constraint and the CRB constraint of the reconnaissance target estimation, the phase shift vector of the RIS and the power distribution coefficient of the feed antenna are jointly optimized to maximize the SEE of the ISAC system, so that the optimization problem can be expressed as:

[0072] ;

[0073] ,

[0074] ,

[0075] ,

[0076]

[0077] Among them, constraint (C1) limits the power of RIS transmission, constraint (C2) ensures that the minimum achievable rate of the legal multicast user k must not be lower than the rate threshold , constraint (C3) represents the phase shift angle constraint and unit modulus constraint of each element of RIS, and constraint (C4) represents the CRB constraint of the DOA estimation of the reconnaissance target.

[0078] Step 208, constructing an unsupervised attention residual network, using a cosine annealing LR scheduling algorithm and a pre-set loss function to train and verify the unsupervised attention residual network, and obtaining a trained supervised deep neural network.

[0079] The unsupervised attention residual network is constructed, the main structure of the network takes layered feature processing as the core and fuses double CAM and deep residual connection, which is composed of four stages of input layer processing and feature preprocessing, double CAM, deep residual feature extraction, and final feature integration and multi-task output layer, realizes the E2E nonlinear mapping of high-dimensional channel features to RIS phase control and power distribution, and the specific architecture of each module of the network structure is as shown in Figure 3 .

[0080] (1) Input layer processing and feature preprocessing:

[0081] The network input is a 384-dimensional channel feature vector spliced from the channel information of the user, eavesdropper and target. Specifically, when the number of users K=2, the number of eavesdroppers E=1, and the number of RIS elements M=64, the channel vector of each entity contains 64-dimensional real and 64-dimensional imaginary parts, the single-entity channel dimension is 2*64=128, and the total input dimension is 128*(2+1)=384. The input layer keeps the dimension unchanged through a linear layer, and unifies the feature space through linear transformation, which is convenient for subsequent nonlinear processing. Batch normalization (BN) is applied synchronously to stabilize the feature distribution and alleviate the gradient vanishing problem to accelerate the training convergence. For example, the real and imaginary parts of the channel features may have different numerical ranges due to path loss differences. BN standardizes them to have a mean of 0 and a variance of 1. Let the network input be Then the output after the linear layer and BN is as follows:

[0082]

[0083] where, is the weight, is the bias.

[0084] Subsequently, a Leaky Rectified Linear Unit (LeakyReLU) with a slope of 0.1 is used to introduce non-linear mapping, avoiding the "death" problem of traditional ReLU neurons and improving the learning ability of negative sample features. After LeakyReLU, the output becomes:

[0085]

[0086] Finally, a 30% probability of random inactivation (Dropout) is used to randomly zero the neuron output, suppressing overfitting and enhancing the model's generalization ability, outputting a 384-dimensional preprocessed feature vector.

[0087] In summary, the input layer processing and feature preprocessing function is to map multi-dimensional channel features (real and imaginary parts of user channel, eavesdropper channel, and target channel) to a high-dimensional feature space, preparing for subsequent processing, where, represents the channel from the RIS to the lth sensing target.

[0088] (2) First CAM:

[0089] For the 384-dimensional pre-processed features, the CAM module extracts the mean and extreme value statistical features of the channel level through the Global Average Pooling (GAP) and Global Max Pooling (GMP) operations, respectively, to form a 384-dimensional pooling feature vector. After two layers of linear transformation (compression ratio 4, i.e. 384→96→384) and Sigmoid activation function, a channel weight vector with a range of 0-1 is generated. The weight vector is multiplied with the input features channel by channel to achieve feature enhancement of key channels, such as strengthening the user LoS propagation path and the high-gain channel corresponding to the target strong scattering path, and suppressing the interference of noise and NLoS path. Through this operation, the network focuses on the feature dimensions that have a significant impact on communication perception performance, and outputs 384-dimensional focused features to provide high discriminative input for deep networks.

[0090] Channel attention weight matrix The calculation formula is as follows:

[0091]

[0092] wherein, is a Sigmoid function, , W i is a learnable weight, and represent GAP and GMP features, respectively.

