Intelligent reflecting surface enhanced security semantic spectrum efficiency maximization method

Abstract

The invention discloses an IRS enhanced security semantic spectrum efficiency maximization method, which is characterized in that under the condition that a malicious eavesdropper exists in a system, a phase shift adjustable IRS is deployed to assist base station semantic transmission so as to achieve the required security semantic expression and security semantic spectrum efficiency. By modeling a joint semantic bit, sub-channel allocation and IRS array element reflection coefficient optimization design problem, the safety semantic spectrum efficiency maximization can be realized. A DDQN-SAC-based hybrid intelligent algorithm is provided in a semantic resource allocation method, DDQN optimizes semantic bit allocation and sub-channel allocation, and SAC optimizes an IRS reflection array element reflection coefficient. Simulation results show that compared with other reference methods, the IRS enhanced security semantic spectrum efficiency maximization method provided by the invention can significantly improve the security semantic spectrum efficiency of the system, and has a good convergence effect at the same time.

Description

Technical Field

[0001] The present invention relates to a method for maximizing the secure semantic spectral efficiency enhanced by an intelligent reflecting surface, belonging to the field of communication technologies. Background Art

[0002] Semantic communication is an effective solution to alleviate network congestion and improve communication efficiency, and is a technology with broad prospects in the sixth-generation (6G) communication technology. By leveraging the powerful knowledge extraction ability of machine learning, semantic coding can use semantic symbols to represent the basic information required for task execution, effectively reducing redundant information and alleviating the problem of spectrum scarcity. However, the openness of the spectrum brings challenges of semantic eavesdropping. Different from traditional physical-layer security, secure semantic communication poses new requirements for privacy protection at the semantic layer, and how to cope with open semantic eavesdropping is a thorny problem.

[0003] Recently, the Intelligent Reflective Surface (IRS) has attracted much attention because it can reshape signal propagation in a low-cost way and is widely used to improve spectral efficiency and security performance. Specifically, the IRS is a planar array composed of numerous programmable passive reflecting elements and a micro control center. Compared with artificial interference, relay base stations, and other power amplification technologies, the IRS can manipulate signals using passive reflection, thus achieving high energy efficiency. There is still little research on IRS-assisted secure semantic communication networks, which can utilize the signal reshaping ability of the IRS to suppress semantic eavesdropping and improve secure semantic performance. The paper "Intelligent resource allocation for transmission security on IRS-assisted spectrum sharing systems with OFDM" by Lingyi Wang et al. (Physical Communication) deeply explores IRS-assisted secure spectrum sharing communication networks, where the IRS can effectively reduce secondary network interference and suppress eavesdropping intensity. The simulation results show that applying the IRS can efficiently handle interference and eavesdropping problems. L. Yan, Z. Qin et al. in "Resource Allocation for Text Semantic Communications" (IEEE Wireless Communications Letters (2022)) further studied the resource allocation problem in semantic communication. The authors proposed an optimized resource allocation scheme based on matching algorithms and exhaustive algorithms. However, in large-scale systems, traditional mathematical algorithms will bring high system latency and cannot meet the requirements of semantic communication systems for intelligence and result real-time performance. Therefore, the optimization scheme based on Deep Reinforcement Learning (DRL) can show higher real-time performance and intelligence and is better compatible with artificial intelligence-driven semantic communication networks. In the industrial and academic fields, DRL has been widely used to quickly solve large-scale complex problems. In addition, existing research has not yet involved applying the IRS in the field of secure semantic communication to enhance the secure transmission efficiency of the semantic layer. Summary of the Invention

[0004] The object of the present invention is to propose a DRL-based intelligent semantic resource allocation method aiming at the defects and deficiencies of the prior art. This method uses a hybrid deep reinforcement learning algorithm based on SAC-DDQN to maximize the secure semantic spectral efficiency. The present invention can efficiently solve non-convex optimization problems containing coupled variables and can significantly improve the secure semantic spectral efficiency of semantic communication systems.

