Multitask learning-based confrontation effectiveness evaluation method

Abstract

The invention relates to a multi-task learning-based confrontation efficiency evaluation method, belongs to the technical field of efficiency evaluation, and solves the problem that the confrontation efficiency is difficult to accurately estimate in the prior art. The method comprises the following steps: acquiring state data sequences and corresponding coping success rates in a countermeasure simulation process of a plurality of active parties and passive parties, and constructing a training sample set according to a final simulation result; constructing a multi-task learning network model, wherein the multi-task learning network model is used for carrying out response success rate regression and final simulation result prediction based on the input data of the sample; training the multi-task learning network model based on the training sample set to obtain a confrontation efficiency evaluation model; and inputting the state data sequences of the initial stages of the active party and the passive party to be evaluated into the adversarial performance evaluation model to predict and obtain a final simulation result. And rapid and accurate adversarial performance evaluation is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of performance evaluation, and in particular to an adversarial performance evaluation method based on multi-task learning. Background Art

[0002] In recent years, with the increasing complexity of system functions and structures, simulation has become an important tool for performance evaluation and analysis. Using simulation technology to conduct performance evaluation and analyze the expected degree of achievement of tasks under specific conditions is of great significance to the design, verification and application of the system. Performance evaluation is an evaluation method to measure the effect of task completion and has been widely used in many fields.

[0003] Existing methods for evaluating adversarial effectiveness mainly include methods based on empirical models, methods based on single-task learning, simulation methods, and linear regression models. Empirical models rely on historical data and are suitable for known environments, but have poor adaptability to emerging threats; single-task learning methods usually only focus on a specific indicator and are difficult to fully reflect system performance; although simulation methods can simulate a variety of scenarios, they are computationally intensive and lack real-time performance, affecting the timeliness of decision-making; linear regression models are difficult to capture complex nonlinear relationships when analyzing the relationship between historical data and adversarial effects, thereby reducing prediction accuracy. Therefore, existing methods are difficult to accurately estimate adversarial effectiveness. Summary of the invention

[0004] In view of the above analysis, the embodiments of the present invention aim to provide a method and system for evaluating adversarial effectiveness based on multi-task learning, so as to solve the existing problem that it is difficult to accurately estimate adversarial effectiveness.

[0005] On the one hand, an embodiment of the present invention provides a method for evaluating adversarial effectiveness based on multi-task learning, comprising the following steps:

[0006] Acquire state data sequences and corresponding response success rates of multiple active and passive parties in the confrontation simulation process, as well as final simulation results to construct a training sample set;

[0007] Constructing a multi-task learning network model, wherein the multi-task learning network model is used to perform response success rate regression and final simulation result prediction based on sample input data;

[0008] Based on the training sample set, the multi-task learning network model is trained to obtain an adversarial effectiveness evaluation model;

[0009] The state data sequences of the active and passive parties to be evaluated in the initial stage are input into the adversarial effectiveness evaluation model to predict and obtain the final simulation results.

[0010] Based on the further improvement of the above method, the state data sequence and corresponding response success rate of the active and passive parties in the confrontation simulation process, as well as the final simulation results, are obtained to construct a training sample set in the following way:

[0011] Extracting a state data sequence in an initial stage of the confrontation simulation process to obtain first data;

[0012] Perform uniform sampling and non-uniform sampling on the state data sequence in the adversarial simulation process to obtain second data and third data respectively;

[0013] The first data, the second data, and the third data are used as input data, and the corresponding response success rate and the final simulation result of the adversarial simulation are used as labels to form a sample, thereby constructing a training sample set.