[0093] Due to the high dimension of channel features and the different degrees of influence of different channels on the optimal configuration, it is necessary to adaptively focus on important features, so it is necessary to introduce an attention mechanism, which has the following effects: first, feature importance evaluation is performed by extracting global feature statistical information through GAP and GMP; second, adaptive weight allocation is performed to assign different importance weights to different channel features; third, key information is enhanced to highlight channel features that are more important to RIS phase and power allocation; fourth, noise is suppressed to reduce the influence of unimportant features on decision-making. In the ISAC system, it can be considered that the channel features of different users, targets and eavesdroppers have different contributions to the optimal configuration, and the attention mechanism can automatically learn this importance difference.

[0094] (3) Deep residual feature extraction module:

[0095] The network stack 4 progressive residual blocks (RBs) constitute a ResNet, which compresses the feature dimension and enhances the semantic expression layer by layer through the "dimension reduction-processing-jump connection" mechanism. The structure and function of each block are as follows: the input of RB1 is 384 dimensions, and the output is 512 dimensions. After the first linear layer (384→512), it is followed by BN, LeakyReLU and 30% Dropout, which reduces the feature dimension and introduces nonlinearity and regularization; the second linear layer (512→512) repeatedly processes to extract the middle layer features. Since the input and output dimensions are inconsistent, the residual connection maps the original input to the target dimension through a linear layer (384→512), and adds it to the main path output, which retains the basic features and reduces the risk of gradient disappearance.

[0096] The input of RB2 is 512 dimensions, and the output is 384 dimensions. The Dropout rate is reduced to 25%, which relieves the overfitting pressure. Two linear transformations (512→384→384) combined with a jump connection (512→384 linear layer) further compress the feature space and strengthen the sparse expression ability of the features.

[0097] The input of RB3 is 384 dimensions, and the output is 256 dimensions. The Dropout rate is 20%, and through two linear layers (384→256→256) and a jump connection (384→256 linear layer), the features containing spatial orientation, path loss and other middle layer semantic information are extracted, which enhances the adaptability of the model to complex channel environment.

[0098] The input of RB4 is 256 dimensions, and the output is 192 dimensions. The Dropout rate is 10%, and two linear layers (256→192→192) and a jump connection (256→192 linear layer) focus on the core features, balance the calculation efficiency and representation ability, and output high-level semantic features (such as BF direction, target angle) required for RIS phase-power joint optimization.

[0099] The output of the qth RB (q=1,2,3,4) is:

[0100]

[0101] Where, Shortcut is a 1×1 convolution (when the dimensions do not match).

[0102] The functions of RB are as follows: 1) promoting gradient flow optimization, solving the gradient disappearance problem of deep network through jump connection; 2) realizing feature reuse, retaining low-level features while learning high-level abstraction; 3) improving network stability: improving training stability and convergence speed.

[0103] (4) Secondary channel attention module:

[0104] For the 192-dimensional high-level features output by RB, CAM is applied again (compression ratio 4). Channel weights are generated by GAP, GMP, linear transformation (192→48→192) and Sigmoid activation, focusing on enhancing feature channels that are strongly related to SEE, communication rate constraint and perception accuracy (such as amplitude difference of user eavesdropper channel and phase gradient of target orientation vector), further filtering redundant information, and outputting 192-dimensional enhanced features.

[0105] (5) Final feature integration and multi-task output layer:

[0106] The feature integration layer reduces the feature dimension to 128 dimensions through a linear layer (192→128), BN and LeakyReLU, forming a compact comprehensive feature representation to adapt to the dimension requirement of subsequent multi-task output.

[0107] Final feature compression and multi-task output are:

[0108]

[0109] Each task is independently output as:

[0110]

[0111] The three-task parallel output module is as follows: Figure 4 , which is specifically introduced as follows:

[0112] Communication phase shift The 64-dimensional real-valued phase parameters are output by a linear layer (128→64). Since the network output does not naturally satisfy the unit modulus constraint , it is necessary to convert it into the form of using Euler's formula to ensure that the phase of each RIS element satisfies the unit modulus constraint, thereby achieving BF optimization for user communication links and maximizing the received signal strength of legitimate users. Perception phase shift The 64-dimensional real-valued phase parameters are output by a linear layer (128→64), which are converted into complex phases in the same way to control the perception sensitivity of RIS to target azimuth and elevation angles and improve the accuracy of target angle estimation. Power allocation The unnormalized power coefficients are output by a linear layer (128→2), and a two-dimensional probability vector (sum = 1) is generated by Softmax to dynamically allocate communication power and perception power under the total power constraint to balance the resource investment of dual tasks and avoid excessive consumption of resources for a single task. The definition of Softmax function is as follows: given an input vector , the i-th output of Softmax is:

[0113] .