[0005] The technical solution adopted by the present invention to solve its technical problems is: a method for maximizing the secure semantic spectral efficiency enhanced by IRS, the method comprising the following steps:

[0006] Step 1: Establish an IRS-assisted semantic communication system model;

[0007] Step 1-1: Establish a secure semantic communication system model, and propose a novel secure semantic spectral efficiency metric that jointly considers physical layer security and semantic layer security

[0008] Step 1-2: Establish an IRS-assisted semantic transmission system. Determine the positions of the base station, D users, E eavesdroppers, and IRS, model the channels between the base station and IRS, IRS and D users, IRS and E eavesdroppers, base station and D users, and base station and E eavesdroppers, and obtain the channel gains;

[0009] Step 2: Establish a problem of maximizing the secure semantic spectral efficiency;

[0010] Step 3: Transform the target problem into a Markov problem, set the base station control center as the agent, which can connect to the IRS control center, design the state space, action space, reward function, and state transition probability, and establish a reinforcement learning basic model;

[0011] Step 4: Use a hybrid algorithm based on DDQN-SAC to jointly optimize the semantic bits, subchannel allocation, and IRS element reflection coefficients to maximize the secure semantic spectral efficiency. Specifically, it includes the following steps:

[0012] Step 4-1: Based on the formed Markov problem, design a secure semantic resource allocation algorithm for DDQN-SAC;

[0013] Step 4-2: Train the semantic coding network to obtain a codebook that records the mapping relationship between physical bits and semantic vectors;

[0014] Step 4-3: In the set simulation environment, the agent interacts with the system to obtain training experience. Specifically, the agent sequentially determines the semantic bits, subchannel allocation, and IRS element reflection coefficients, verifies the effectiveness of the actions in the IRS-assisted secure semantic communication system, and calculates the reward value. The system state sequence, action selection situation, and reward value generated in each iteration are stored in the experience pool;

[0015] Step 4-4: Update the algorithm parameters. When the agent and the environment interact for a set number of times, by replaying the experience stored in the experience pool, the agent learns with the goal of minimizing the loss function, updates the network parameters through backpropagation gradients, and returns to Step 4-3 until the training ends;

[0016] Step 4-5: The algorithm converges, and each network parameter is locally saved.

[0017] Step 5: In a real communication system, semantic bits, sub-channel allocation, and IRS element reflection coefficients are obtained through the secure semantic resource allocation method based on DDQN-SAC, and the results are verified.

[0018] Beneficial effects:

[0019] 1. The present invention applies IRS to secure semantic communication. IRS can enhance the semantic task performance of legitimate users while suppressing eavesdroppers from eavesdropping on semantic information. Compared with traditional relay base station security schemes, the IRS-assisted secure semantic transmission system can enhance the secure semantic spectral efficiency in a low-power consumption manner.

[0020] 2. The present invention proposes a novel secure semantic spectral efficiency metric that jointly considers physical layer security and semantic layer security.

[0021] 3. The present invention proposes a secure semantic resource allocation algorithm based on DDQN-SAC. This algorithm can intelligently jointly optimize semantic bits, sub-channel allocation, and IRS element reflection coefficients to maximize the secure semantic spectral efficiency. Simulation results show that the secure semantic resource allocation algorithm based on DDQN-SAC of the present invention has good convergence and can efficiently handle the target non-convex problem.

[0022] 4. The present invention deploys a programmable IRS to assist the base station in semantic transmission of intelligent devices to achieve the required secure semantic communication efficiency. By modeling a joint optimization design problem of semantic bits, sub-channel allocation, and IRS element reflection coefficients, the maximum secure semantic spectral efficiency of the system is achieved. Using the intelligent algorithm based on DDQN-SAC in the resource allocation method can work in coordination with the machine learning-driven semantic communication network. Simulation results show that compared with multiple other benchmark methods, the IRS-assisted secure semantic communication system and the secure semantic resource allocation method proposed by the present invention can significantly improve the enhanced secure semantic spectral efficiency and have good convergence effects at the same time. Description of the Drawings

[0023] Figure 1 It is the system diagram of the present invention.

[0024] Figure 2 It is the graph showing the convergence performance of the secure semantic resource allocation algorithm in the present invention.