[0014] Based on the further improvement of the above method, the sampling points of non-uniform sampling are obtained in the following way:

[0015] S1211, taking the first data point of the state data sequence in the adversarial simulation process as the first sampling point and the last data point as the last sampling point, to obtain a sampling point set;

[0016] S1212, each segment obtained by dividing the state data sequence by the sampling points in the current sampling point set is a current segment;

[0017] S1213, traverse each time step in the current segment in turn, calculate the maximum likelihood estimate sum of the data point of each time step and the two sampling points of the current segment; take the time point with the largest maximum likelihood estimate sum as the candidate sampling point of the current segment;

[0018] S1214. If there is only one candidate sampling point currently, the candidate sampling point is directly used as a sampling point and added to the sampling point set. Otherwise: for each candidate sampling point, the maximum likelihood estimate sum between adjacent sampling points is calculated if the candidate sampling point is used as a sampling point, and the candidate sampling point with the largest maximum likelihood estimate sum between adjacent sampling points is used as the sampling point and added to the sampling point set.

[0019] S1215. If the current number of iterations is reached, stop the iteration to obtain the sampling point combination; otherwise, return to step S1212.

[0020] Based on the further improvement of the above method, the following formula is used to calculate the maximum likelihood estimate of the data point at each time step and the two sampling points of the current segment:

[0021]

[0022] Among them, ψ(b0,t) represents the maximum likelihood estimate of the sampling point b0 and the data point at the t-th time step, ψ(t,b2) represents the maximum likelihood estimate of the data point at the t-th time step and the sampling point b2, b0 and b2 represent the sampling points at the left and right ends of the current segment respectively, S (1) represents the covariance matrix of the data sequence between b0 and t, S (2) represents the covariance matrix of the data sequence between t and b2, logdet(·) represents the logarithmic-determinant function, Tr(·) represents the trace of the covariance matrix, and η represents the regularization parameter

[0023] Based on the further improvement of the above method, the multi-task learning network model includes: a classification network and a regression network;

[0024] The classification network includes: a first feature extraction module, used to extract global features of input data; a classification head, used to predict the final simulation results based on the global features;

[0025] The regression network includes: a second feature extraction module, which is used to extract the local features of each time step of the input data; and a regression unit, which is used to predict the response success rate of each time step based on the local features of each time step.

[0026] Based on the further improvement of the above method, the multi-task learning network model is trained based on the training sample set to obtain an adversarial effectiveness evaluation model, including:

[0027] For each sample in the current training batch, the first data, the second data and the third data are respectively input into the classification network to predict the final simulation result; the first data, the second data and the third data are respectively input into the regression network to predict the response success rate;

[0028] The parameters of the multi-task learning network model are updated based on the loss of the current training batch samples. When the loss reaches a preset accuracy or a preset number of iterations, the training is stopped to obtain an adversarial effectiveness evaluation model.

[0029] Based on the further improvement of the above method, the following formula is used to calculate the training loss:

[0030]

[0031] in, represents the similarity constraint, represents the prediction loss of the classification network, represents the regression loss of the regression network, ω1 and ω2 are weight coefficients.

[0032] Based on the further improvement of the above method, the similarity constraint is calculated using the following formula:

[0033]

[0034] in, represents the similarity between the global features of the first data and the global features of the second data of the i-th sample, represents the similarity between the global features of the first data and the global features of the third data of the i-th sample, z1 represents the global features of the first data, z2 represents the global features of the second data, z3 represents the global features of the third data, α, β and γ represent weights, and n represents the number of samples in the current training batch.

[0035] Based on the further improvement of the above method, the following formula is used to calculate the prediction loss of the classification network:

[0036]

[0037] in, represents the final simulation result prediction loss of the first data of the i-th sample, represents the final simulation result prediction loss of the second data of the i-th sample, represents the prediction loss of the final simulation result of the third data of the i-th sample, n represents the number of samples in the current training batch, α t represents the sample difficulty weight parameter, λ represents the adjustment factor, and δ represents the weight; y i represents the true performance result of i samples, y i =1, y i = 0, represents the prediction result of the first data of the i-th sample, represents the prediction result of the second data of the i-th sample, represents the prediction result of the third data of the i-th sample; P represents the distribution of the true effectiveness of the current batch of training samples, Q represents the distribution of the predicted effectiveness of the current batch of training samples, and KL(P|Q) represents the calculation of the divergence of p and q.