[0114] During model training, the linear layer is initialized with Kaiming Normal initialization (adapted to LeakyReLU activation function with parameter a=0.1), the initial linear increase of LR is used to accelerate convergence, and the cosine curve decay of LR is used in the later stage to avoid model oscillation or difficulty in convergence due to too high LR, and the BN layer weight is initialized to 1 and the bias is 0 to ensure stable gradient in the initial training. Gradient clipping (norm ≤ 1.0) is performed on model parameters during training to suppress gradient explosion; cosine annealing LR scheduling with 5 warm-up rounds is used, the initial linear increase of LR is used to accelerate convergence, and the cosine curve decay of LR is used in the later stage to achieve smooth decay of LR to improve convergence quality. The formula of cosine annealing LR scheduling with 5 warm-up rounds is:

[0115]

[0116] wherein, is the initial LR, is the warm-up round number 5, T is the total training round number, and t is the current round number.

[0117] Through residual connection to alleviate gradient vanishing by cross-layer feature reuse, key features are strengthened by dual channel attention, and multi-layer Dropout (0.1~0.3) is used to suppress overfitting, so that the network can handle the multi-objective conflict of communication rate constraint, perception CRB constraint and SEE optimization maximization at the same time, and realize the joint optimization of RIS phase and power allocation.

[0118] The design idea of the loss function is to maximize SEE under the premise of meeting the communication rate and CRB constraints. In DL, the model usually optimizes parameters by minimizing the loss function. In order to convert the original maximization SEE problem into a minimization problem that can be processed by the DL framework, the negative value of the objective function is taken as the loss function, and at the same time, by adding a penalty term, the constraint condition is embedded into the objective optimization process to form an unsupervised learning framework. The mathematical expression of the loss function is: i.e.

[0119]

[0120] wherein, S is the mini-batch size (mbs) in the training stage, is the SEE, and are the constraint violation amounts of communication rate and CRB, respectively, and The loss function is the average of the rounds, and the penalty coefficient is updated based on the average violation of the round. By taking the negative of the objective function, the model can approximate the optimal phase and power allocation by minimizing the loss function without labels, avoiding the high cost of obtaining labels in supervised learning (such as generating the optimal phase and power allocation through exhaustive search). This design embeds physical layer constraints into the neural network optimization process, enabling the network to autonomously learn the nonlinear mapping of channel characteristics and optimal resource allocation without enumerating the feasible solution space, significantly improving optimization efficiency under complex constraints. During training, the loss function calculates the gradient through the PyTorch automatic gradient mechanism, BP to the parameters of each layer of the network (such as linear layer weights, BN parameters), and updates the parameters through the AdamW optimizer. Gradient clipping ensures training stability and prevents gradient explosion.

[0121] To meet the requirement that the rate of legal users is not lower than the threshold , the ReLU function is used to calculate the communication rate constraint violation , and only when the rate is not up to standard, the penalty is applied to ensure the minimum communication rate requirement.

[0122] The diagonal elements of the CRB matrix are extracted as the CRB of azimuth and elevation, and the maximum CRB of all targets is finally taken as the constraint index:

[0123]

[0124] The constraint violation is:

[0125]

[0126] To balance the relationship between optimization objectives and constraint conditions, a dynamic adjustment mechanism is designed to update the penalty term coefficients and : the initial value is set to 1 and is adjusted adaptively through training, and the update rule is , (where LR eta =0.01) to gradually increase the penalty weight for violating constraints in a gradient descent manner, while limiting it to the interval [0.1, 10] to avoid excessive punishment leading to unstable training and ensure that the optimization process searches within the feasible solution space and gradually approaches the optimal solution.