[0025] Figure 3 It is the comparison graph of the secure semantic performance of the intelligent resource allocation method of the present invention and other resource allocation methods under different numbers of IRS reflection elements.

[0026] Figure 4 Comparison chart of the secure semantic spectral efficiency of the intelligent resource allocation method of the present invention and other resource allocation methods under different numbers of IRS reflecting array elements. Detailed implementation mode

[0027] The invention will be further described in detail below in conjunction with the accompanying drawings of the specification.

[0028] Embodiment 1

[0029] As Figure 1 shown, the present invention provides an IRS-enhanced method for maximizing secure semantic spectral efficiency, which includes the following steps:

[0030] Step 1: Establish an IRS-assisted semantic communication system model;

[0031] Based on the Transformer framework, an image semantic communication system model is established. Specifically, the sender is equipped with a semantic encoder and a channel encoder, and the receiver is equipped with a semantic decoder and a channel decoder. Based on the semantic encoder, the sender extracts features from the source information to obtain semantic features, and inputs them into the channel encoder to obtain semantic signals. After being transmitted through the wireless physical channel, the semantic signals pass through the semantic decoder and the channel decoder in sequence at the receiver and are transformed into the target image. The source image can be expressed as S = [s 1 ,..., s i ,..., s l×l , where l×l is the number of image blocks, and s i represents the i-th image block in the image. The semantic encoder is expressed as E ω (·), and the semantic decoder at the legitimate user is expressed as , where ω and η d are neural network parameters. In addition, the present invention considers the existence of eavesdroppers, and the semantic decoder at the eavesdropper is expressed as , where η e is a neural network parameter.

[0032] Step 1-1: Establish a secure semantic communication system model, and propose a novel secure semantic spectral efficiency metric that jointly considers physical layer security and semantic layer security

[0033] The present invention considers semantic layer security based on semantic structural similarity and perceptual similarity. The semantic structural similarity is expressed as

[0034]

[0035] where, is the restored picture at the receiving end, is the brightness difference, is the contrast difference and For the structural differences of pictures. The perceptual similarity is expressed as

[0036]

[0037] where represents the pre-trained features of the l-th layer, and w l represents the weight factor of the l-th layer, and H l and W l represent the height and width of the picture respectively.

[0038] The semantic layer metric proposed by the present invention is defined as

[0039] φ(S,S′) = SSIM - ξLPIPS,

[0040] where ξ is the weight factor. Therefore, the achievable semantic layer security metric for the legitimate user d on the sub-channel c is

[0041]

[0042] where represents the eavesdropping effectiveness factor, represents the recovered picture at the eavesdropper. If then , otherwise

[0043] The physical layer security metric is based on the secure transmission rate. The secure transmission rate for the legitimate user d on the sub-channel c can be expressed as

[0044]

[0045] where, R d，c represents the achievable transmission rate for the legitimate user d on the sub-channel c, and R e，c represents the achievable transmission rate for the eavesdropper e on the sub-channel c. Considering both the physical layer security and the semantic layer security jointly, the achievable secure semantic spectral efficiency for the legitimate user d on the sub-channel c is

[0046]

[0047] where, |I| represents the dimension of the semantic vector, and b represents the number of bits used to represent the semantic vector.

[0048] Step 1 - 2: Establish an IRS-assisted semantic transmission system. Determine the positions of the base station, D users, E eavesdroppers, and the IRS, and model the channels between the base station and the IRS, the IRS and the D users, the IRS and the E eavesdroppers, the base station and the D users, and the base station and the E eavesdroppers to obtain the channel gains;

[0049] The channel set is expressed as The user set is represented as From the base station to the target user d, the channels to the eavesdropper and the IRS are respectively represented as h d , w e and h r , and the channels from the IRS to the target user and the eavesdropper are respectively represented as g d and g e . Define the IRS element offset as the diagonal matrix Θ, where α n represents the phase coefficient of the nth reflecting element, and φ n represents the offset coefficient of the nth reflecting element. The signal received by the legitimate user d on the sub-channel c can be expressed as The signal eavesdropped by the eavesdropper on the sub-channel c can be expressed as where n d and n e are additive white Gaussian noise. The transmission rate of the legitimate user d on the sub-channel c is expressed as The eavesdropping rate of the eavesdropper on the sub-channel c is where, W c represents the bandwidth resource, f d，c represents the beamforming transmitted by the base station, and represent the noise variance.