[0038] Based on the further improvement of the above method, the following formula is used to calculate the regression loss of the regression network:

[0039]

[0040] in, represents the regression loss of the first data of the i-th sample, represents the regression loss of the second data of the i-th sample, represents the regression loss of the third data of the i-th sample, n represents the number of samples in the current training batch, and m represents the number of time steps of the third data; when j=1, Represents the label value of the kth time step of the first data of the i-th sample, represents the regression value of the kth time step of the first data of the i-th sample; when j=2, represents the label value of the kth time step of the second data of the i-th sample, represents the regression value of the kth time step of the second data of the i-th sample; when j=3, represents the label value of the kth time step of the third data of the i-th sample, Represents the regression value of the kth time step of the third data of the i-th sample, ε and μ are adjustable parameters.

[0041] Compared with the prior art, the present invention constructs a training sample set by acquiring the state data sequence and the corresponding response success rate in the confrontation simulation process, as well as the final simulation result, and trains a multi-task learning network model based on the constructed training sample set, thereby combining the final simulation effectiveness result prediction with the intermediate response success rate regression, and assisting the final simulation result prediction task through the response success rate regression task, extracting features that are conducive to predicting the simulation results, making the final simulation result prediction more accurate, and inputting the state data sequence of the initial stage of the active and passive parties to be evaluated into the confrontation effectiveness evaluation model to quickly and accurately predict the final simulation result, thereby improving the accuracy and efficiency of the confrontation effectiveness evaluation.

[0042] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0044] Figure 1 The flowchart of the adversarial effectiveness evaluation method based on multi-task learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0046] A specific embodiment of the present invention discloses a method for evaluating adversarial effectiveness based on multi-task learning, such as Figure 1 As shown, the following steps are included:

[0047] S1, obtaining state data sequences and corresponding response success rates of multiple active and passive parties in the confrontation simulation process, as well as the final simulation results to construct a training sample set;

[0048] S2. Construct a multi-task learning network model, wherein the multi-task learning network model is used to perform response success rate regression and final simulation result prediction based on sample input data;

[0049] S3, training the multi-task learning network model based on the training sample set to obtain an adversarial effectiveness evaluation model;

[0050] S4. Input the state data sequence of the active party and the passive party to be evaluated in the initial stage into the confrontation effectiveness evaluation model to predict and obtain the final simulation result.

[0051] It should be noted that the final simulation result is the confrontation effectiveness, which is 1 if the confrontation is successful and 0 if it is unsuccessful.

[0052] Compared with the prior art, the adversarial effectiveness evaluation method based on multi-task learning provided in this embodiment constructs a training sample set by acquiring the state data sequence and the corresponding response success rate in the adversarial simulation process, as well as the final simulation result, and trains the multi-task learning network model based on the constructed training sample set, thereby combining the final simulation effectiveness result prediction with the intermediate response success rate regression, and assisting the final simulation result prediction task through the response success rate regression task, extracting features that are conducive to predicting the simulation results, and making the final simulation result prediction more accurate. The state data sequence of the active and passive parties to be evaluated in the initial stage is input into the adversarial effectiveness evaluation model to quickly and accurately predict the final simulation result, thereby improving the accuracy and efficiency of the adversarial effectiveness evaluation.

[0053] It should be noted that the status data in the confrontation simulation process includes the number of entities, behavioral strategies, speed, position and operation mode, information processing efficiency of each active party, and the number of entities, behavioral strategies, orientation and information processing capabilities of the passive party.

[0054] The response success rate is the response success rate at each simulation moment, which is calculated using the following formula:

[0055]

[0056] Where N represents the number of active entities at the current simulation time, p i represents the probability that the i-th active party is discovered by the passive party, and ρ represents the success rate of the passive party’s response action.