[0127] The parameters optimized in the training process are divided into two categories, including neural network parameters and dynamic penalty term coefficients, which are introduced as follows: Neural network learnable parameters: first, linear layer weights and biases, such as those in the input layer, RB, and linear layers in the output layer, which learn the non-linear mapping from channel features to the optimal configuration through BP; second, BN parameters, such as the scaling factor γ and the offset factor β of each layer BN, which are used to stabilize the feature distribution and accelerate convergence; third, channel attention module parameters, in which the FC weights are used to generate a channel weight vector to strengthen key features (such as LoS path channel features). Dynamic penalty term coefficients, including and They are dynamically adjusted through gradient descent to balance the optimization objective and the constraint condition. The initial value is 1.0, and the limit is [0.1, 10.0] to avoid constraint violation or excessive punishment.

[0128] In the iterative optimization process of the unsupervised learning model, FP and BP form a closed-loop feedback mechanism. In the FP stage, high-dimensional channel features are first input into the network, and through the attention module, global feature statistics are extracted through average pooling and max pooling to adaptively learn the contribution differences of different channels (user, target, and eavesdropper channels) to RIS phase and power allocation, assign high weights to key features to enhance information expression, and suppress the interference of noise features; then, RB builds a gradient flow shortcut through a skip connection to alleviate the problem of gradient vanishing in deep networks, while realizing the reuse of low-level detailed features and high-level abstract features, improving training stability and convergence efficiency, and finally directly mapping the continuous configuration scheme of RIS phase and power in the output layer. In the BP stage, the violation amount of SEE, communication rate constraint, and CRB constraint is used as the core to build the loss function, a dynamic penalty term (gradient adaptive update penalty coefficient based on constraint violation amount) is introduced, the gradient of each layer parameter is reversely deduced through the chain rule, and the weight matrix of the attention module and the connection parameters of RB are iteratively updated using the optimizer, so that the model can gradually learn the non-linear mapping relationship between channel features and RIS optimal configuration in the scene without labeled data, realizing E2E joint optimization. In the BP process, the loss function calculates the partial derivative of each network parameter, starting from the output layer and calculating the gradient layer by layer. The gradient of the output layer is directly determined by the partial derivative of the output by the loss function, while the gradient of the hidden layer (such as the attention module and RB) is determined by the gradient passed from the next layer and the derivative of the current layer activation function - for example, the derivative of the Sigmoid activation function in the attention module and the derivative of the LeakyReLU in the RB, which will multiply the error signal BP to the front end of the network through the chain, providing accurate gradient direction for the optimizer to update the fully connected weights of the attention module and the linear transformation parameters of the RB.

[0129] The overall process is as follows: in the training stage, data preparation is first performed, and a legitimate user and an eavesdropper are generated in a three-dimensional scene to generate a Rician channel dataset, and the real and imaginary parts of each channel vector are extracted and combined into a feature vector After real and imaginary part decomposition and normalization processing, Gaussian noise is added to the training set to enhance the data generalization ability, and the training set and the validation set are divided in proportion. Then an unsupervised learning network is constructed, which includes an input preprocessing layer, a double channel attention module, a residual block and a multi-task output layer. The input preprocessing layer performs linear transformation, batch normalization and LeakyReLU activation on the features, the double attention module extracts key channel features through pooling operation, the residual block uses jump connection to keep the integrity of the features, and the output layer generates RIS phase shift and power allocation parameters. After entering the iterative training process, all training batches are traversed in each training cycle (epoch): forward propagation generates RIS communication phase shift, perception phase shift and power allocation coefficient, where the phase is converted into a unit modulus complex number through a trigonometric function, and the power coefficient is normalized through Softmax; the SEE, communication rate and CRB constraint violation of the ISAC system are calculated synchronously, and the penalty term coefficient is dynamically updated to balance the optimization objective and the constraint condition; based on the total loss function, back propagation is performed, the gradient norm is constrained within 1 through gradient clipping to stabilize the training, and the learning rate is adjusted using a cosine annealing strategy with a 5-round warm-up period. The validation set loss is continuously monitored during training. If the loss improvement amplitude is less than 0.0001 for 30 consecutive cycles, the early stopping mechanism is started to prevent overfitting.