[0050] Step 2: Establish the problem of maximizing the secure semantic spectral efficiency;

[0051] Define ρ d，c to represent whether the dth user occupies the cth sub-channel. If the dth user occupies the cth sub-channel, then ρ d，c = 1, otherwise, ρ d，c = 0. The secure semantic spectral efficiency can be expressed as

[0052]

[0053] By jointly optimizing the semantic bits, sub-channel allocation, and IRS element reflection coefficients, the system objective is to maximize γ.

[0054] Constraints of the resource allocation problem: The first constraint is the limit on the number of bits used by each user to represent semantics, which can be expressed as b d ∈{b min ,b min +1,…,b max}. The second constraint is the IRS reflection factor constraint, which can be expressed as 0 ≤ |[Ψ] n,n | ≤ 1 and φ n ∈{0,2π}, where [Ψ] n,nRepresents the diagonal elements of the reflection array element. The third constraint is the user-occupied subchannel constraint. Each user and subchannel only have an occupied or unoccupied relationship, and each user can only occupy one channel, that is and

[0055] Step 3: Transform the target problem into a Markov problem. Set the base station control center as the agent, which can connect to the IRS control center, design the state space, action space, reward function, and state transition probability, and establish a basic reinforcement learning model;

[0056] The Markov process is the basis for constructing a reinforcement learning model. The present invention models the problem of maximizing the secure semantic spectral efficiency as a Markov problem. The present invention uses the semantic communication system as the interaction environment, designs the base station control center as an intelligent agent, and the base station center can connect to the IRS intelligent controller.

[0057] The action space consists of bit allocation, IRS transmission factor, and subchannel allocation. At time t, the action space can be expressed as:

[0058] a (t) ={B (t) , Ψ (t) , ρ (t)},

[0059] where B (t) represents bit allocation, Ψ (t) represents the IRS transmission factor, and ρ (t) represents subchannel allocation. The state space consists of the previous moment's action selection, legitimate user channel factor, and eavesdropper channel factor. At time t, the state space can be expressed as:

[0060]

[0061] where a t-1 represents the previous moment's action selection, represents the legitimate user channel factor, represents the eavesdropper channel factor. The reward value function is designed as

[0062] r(t)=Υ (t) +w(υ d -υ e ),

[0063] where Υ (t) represents the secure semantic spectral efficiency, and w is a given weight factor. υ d represents the legitimate user channel gain, which can be expressed as υ e represents the eavesdropping channel gain, which can be expressed as

[0064] Step 4: Use the hybrid algorithm based on DDQN-SAC to jointly optimize semantic bits, sub-channel allocation, and IRS element reflection coefficients to maximize the secure semantic spectral efficiency. Specifically, it includes the following steps:

[0065] Step 4-1: Design a secure semantic resource allocation algorithm for DDQN-SAC based on the formed Markov problem;

[0066] Based on the reinforcement learning modeling in Step 3 above, the present invention proposes a semantic resource allocation algorithm based on DDQN-SAC to jointly optimize semantic bits, sub-channel allocation, and IRS element reflection coefficients to maximize the secure semantic spectral efficiency. The Q value can be calculated as The V value can be expressed as The advantage function is expressed as A π (s,a) = V π (s) - Q π (s,a), and the optimal policy π * can be expressed as

[0067] DDQN obtains discrete actions by maximizing the Q value, which can be expressed as

[0068]

[0069] where θ and ζ are the parameters of two independent Q networks in DDQN respectively. SAC obtains continuous actions through maximum entropy, which can be expressed as

[0070] a (t+1) (s (t) ; λ) = tanh(μ(s (t) ; λ)),

[0071] where λ is the SAC network parameter.