[0057] Specifically, in step S1, the state data sequence and the corresponding response success rate of the active party and the passive party in the confrontation simulation process, as well as the final simulation results are obtained to construct a training sample set in the following manner:

[0058] S11, extracting a state data sequence in an initial stage of a confrontation simulation process to obtain first data;

[0059] S12, performing uniform sampling and non-uniform sampling on the state data sequence in the adversarial simulation process to obtain second data and third data respectively;

[0060] S13, taking the first data, the second data, and the third data as input data, and the corresponding response success rate and the final simulation result of the adversarial simulation as labels to form a sample, thereby constructing a training sample set.

[0061] According to the data of the confrontation simulation process, the change range of the data state characteristics under the full trajectory of the confrontation parties is not significant, and there is no short-term mutation. Therefore, the initial stage data set with a specific step size can be extracted from local features to global trajectory features.

[0062] During implementation, the length of the initial phase is set according to the specific simulation task length.

[0063] In order to enable the trained model to predict the final simulation results based on the simulation data in the initial stage, the present invention first extracts the state data sequence in the initial stage of the adversarial simulation process to obtain the first data X1. In order to reduce the impact of the simulation process sampling on the data consistency, the present invention performs uniform sampling and non-uniform sampling on the state data sequence in the adversarial simulation process to obtain the second data X2 and the third data X3 respectively. The first data, the second data and the third data are all input data of the sample, and the response success rate corresponding to each time step of the first data, the second data and the third data and the final simulation result are used as the sample label. The second data and the third data help the model to extract features that can accurately predict the final simulation results from the first data.

[0064] If the state data sequence in the initial stage includes M time steps, and the state data of each time step is D-dimensional, then X1, X2, X3∈R D*M .

[0065] Specifically, the sampling points of non-uniform sampling are obtained in the following manner:

[0066] S1211, taking the first data point of the state data sequence in the adversarial simulation process as the first sampling point and the last data point as the last sampling point, to obtain a sampling point set;

[0067] S1212, each segment obtained by dividing the state data sequence by the sampling points in the current sampling point set is a current segment;

[0068] S1213, traverse each time step in the current segment in turn, calculate the maximum likelihood estimate sum of the data point of each time step and the two sampling points of the current segment; take the time point with the largest maximum likelihood estimate sum as the candidate sampling point of the current segment;

[0069] S1214. If there is only one candidate sampling point currently, the candidate sampling point is directly used as a sampling point and added to the sampling point set. Otherwise: for each candidate sampling point, the maximum likelihood estimate sum between adjacent sampling points is calculated if the candidate sampling point is used as a sampling point, and the candidate sampling point with the largest maximum likelihood estimate sum between adjacent sampling points is used as the sampling point and added to the sampling point set.

[0070] S1215. If the current number of iterations is reached, stop the iteration to obtain the sampling point combination; otherwise, return to step S1212.

[0071] For example, the state data sequence in the adversarial simulation process has a total of T time steps S = {x1, x2, ... x T}, the data point x at each time step t The dimension of is D. The state data sequence in the initial stage includes M time steps, so the state data sequence in the adversarial simulation process needs to be divided into M-1 segments.

[0072] The rth segment is represented by [b r-1 ,b r ), where each data point obeys a multi-source Gaussian distribution, and its mean vector and covariance matrix are:

[0073]

[0074] First, in step S1211, the first data point of the state data sequence is used as the first sampling point, and the last data point is used as the last sampling point, that is, b0 = 0, b2 = T-1. The sampling points in the current sampling point set {b0, b2} divide the state data sequence into a segment, that is, the state data sequence as a whole is taken as a segment.

[0075] Take the state data sequence as the current segment, and traverse each time step in the current segment in turn, that is, calculate each time step in the current segment except the sampling point, and calculate the maximum likelihood estimate and of the data point of each time step and the two sampling points of the current segment.