[0130] In the verification stage, each validation sample is processed through an infinite loop: first, construct the channel feature vector H of the current scene, and generate RIS configuration parameters and power allocation scheme through forward inference of the trained network; then calculate the communication rate, eavesdropping rate, secrecy rate, CRB matrix and SEE, etc. At the same time, it is verified whether the communication rate meets the threshold and the CRB is lower than the set threshold; if the constraint conditions are met, the current configuration is applied and the SEE is recorded, otherwise the model retraining process is triggered to adapt to dynamic channel changes, ensuring the effectiveness and generalization ability of the ISAC system in actual application.

[0131] Step 210, according to the trained supervised deep neural network, the RIS-aided ISAC system optimization model is solved, and the RIS phase shift and power allocation parameters are obtained.

[0132] The above safety energy efficiency maximization method based on unsupervised attention residual network first lays the foundation through accurate channel modeling and signal analysis. The RIS transmission signal is obtained and the channel from the RIS to the legitimate user is modeled, and the signal-to-interference-and-noise ratio (SINR) and achievable rate of the legitimate user and the eavesdropper are calculated respectively to provide accurate data support for the design of the secure capacity. The multicast secure capacity designed in this way, combined with the total power consumption calculation of the RIS transmission power and the hardware dissipation, can accurately quantify the energy utilization efficiency and avoid the error accumulation of the traditional evaluation method. With the energy utilization efficiency as the target, the legitimate user rate, RIS power and CRB constraint are included, which not only guarantees the communication security and sensing accuracy, but also avoids the waste of resources caused by unconstrained optimization. This solves the complexity problem caused by high-dimensional optimization in traditional mathematical methods, and the constraint condition also reduces the dimension of the solution space. By introducing the unsupervised attention residual network framework, the SEE maximization of the ISAC system in a dynamic environment can be realized without pre-labeled data. The double-channel attention mechanism effectively extracts key channel features, the residual connection ensures the gradient flow of the deep network, and the dynamic penalty mechanism adaptively balances SEE optimization and constraint satisfaction. Compared with the traditional zero forcing (ZF) method, this method significantly reduces the computational complexity while ensuring similar SEE performance, is suitable for large-scale RIS deployment in dynamic vehicle scenarios, and balances the computational efficiency and optimization performance. It can maximize SEE in a dynamic vehicle environment and solve the technical problems of traditional methods such as high computational complexity and dependence on labels.

[0133] In one embodiment, the received signal-to-interference-and-noise ratio (SINR) at the legitimate user and the achievable rate of the legitimate user are calculated based on the transmitted integrated RIS transmission signal and the channel from the RIS to the legitimate user, including:

[0134] The received signal-to-interference-and-noise ratio (SINR) at the legitimate user and the achievable rate of the legitimate user are calculated based on the transmitted integrated RIS transmission signal and the channel from the RIS to the legitimate user, respectively:

[0135]

[0136] wherein, is the received signal-to-interference-and-noise ratio (SINR) at the kth legitimate multicast legitimate user end, is the conjugate transpose of the channel from the RIS to the legitimate user k, and the superscript H represents the conjugate transpose, , respectively represent the target transmission power allocated to the legitimate user and the target transmission power by the RF feed antenna, , respectively represent the phase shift vectors of the RIS for communication and sensing, is the noise power at the legitimate user end;

[0137] Then the achievable rate of the kth legal user is calculated as:

[0138] .

[0139] In one embodiment, the communication signal received by the eavesdropper is

[0140]

[0141] in, 、 They represent the communication signals sent by RIS to legitimate users and the perception signals sent to the target, 、 denote the transmit power allocated to the legitimate user and the target by the RF feeding antenna, respectively, represents the phase shift vector used by RIS for communication, The noise power is AWGN, Represents the eavesdropper channel.

[0142] In one embodiment, calculating the signal-to-interference-and-noise ratio received by the eavesdropped legitimate user terminal and the achievable rate of the eavesdropped link based on the communication signal received at the eavesdropper includes:

[0143] Based on the communication signal received by the eavesdropper, the signal-to-interference-and-noise ratio received by the eavesdropping legitimate user terminal and the achievable rate of the eavesdropping link are calculated as follows:

[0144] ;

[0145] ;

[0146] in, 、 denote the transmit power allocated to the legitimate user and the target by the RF feeding antenna, respectively, 、 denote the phase shift vectors used by RIS for communication and perception, respectively. For eavesdroppers The noise power, represents the eavesdropper channel, Indicates the signal-to-interference-and-noise ratio (SINR) received by the legitimate eavesdropping user.