[0072] Step 4-2: Train the semantic coding network to obtain a codebook recording the mapping relationship between physical bits and semantic vectors;

[0073] Under a given data set, without considering the physical transmission stage, train the semantic coding network. After the semantic coding network converges, further consider the physical channel transmission. The codebook is a combination of a set of index vectors, which can convert continuous semantic feature vectors into physical bits. Specifically, through the nearest neighbor search method, find the codebook vector closest to the semantic information to be transmitted and transmit its integer index, where e j is the vector in the codebook and I is the target semantic information. Given a data set and a pre-trained semantic coding network, train the vector codebook and achieve the convergence of the semantic communication network.

[0074] Step 4-3: In the set simulation environment, the agent interacts with the system to obtain training experience. Specifically, the agent sequences the decision of semantic bits, sub-channel allocation, and IRS element reflection coefficients, verifies the effectiveness of the actions in the IRS-assisted secure semantic communication system, and calculates the obtained reward value. The system state sequence, action selection situation, and reward value generated in each iteration are stored in the experience pool;

[0075] Step 4-4: Update the algorithm parameters. When the number of samples in the experience pool meets the requirements, the agent can utilize the samples in the experience pool, aim at minimizing the loss function, conduct batch learning, and update the network parameters through backpropagation gradients, then return to Step 4-3 until the training ends;

[0076] Step 4-5: The algorithm converges, and all network parameters are locally saved.

[0077] Step 5: In the real communication system, obtain the semantic bits, sub-channel allocation, and IRS element reflection coefficients through the secure semantic resource allocation method based on DDQN-SAC, and verify the results.

[0078] The effects of the present invention are further described in detail below in combination with simulation experiments, specifically including:

[0079] 1. Simulation hardware conditions

[0080] The simulation experiment of the present invention is carried out on a simulation platform of Python 3.6 and TensorFlow 1.13. The computer CPU model is E5-2680v4, and the number is 6. The GPU model is NVIDIA GeForce RTX 4070. The video memory is 24GB.

[0081] 2. Simulation system parameters

[0082] The number of users D = 3, and the number of sub-channels C = 3. The base station location is (0, 0, 30), the target users are deployed at (80, 80, 0), (100, 0, 0), and (0, 70, 30), the IRS is deployed at (40, 40, 20), the eavesdropper is deployed at (40, 40, 20), the number of IRS reflection elements is 64, the number of base station antennas M = 6, the maximum transmit power of the base station TP = 20 dBm, and the noise variance is set to 0.02. The path loss factors from the base station to the users, from the base station to the IRS, and from the IRS to the users are respectively set to 3.6, 2.0, and 2.1.

[0083] The dataset used in this invention is the publicly available dataset PASCAL VOC-2007, and the range of the number of semantic bits is 4 - 8 bits. DDQN consists of two Q networks, each network has 5 hidden layers, with 512 neurons in each layer, and the learning rate of the Q network is set to 0.003. SAC consists of two critic networks, two target critic networks and one policy network, each network has 5 hidden layers, with 512 neurons in each layer, and the learning rate of the network is set to 0.003.

[0084] 3. Simulation Content

[0085] Figure 2 Shows the periodic convergence of the IRS-enhanced secure semantic spectrum efficiency maximization method proposed in this invention. As shown in the figure, the proposed secure semantic spectrum efficiency maximization method based on DDQN-SAC can reach convergence within a short iteration period, proving the efficiency of the method.

[0086] Figure 3 Shows the semantic security performance of the IRS-enhanced secure semantic spectrum efficiency maximization method proposed in this invention and other benchmark methods as the number of IRS reflecting elements changes. It can be observed from the figure that the secure semantic spectrum efficiency increases significantly with the increase in the number of IRS reflecting elements and finally tends to converge. In addition, compared with only considering physical layer security or semantic layer security, the proposed secure semantic spectrum efficiency maximization method in this invention can jointly consider physical layer security and semantic layer security, and is superior to the benchmark methods in terms of semantic security performance.