[0076] Specifically, the following formula is used to calculate the maximum likelihood estimate of the data point at each time step and the two sampling points of the current segment:

[0077]

[0078] Among them, ψ(b0,t) represents the maximum likelihood estimate of the sampling point b0 and the data point at the t-th time step, ψ(t,b2) represents the maximum likelihood estimate of the data point at the t-th time step and the sampling point b2, b0 and b2 represent the sampling points at the left and right ends of the current segment respectively, S (1) represents the covariance matrix of the data sequence between b0 and t, S (2) represents the covariance matrix of the data sequence between t and b2, logdet(·) represents the logarithm-determinant function, Tr(·) represents the trace of the covariance matrix, η represents the regularization parameter, η≥0.

[0079] Traverse each time step between b0 and b2 and calculate ψ t , take ψ t The largest time point is the candidate sampling point. Since there is only one segment at present, the candidate sampling point set is directly obtained. If there are multiple segments at present, step S1213 needs to be executed for all the current segments to obtain the corresponding candidate sampling points, and then the candidate sampling point set is formed, and then step S1214 is executed. Since there is only one segment at present, corresponding to one candidate sampling point, the obtained candidate sampling point is directly used as a sampling point and added to the sampling point set, and the sampling point set is updated to {b0, b1, b2} (arranged in time order).

[0080] Since the number of iterations has not been reached, the process returns to step S1212. At this time, the sampling points in the current sampling point set divide the state data sequence into two segments, namely the current segments [b0, b1) and [b1, b2). Then, step S1213 is executed on the segment [b0, b1) to obtain the corresponding candidate sampling point t1, and step S1213 is executed on the segment [b1, b2) to obtain the corresponding candidate sampling point t2.

[0081] In step S1214, if the candidate sampling point t1 is used as the sampling point, the sampling point set is {b0, t1, b1, b2}, and the maximum likelihood estimate and the maximum likelihood estimate between adjacent sampling points are calculated according to the above formula, that is, ψ(b0, t1) + ψ(t1, b1) + ψ(b1, b2); if the candidate sampling point t2 is used as the sampling point, the sampling point set is {b0, b1, t2, b2}, and the maximum likelihood estimate and the maximum likelihood estimate between adjacent sampling points are calculated according to the above formula, that is, ψ(b0, b1) + ψ(b1, t2,) + ψ(t2, b2); the candidate sampling point corresponding to the maximum likelihood estimate and is taken as the sampling point, added to the sampling point set, and the sampling point set is updated to {b0, b1, b2, b3} (arranged in chronological order), and so on, iterate M-2 times to obtain the sampling point.

[0082] By adopting non-uniform sampling, the trend of the state sequence can be better sampled, laying the foundation for accurately extracting prediction features.

[0083] During implementation, uniform sampling is to divide the state data sequence into M-1 equal parts to obtain sampling points.

[0084] Specifically, the constructed multi-task learning network model includes: classification network and regression network;

[0085] The classification network includes: a first feature extraction module, used to extract global features of input data; a classification head, used to predict the final simulation results based on the global features;

[0086] The regression network includes: a second feature extraction module, which is used to extract the local features of each time step of the input data; and a regression unit, which is used to predict the response success rate of each time step based on the local features of each time step.

[0087] The classification network is used to predict the final simulation results, and the regression network is used to regress the response success rate value. Performing classification and regression tasks simultaneously improves the application flexibility and performance of the model.

[0088] Specifically, the multi-task learning network model is trained based on the training sample set to obtain an adversarial effectiveness evaluation model, including:

[0089] For each sample in the current training batch, the first data, the second data and the third data are respectively input into the classification network to predict the final simulation result; the first data, the second data and the third data are respectively input into the regression network to predict the response success rate;

[0090] The parameters of the multi-task learning network model are updated based on the loss of the current training batch samples. When the loss reaches a preset accuracy or a preset number of iterations, the training is stopped to obtain an adversarial effectiveness evaluation model.