[0147] In one embodiment, calculating the efficiency of energy resource utilization of the ISAC system using the safety capacity and the total power consumption includes:

[0148] The energy resource utilization efficiency of the ISAC system is calculated using the safety capacity and total power consumption:

[0149]

[0150] in, represents the dynamic power consumption of each RIS unit, It represents the static power required by RIS to maintain basic circuit operation. Indicates the total power consumption of the hardware of legitimate users, represents the safety capacity, Indicates the total power consumption, 、 denote the transmit power allocated to the legitimate user and the target by the RF feeding antenna, respectively, represents the achievable rate of the eavesdropping link, Indicates the achievable rate of legitimate users.

[0151] In one embodiment, the ISAC system's energy resource utilization efficiency is used as the objective function, and a RIS-assisted ISAC system optimization model is constructed using pre-set legitimate user reachable rate constraints, RIS amplification power constraints, and reconnaissance target estimation CRB constraints:

[0152]

[0153] ,

[0154] ,

[0155] ,

[0156]

[0157] in, Indicates the efficiency of ISAC system in utilizing energy resources, represents the safety capacity, Indicates the total power consumption, 、 denote the transmit power allocated to the legitimate user and the target by the RF feeding antenna, respectively, represents the achievable rate of the kth legal user, Indicates the CRB for DOA estimation of reconnaissance targets, Indicates the actual power transmitted by RIS, Indicates the maximum threshold of transmit power. represents the unit mode phase shift, Indicates the number of RIS elements, Indicates the maximum CRB threshold, Indicates the number of perceived targets, 、 denote the phase shift vectors used by RIS for communication and perception, Rk is the minimum achievable rate of the kth legitimate user.

[0158] In one embodiment, an unsupervised attention residual network is constructed, including an input preprocessing layer, a double channel attention module, a residual block, and a multi-task output layer. The input preprocessing layer is used for linear transformation, batch normalization, and LeakyReLU activation of the features. The double attention module is used to extract key channel features through a pooling operation. The residual block is used to maintain feature integrity using a skip connection. The output layer is used to generate RIS phase shift and power allocation parameters.

[0159] In one embodiment, the unsupervised attention residual network is trained and verified using a cosine annealing LR scheduling algorithm and a pre-set loss function, including:

[0160] In a three-dimensional scene, a Rician channel dataset containing legitimate users and eavesdroppers is generated. The real and imaginary parts of each channel vector are extracted and combined into a feature vector. After real and imaginary part decomposition and normalization, Gaussian noise is added to the training set and the training set and validation set are divided in proportion. The iteration training process is entered. In each training cycle, all training batches are traversed, and the RIS communication phase shift, perception phase shift, and power allocation coefficient are generated by forward propagation. The phase is converted to a unit modulus complex number by a trigonometric function, and the power coefficient is normalized by Softmax. The safety energy efficiency, communication rate, and violation amount of the CRB constraint of the ISAC system are calculated simultaneously, and the penalty term coefficient is dynamically updated to balance the optimization objective and the constraint condition.

[0161] Based on the pre-set loss function, back propagation is performed, the gradient norm is constrained within 1 by gradient clipping to stabilize training, and the learning rate is adjusted using a cosine annealing strategy with a 5-round warm-up period. When the pre-set training termination condition is reached, the training is stopped.

[0162] In one embodiment, the cosine annealing strategy with a 5-round warm-up period includes a cosine annealing LR scheduling formula with a 5-round warm-up. The cosine annealing LR scheduling formula with a 5-round warm-up is

[0163]

[0164] wherein, is the initial LR, is the warm-up round number 5, T is the total training round number, and t is the current round number.

[0165] In one embodiment, the pre-set loss function is:

[0166]

[0167] wherein, S is the small batch size in the training stage, is the safety energy efficiency, and are the constraint violation amounts of the communication rate and the CRB, respectively, and are the coefficients of the penalty terms adjusted dynamically.