[0087] Figure 4 Shows the secure semantic spectrum efficiency performance of the IIRS-enhanced secure semantic spectrum efficiency maximization method proposed in this invention and other benchmark methods as the number of IRS reflecting elements changes. Compared with the method without IRS-assisted secure transmission, IRS-assisted semantic communication can achieve higher secure semantic spectrum efficiency. In addition, as the number of IRS reflecting elements increases, the secure semantic spectrum efficiency is significantly improved. This is because the increase in IRS reflecting elements can more precisely enhance beamforming, significantly enhance the semantic gain of legitimate users, and suppress eavesdropping semantics, thereby enhancing the secure semantic spectrum efficiency. This also shows that the intelligent method proposed in this invention can make full use of a large number of IRS reflecting elements to achieve secure semantic transmission.

[0088] Based on the above simulation results and analysis, the IRS-enhanced secure semantic spectrum efficiency maximization method proposed in this invention can significantly improve the secure semantic spectrum efficiency compared with other benchmark methods, verifying that the semantic resource allocation algorithm based on DDQN-SAC in this invention can efficiently find the optimal security solution. In addition, the simulation results also show that applying IRS in a secure semantic communication system can enhance both physical layer security and semantic layer security.

Claims

1. A method for maximizing the safety semantics spectrum efficiency enhanced by intelligent reflective surface, characterized in that: The method comprises the following steps: Step 1: Establish an IRS-assisted semantic communication system model; Step 1-1: Establish a secure semantic communication system model and propose a novel secure semantic spectrum efficiency index that jointly considers physical layer security and semantic layer security Step 1-2: Establish an IRS-assisted semantic transmission system, determine the base station location, D user locations, E eavesdropper locations and IRS location, model the channels from the base station to IRS, IRS to D users, IRS to E eavesdroppers, base station to D users and base station to E eavesdroppers, and obtain channel gains; Step 2: Establish the problem of maximizing the security semantic spectrum efficiency; Step 3: The target problem is transformed into a Markov problem. The base station control center is set as an intelligent agent that can connect to the IRS control center. The state space, action space, reward function and state transition probability are designed to establish a basic reinforcement learning model. Step 4: Use a hybrid algorithm based on DDQN-SAC to jointly optimize semantic bits, subchannel allocation, and IRS array element reflection coefficient to maximize the security semantic spectrum efficiency; Step 4-1: Based on the formed Markov problem, design a secure semantic resource allocation algorithm for DDQN-SAC; Step 4-2: Train the semantic coding network to obtain a codebook that records the mapping relationship between physical bits and semantic vectors; Step 4-3: In the set simulation environment, the agent interacts with the system to gain training experience; The agent sequence decides semantic bits, subchannel allocation, and IRS array element reflection coefficients, verifies the effectiveness of actions in the IRS-assisted safety semantic communication system, calculates the reward value, and stores the system state sequence, action selection, and reward value generated by each iteration into the experience pool; Step 4-4: Algorithm parameter update; When the agent interacts with the environment for the set number of times, the agent learns with the goal of minimizing the loss function by replaying the experience stored in the experience pool, and updates the network parameters through reverse gradient, returning to the above step 4-3 until the training ends; Step 4-5: After the algorithm converges, each network parameter is saved locally; Step 5: In a real communication system, the semantic bits, subchannel allocation, and IRS element reflection coefficients are obtained through the secure semantic resource allocation method based on DDQN-SAC, and the results are verified.

2. The method for maximizing the safety semantic spectrum efficiency of intelligent reflective surface enhancement according to claim 1, characterized in that: The step 1-1 includes: establishing a secure semantic communication system model, proposing a novel secure semantic spectrum efficiency index that jointly considers physical layer security and semantic layer security Semantic layer security is considered based on semantic structure similarity and perceptual similarity. The semantic structure similarity is expressed as in, Restore the image for the receiving end, is the brightness difference, is the contrast difference and For the structural difference of the pictures, the perceptual similarity is expressed as: in represents the pre-trained features of the lth layer, w l represents the weight factor of the lth layer, H l and W l Represents the image height and width respectively. The semantic layer index is defined as: φ(S, S′)=SSIM-ξLPIPS, Where ξ is the weight factor. Therefore, the semantic layer security index that a legitimate user d can reach on subchannel c is: in represents the eavesdropping effectiveness factor, Indicates that the eavesdropper restores the image. If but otherwise The physical layer security index is based on the secure transmission rate. The secure transmission rate of a legitimate user d on a subchannel c is expressed as: Among them, R d,c represents the achievable transmission rate of legitimate user d on subchannel c, R e，c represents the transmission rate that eavesdropper e can achieve on subchannel c. Considering the physical layer security and semantic layer security, the spectral efficiency of the semantic security that legitimate user d can achieve on subchannel c is Here, |I| represents the dimension of the semantic vector, and b represents the number of bits used to represent the semantic vector.