[0091] In implementation, the classification head consists of two fully connected layers and a Softmax layer. The regression unit consists of a fully connected layer and a linear activation layer for the output unit.

[0092] Both the first feature extraction module and the second feature extraction module can adopt the ResNet structure, except that the input of the first feature extraction module is the complete data of the first data, the second data or the third data, while the input of the second feature extraction module is the data of each time step of the first data, the second data or the third data. The training loss of the model is calculated according to the prediction results of the classification network for the first data, the second data and the third data and the regression results of the regression network for the first data, the second data and the third data, and the model training is performed using the back propagation algorithm and the optimizer (such as the Adam optimizer), and the model parameters are iteratively updated by minimizing the loss until the condition for stopping training is reached to obtain the adversarial effectiveness evaluation model.

[0093] Specifically, the training loss is calculated using the following formula:

[0094]

[0095] in, represents the similarity constraint, represents the prediction loss of the classification network, represents the regression loss of the regression network, ω1 and ω2 are weight coefficients.

[0096] Based on the classification prediction loss and the regression loss, the present invention minimizes the difference between the features of the first data and the second data and the third data through similarity constraints in order to enable the model to learn the similarities between the features of the first data and the second data and the third data.

[0097] Specifically, the similarity constraint is calculated using the following formula:

[0098]

[0099] in, Represents the difference between the global features of the first data and the global features of the second data of the i-th sample, represents the difference between the global features of the first data and the global features of the third data of the i-th sample, z1 represents the global features of the first data, z2 represents the global features of the second data, z3 represents the global features of the third data, α, β and γ represent weights, and n represents the number of samples in the current training batch.

[0100] The first term in the difference calculation formula represents the distance between two global features; the second term represents the angular difference between two global features, which helps improve the discriminative ability of feature space, especially in high-dimensional space; the third term is the regularization term, which is used to control the scale of feature vectors to avoid calculation instability caused by too large values; α, β, γ are adjustable parameters used to control the influence of different terms. By comprehensively considering distance, angle and normalization, the model not only focuses on minimizing the direct distance between features during learning, but also considers directionality and scale, which helps improve the generalization ability and stability of the model.

[0101] Specifically, the following formula is used to calculate the prediction loss of the classification network:

[0102]

[0103] in, represents the final simulation result prediction loss of the first data of the i-th sample, represents the final simulation result prediction loss of the first data of the i-th sample, represents the final simulation result prediction loss of the second data of the i-th sample, represents the prediction loss of the final simulation result of the third data of the i-th sample, n represents the number of samples in the current training batch, α t represents the sample difficulty weight parameter, λ represents the adjustment factor, and δ represents the weight; y i represents the true performance result of i samples, y i =1, y i = 0, represents the prediction result of the first data of the i-th sample, represents the prediction result of the second data of the i-th sample, represents the prediction result of the third data of the i-th sample; P represents the distribution of the true effectiveness of the current batch of training samples, Q represents the distribution of the predicted effectiveness of the current batch of training samples, and KL(P|Q) represents the calculation of the divergence of p and q.

[0104] During implementation, taking into account the imbalance of samples, the present invention calculates the prediction loss of each sample by adopting an improved dynamic balance loss function, which not only takes into account the difficulty of the sample, but also dynamically adjusts the weight of each category to more effectively handle category imbalance.

[0105] Specifically, the regression loss of the regression network is calculated using the following formula:

[0106]

[0107] in, represents the regression loss of the first data of the i-th sample, represents the regression loss of the second data of the i-th sample, represents the regression loss of the third data of the i-th sample, n represents the number of samples in the current training batch, and m represents the number of time steps of the third data; when j=1, Represents the label value of the kth time step of the first data of the i-th sample, represents the regression value of the kth time step of the first data of the i-th sample; when j=2, represents the label value of the kth time step of the second data of the i-th sample, represents the regression value of the kth time step of the second data of the i-th sample; when j=3, represents the label value of the kth time step of the third data of the i-th sample, Represents the regression value of the kth time step of the third data of the i-th sample, ε and μ are adjustable parameters.