[0168] It should be understood that although the steps in the flowcharts of Figure 1 are shown in sequential order, such steps are not necessarily performed in the order shown. Unless explicitly stated, the steps of the methods described herein are not necessarily performed in the order described. Furthermore, Figure 1 at least some of the steps in the flowcharts of may include multiple sub-steps or stages, which are not necessarily performed at the same time, and which can be performed in different orders, in parallel or in an alternating manner.

[0169] The technical features of the above embodiments can be combined in any manner. For brevity, not all possible combinations of the technical features in the above embodiments are described, but it should be understood that any combination of the technical features is within the scope of the present disclosure, as long as the combination does not result in contradictions.

[0170] The above embodiments merely express several implementation manners of the present application, and the description is relatively specific and detailed, but should not be construed as a limitation on the scope of the present application. It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these are within the scope of the present application. Therefore, the scope of the present application should be subject to the appended claims.

Claims

1. A method for maximizing security and energy efficiency based on unsupervised attention residual networks, characterized by: The method is applied to an ISAC system including a transmission integrated RIS configured with a reflection unit, a legitimate user, a sensing target, and an eavesdropper's RIS assistance; the method comprises: Acquire a transmission signal of the transmitting integrated RIS and model a channel from the RIS to the legitimate user; calculate a signal-to-interference-plus-noise ratio (SINR) received at the legitimate user and a achievable rate for the legitimate user based on the transmission signal of the transmitting integrated RIS and the channel from the RIS to the legitimate user; and calculate a signal-to-interference-plus-noise ratio (SINR) received at the eavesdropped legitimate user end and a achievable rate for the eavesdropped link based on a communication signal received at the eavesdropper; The security capacity of the multicast system is designed based on the minimum achievable rate of the legitimate users and the achievable rate of the eavesdropping link, and the total power consumption of the ISAC system is composed of two parts: the transmit power of the RIS and the dissipation of related hardware components; The ISAC system's energy resource utilization efficiency is calculated using the security capacity and the total power consumption. A RIS-assisted ISAC system optimization model is constructed using the ISAC system's energy resource utilization efficiency as an objective function and pre-set legitimate user reachable rate constraints, RIS amplification power constraints, and reconnaissance target estimation CRB constraints. Constructing an unsupervised attention residual network, training and verifying the unsupervised attention residual network using a cosine annealing LR scheduling algorithm and a preset loss function to obtain a trained supervised deep neural network; The RIS-assisted ISAC system optimization model is solved according to the trained supervised deep neural network to obtain RIS phase shift and power allocation parameters.

2. The method according to claim 1, characterized in that Calculating a signal-to-interference-plus-noise ratio received at a legitimate user and a achievable rate of the legitimate user based on a transmission signal of the transmit integrated RIS and a channel from the RIS to the legitimate user, including: The signal-to-interference-plus-noise ratio (SINR) received at the legal user and the achievable rate of the legal user are calculated based on the transmission signal of the transmission integrated RIS and the channel from the RIS to the legal user, respectively: in, is the signal-to-interference-and-noise ratio received at the kth legal multicast legal user terminal, is the conjugate transpose of the channel from RIS to the legitimate user k, and the superscript H represents the conjugate transpose. 、 denote the transmit power allocated to the legitimate user and the target by the RF feeding antenna, respectively, 、 denote the phase shift vectors used by RIS for communication and perception, respectively, For legitimate users Noise power at Then the achievable rate of the kth legal user is calculated as: 。 3. The method according to claim 2, characterized in that The communication signal received by the eavesdropper is in, 、 They represent the communication signals sent by RIS to legitimate users and the perception signals sent to the target, 、 denote the transmit power allocated to the legitimate user and the target by the RF feeding antenna, respectively, represents the phase shift vector used by RIS for communication, The noise power is AWGN, Represents the eavesdropper channel.