3. According to claim 1, a method for maximizing the safety semantics spectrum efficiency enhanced by an intelligent reflective surface is characterized in that: The step 1-2 includes: establishing an IRS-assisted semantic transmission system, determining the base station location, D user locations, E eavesdropper locations and the IRS location, modeling the channels from the base station to the IRS, the IRS to the D users, the IRS to the E eavesdroppers, the base station to the D users and the base station to the E eavesdroppers, and obtaining the channel gain, which is defined as follows; The channel set is represented as The user set is represented as The channels from the base station to the target user d, the eavesdropper and the IRS are denoted as h d , w e and h r , the channels from IRS to the target user and the eavesdropper are denoted as g d and g e , define the IRS array element offset as a diagonal matrix Θ, where α n represents the phase coefficient of the nth reflection array element, φ n represents the offset coefficient of the nth reflection element. The signal received by the legitimate user d on the subchannel c can be expressed as The eavesdropper's eavesdropping signal on subchannel c can be expressed as Where n d and n e is additive Gaussian white noise, and the transmission rate of legitimate user d on subchannel c is expressed as The eavesdropper’s eavesdropping rate on subchannel c is Among them, W c represents bandwidth resources, f d,c represents the beamforming transmitted by the base station, and represents the noise variance.

4. The method for maximizing the safety semantics spectrum efficiency enhanced by intelligent reflective surface according to claim 1, characterized in that: The step 2 includes: defining ρ d，c represents whether the dth user occupies the cth subchannel. If the dth user occupies the cth subchannel, then ρ d,c =1, otherwise, ρ d，c =0, the safety semantics spectral efficiency is expressed as: By jointly optimizing semantic bits, subchannel allocation, and IRS array element reflection coefficients, the system goal is to maximize Y; Resource allocation problem constraints: The first constraint is the number of bits used by each user to represent semantics, denoted as b d ∈{b min , b min +1,…,b max }, the second constraint is the IRS reflection factor constraint expressed as 0≤|[ψ] n，n |≤1 and φ n ∈{0,2π}, where [ψ] n，n represents the diagonal element of the reflector array. The third constraint is the user occupied subchannel constraint. Each user and subchannel only have an occupied or unoccupied relationship, and each user can only occupy one channel, that is, and 5. The method for maximizing the safety semantic spectrum efficiency of intelligent reflective surface enhancement according to claim 1, characterized in that: The step 3 includes: converting the target problem into a Markov problem, setting the base station control center as an intelligent agent that can connect to the IRS control center, designing the state space, action space, reward function and state transition probability, and establishing a reinforcement learning basic model; Markov process is the basis for building reinforcement learning model. The problem of maximizing the safety semantic spectrum efficiency is modeled as a Markov problem. The semantic communication system is used as the interactive environment. The base station control center is designed as an intelligent agent, and the base station center is connected to the IRS intelligent controller. The action space consists of bit allocation, IRS transmission factor and subchannel allocation. At time t, the action space is expressed as: a (t) ={B (t) ,ψ (t) ,r (t) }, Among them, B (t) represents the bit allocation, ψ (t) represents the IRS emission factor and ρ (t) represents the subchannel allocation. The state space consists of the action selection at the previous moment, the legitimate user channel factor, and the eavesdropping user channel factor. At time t, the state space is expressed as: Among them, a t-1 represents the action selection at the previous moment, represents the legitimate user channel factor, represents the eavesdropper channel factor, and the reward value function is designed to be expressed as: r (t) =Y (t) +w(υ d -u e ), Among them, (t) represents the safety semantics spectral efficiency, w is a given weight factor, υ d represents the legitimate user channel gain, which can be expressed as υ e The eavesdropping channel gain is expressed as

Images