[0108] The first term of the regression loss is the conventional MSE mean square error loss, which is used to calculate the basic regression task loss. The second term is an exponential decay term, which is used to impose additional penalties on samples with large differences between the predicted values ​​and the target values, especially samples near the data boundary or outliers, thereby improving training efficiency.

[0109] The back-propagation algorithm is used to calculate the gradient of the loss function relative to the model parameters, and the model parameters are iteratively updated to minimize the loss function. When the loss reaches the preset accuracy or the predetermined number of iterations, the training is completed and the adversarial effectiveness evaluation model is obtained.

[0110] For the adversarial simulation task to be evaluated, the state data sequence of the active and passive parties in the initial stage is obtained and input into the adversarial effectiveness evaluation model to quickly and accurately predict the final simulation results. Based on these prediction results, decision makers can adjust the adversarial strategy to improve the success rate.

[0111] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0112] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for evaluating adversarial effectiveness based on multi-task learning, characterized in that: The following steps are involved: Acquire state data sequences and corresponding response success rates of multiple active and passive parties in the confrontation simulation process, as well as final simulation results to construct a training sample set; Constructing a multi-task learning network model, wherein the multi-task learning network model is used to perform response success rate regression and final simulation result prediction based on sample input data; Based on the training sample set, the multi-task learning network model is trained to obtain an adversarial effectiveness evaluation model; The state data sequences of the active and passive parties to be evaluated in the initial stage are input into the adversarial effectiveness evaluation model to predict and obtain the final simulation results.

2. The method for evaluating adversarial effectiveness based on multi-task learning according to claim 1, characterized in that: The state data sequence and corresponding response success rate of the active and passive parties in the confrontation simulation process, as well as the final simulation results, are obtained by constructing a training sample set in the following way: Extracting a state data sequence in an initial stage of the confrontation simulation process to obtain first data; Perform uniform sampling and non-uniform sampling on the state data sequence in the adversarial simulation process to obtain second data and third data respectively; The first data, the second data, and the third data are used as input data, and the corresponding response success rate and the final simulation result of the adversarial simulation are used as labels to form a sample, thereby constructing a training sample set.

3. The method for evaluating adversarial effectiveness based on multi-task learning according to claim 2, characterized in that: The sampling points of non-uniform sampling are obtained in the following way: S1211, taking the first data point of the state data sequence in the adversarial simulation process as the first sampling point and the last data point as the last sampling point, to obtain a sampling point set; S1212, each segment obtained by dividing the state data sequence by the sampling points in the current sampling point set is a current segment; S1213, traverse each time step in the current segment in turn, calculate the maximum likelihood estimate sum of the data point of each time step and the two sampling points of the current segment; take the time point with the largest maximum likelihood estimate sum as the candidate sampling point of the current segment; S1214. If there is only one candidate sampling point currently, the candidate sampling point is directly used as a sampling point and added to the sampling point set. Otherwise: for each candidate sampling point, the maximum likelihood estimate sum between adjacent sampling points is calculated if the candidate sampling point is used as a sampling point, and the candidate sampling point with the largest maximum likelihood estimate sum between adjacent sampling points is used as the sampling point and added to the sampling point set. S1215. If the current number of iterations is reached, stop the iteration to obtain the sampling point combination; otherwise, return to step S1212.