4. The method according to claim 3, characterized in that The signal-to-interference-and-noise ratio (SINR) received by the legitimate eavesdropping user and the achievable rate of the eavesdropping link are calculated based on the communication signal received by the eavesdropper, including: Based on the communication signal received by the eavesdropper, the signal-to-interference-and-noise ratio received by the eavesdropping legitimate user terminal and the achievable rate of the eavesdropping link are calculated as follows: in, 、 denote the transmit power allocated to the legitimate user and the target by the RF feeding antenna, respectively, 、 denote the phase shift vectors used by RIS for communication and perception, respectively, For eavesdroppers The noise power, represents the eavesdropper channel, Indicates the signal-to-interference-and-noise ratio (SINR) received by the legitimate eavesdropping user.

5. The method according to claim 1, wherein Calculating the utilization efficiency of energy resources by the ISAC system using the safety capacity and the total power consumption includes: The utilization efficiency of the ISAC system on energy resources is calculated using the safety capacity and the total power consumption: in, represents the dynamic power consumption of each RIS unit, It represents the static power required by RIS to maintain basic circuit operation. Indicates the total power consumption of the hardware of legitimate users, represents the safety capacity, Indicates the total power consumption, 、 denote the transmit power allocated to the legitimate user and the target by the RF feeding antenna, respectively, represents the achievable rate of the eavesdropping link, Indicates the achievable rate of legitimate users.

6. The method according to claim 1, characterized in that Taking the energy resource utilization efficiency of the ISAC system as the objective function, the RIS-assisted ISAC system optimization model is constructed using the pre-set reachable rate constraints of legitimate users, RIS amplification power constraints, and CRB constraints of reconnaissance target estimation: in, Indicates the efficiency of ISAC system in utilizing energy resources, represents the safety capacity, Indicates the total power consumption, 、 denote the transmit power allocated to the legitimate user and the target by the RF feeding antenna, respectively, represents the achievable rate of the kth legal user, Indicates the CRB for DOA estimation of reconnaissance targets, Indicates the actual power transmitted by RIS, Indicates the maximum threshold of transmit power. represents the unit mode phase shift, Indicates the number of RIS elements, Indicates the maximum CRB threshold, Indicates the number of perceived targets, 、 denote the phase shift vectors used by RIS for communication and perception, respectively, represents the minimum achievable rate of the kth legal user.

7. The method according to claim 1, characterized in that An unsupervised attention residual network is constructed, including an input preprocessing layer, a dual-channel attention module, a residual block, and a multi-task output layer. The input preprocessing layer is used to perform linear transformation, batch normalization, and LeakyReLU activation on the features. The dual-attention module is used to extract key channel features through pooling operations. The residual block is used to maintain feature integrity using skip connections. The output layer is used to generate RIS phase shift and power allocation parameters.

8. The method according to claim 1, characterized in that The unsupervised attention residual network is trained and verified using the cosine annealing LR scheduling algorithm and a preset loss function, including: A Rice channel dataset containing legitimate users and eavesdroppers is generated in a three-dimensional scene. The real and imaginary parts of each channel vector are extracted and combined into a feature vector. After real and imaginary decomposition and normalization, Gaussian noise is added to the training set and the training set and validation set are divided proportionally. The iterative training process begins, traversing all training batches in each training cycle. The RIS communication phase shift, perception phase shift, and power allocation coefficient are generated through forward propagation. The phase is converted to a unit modulus complex number using trigonometric functions, and the power coefficient is normalized using Softmax. The ISAC system's security energy efficiency, communication rate, and CRB constraint violations are simultaneously calculated, and the penalty term coefficient is dynamically updated to balance the optimization objectives and constraints. Backpropagation is performed based on a preset loss function. Gradient clipping is used to constrain the gradient norm to 1 to stabilize training. The learning rate is adjusted using a cosine annealing strategy with a 5-round warm-up period until the preset training termination condition is reached, at which point training is stopped.

9. The method according to claim 8, characterized in that The cosine annealing strategy with 5 warm-up rounds includes a cosine annealing LR scheduling formula with 5 warm-up rounds; the cosine annealing LR scheduling formula with 5 warm-up rounds is in, is the initial LR, is the number of warm-up rounds, 5 rounds, T is the total number of training rounds, and t is the current round number.

10. The method according to claim 8, characterized in that The preset loss function is: Where S is the mini-batch size during training, For safety and energy efficiency, and are the communication rate and the CRB constraint violation, and is the penalty coefficient that is adjusted dynamically.

Images