4. The method for evaluating adversarial effectiveness based on multi-task learning according to claim 3, characterized in that: The following formula is used to calculate the maximum likelihood estimate of the data point at each time step and the two sampling points of the current segment: Among them, ψ(b0,t) represents the maximum likelihood estimate of the sampling point b0 and the data point at the t-th time step, ψ(t,b2) represents the maximum likelihood estimate of the data point at the t-th time step and the sampling point b2, b0 and b2 represent the sampling points at the left and right ends of the current segment respectively, S (1) represents the covariance matrix of the data sequence between b0 and t, S (2) represents the covariance matrix of the data sequence between t and b2, logdet(·) represents the logarithm-determinant function, Tr(·) represents the trace of the covariance matrix, and η represents the regularization parameter.

5. The method for evaluating adversarial effectiveness based on multi-task learning according to claim 2, characterized in that: The multi-task learning network model includes: a classification network and a regression network; The classification network includes: a first feature extraction module, used to extract global features of input data; a classification head, used to predict the final simulation result based on the global features; The regression network includes: a second feature extraction module, which is used to extract the local features of each time step of the input data; and a regression unit, which is used to predict the response success rate of each time step based on the local features of each time step.

6. The method for evaluating adversarial effectiveness based on multi-task learning according to claim 5, characterized in that: The multi-task learning network model is trained based on the training sample set to obtain an adversarial effectiveness evaluation model, including: For each sample in the current training batch, the first data, the second data and the third data are respectively input into the classification network to predict the final simulation result; the first data, the second data and the third data are respectively input into the regression network to predict the response success rate; The parameters of the multi-task learning network model are updated based on the loss of the current training batch samples. When the loss reaches a preset accuracy or a preset number of iterations, the training is stopped to obtain an adversarial effectiveness evaluation model.

7. The method for evaluating adversarial effectiveness based on multi-task learning according to claim 6, characterized in that: The training loss is calculated using the following formula: in, represents the similarity constraint, represents the prediction loss of the classification network, represents the regression loss of the regression network, ω1 and ω2 are weight coefficients.

8. The method for evaluating adversarial effectiveness based on multi-task learning according to claim 7, characterized in that: The similarity constraint is calculated using the following formula: in, represents the similarity between the global features of the first data and the global features of the second data of the i-th sample, represents the similarity between the global features of the first data and the global features of the third data of the i-th sample, represents the global characteristics of the first data, represents the global characteristics of the second data, represents the global features of the third data, α, β and γ represent weights, and n represents the number of samples in the current training batch.

9. The method for evaluating adversarial effectiveness based on multi-task learning according to claim 7, characterized in that: The prediction loss of the classification network is calculated using the following formula: in, represents the final simulation result prediction loss of the first data of the i-th sample, represents the final simulation result prediction loss of the second data of the i-th sample, represents the prediction loss of the final simulation result of the third data of the i-th sample, n represents the number of samples in the current training batch, α t represents the sample difficulty weight parameter, λ represents the adjustment factor, and δ represents the weight; y i represents the true performance result of i samples, y i =1, y i = 0, represents the prediction result of the first data of the i-th sample, represents the prediction result of the second data of the i-th sample, represents the prediction result of the third data of the i-th sample; P represents the distribution of the true effectiveness of the current batch of training samples, Q represents the distribution of the predicted effectiveness of the current batch of training samples, and KL(P|Q) represents the calculation of the divergence of p and q.

10. The method for evaluating adversarial effectiveness based on multi-task learning according to claim 7, characterized in that: The regression loss of the regression network is calculated using the following formula: in, represents the regression loss of the first data of the i-th sample, represents the regression loss of the second data of the i-th sample, represents the regression loss of the third data of the i-th sample, n represents the number of samples in the current training batch, and m represents the number of time steps of the third data; when j=1, represents the label value of the kth time step of the first data of the i-th sample, represents the regression value of the kth time step of the first data of the i-th sample; when j=2, represents the label value of the kth time step of the second data of the i-th sample, represents the regression value of the kth time step of the second data of the i-th sample; when j=3, represents the label value of the kth time step of the third data of the i-th sample, Represents the regression value of the kth time step of the third data of the i-th sample, ε and μ are adjustable parameters.

Images