Method for improving network model performance based on optimization of training sample quality
By generating adversarial samples based on guid-Xmage interpretability method and optimizing the training set, the convolutional neural network model has solved the problems of high computing resource consumption, unstable generalization ability and poor interpretability in terms of poor interpretation, and the model performance is improved and the migration is achieved. Enhanced.
Patent Information
- Application Number
- CN202411151695.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-21
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-08-21
AI Technical Summary
The existing technology consumes a lot of computing resources, is difficult to select hyperparameters, is unstable in generalization, poor interpretability of model, high data demand and poor migration, and traditional data enhancement methods lack clear optimization goals, resulting in The model performs well on the training set but has poor generalization ability on the test set.
The guid-Xmage interpretive method is used to output significant graphs for each original training sample, tamper with background pixels to generate adversarial samples, build a new training set to improve network model performance, reduce background classification contribution value, and improve model accuracy and robustness sex.
It effectively improves the accuracy and robustness of the network model. The optimized training samples are transferable and can adapt to different tasks and fields, reducing the model's demand for data and improving the interpretability of the model.
Smart Images

Figure CN118981641A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of deep learning technology, and in particular relates to a method for improving the performance of network models based on optimizing the quality of training samples. Background Technology
[0002] Generally speaking, the performance of a convolutional neural network model, such as accuracy and robustness, is mainly determined by the network structure and the quality of the training set.
[0003] In the field of research on optimizing network structure to improve network performance, many excellent methods have been proposed. These include network pruning, which reduces network size by removing redundant connections or parameters, thereby improving network efficiency and generalization ability; knowledge distillation, which improves student network performance by training on a large, slow teacher network and then transferring its "knowledge" to a smaller, faster student network; and network ensemble, which improves performance and generalization ability by integrating multiple networks with different structures or the same structure but different initializations. While methods for optimizing network structure to improve network performance can achieve significant improvements, they also face some drawbacks and challenges. First, these methods consume significant computational resources, especially some optimization methods that require substantial computational resources and time, particularly in methods like neural architecture search that require extensive experimentation. Second, hyperparameter selection is difficult; optimizing network structure typically requires adjusting many hyperparameters, such as learning rate and regularization parameters, and choosing appropriate hyperparameter combinations is a challenging task. Furthermore, the enhanced model's generalization ability is unstable. Optimizing the network structure may result in good performance on the training set but poor generalization on the test set, leading to overfitting. Additionally, the model's interpretability is poor; some optimization methods may result in complex network structures that are difficult to explain in terms of feature extraction and decision-making processes, reducing the model's interpretability. Data requirements are high; some methods require large amounts of training data to effectively optimize the network structure, which may not be suitable for tasks with scarce data. Finally, this enhanced model has poor transferability; some optimized network structures have poor transferability across different tasks or domains and cannot adapt well to new tasks or data.
[0004] Besides optimizing network structure, optimizing training set quality is crucial for improving network model performance, and traditional data augmentation methods remain widely used. Among these, scaling, rotating, and changing the background of training samples are common and effective techniques. These methods increase the diversity of training data and improve the model's generalization ability across different scenarios. Scaling allows the model to learn features at different scales, increasing its robustness. Rotation allows the model to learn the representation of objects from different angles, improving its adaptability to rotational transformations. Changing the background increases data complexity, enabling the model to better adapt to complex real-world scenarios. In addition to traditional methods, many novel data augmentation techniques have been proposed, such as random cropping, color distortion, and blurring. These methods further enrich training data, improving the model's generalization ability and robustness. However, traditional training set optimization methods essentially rely on increasing the number of samples as the basis for optimization, lacking a clear optimization objective and exhibiting a degree of arbitrariness. Often, a certain number of new training samples are needed to achieve the goal of improving network model performance. Furthermore, some methods may alter the content or features of the image, making the model's predictions difficult to interpret and understand, thus reducing the model's interpretability. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention proposes a method for improving network model performance based on optimizing training sample quality, thereby resolving the issues present in the prior art.
[0006] To achieve the above objectives, this invention provides a method for improving network model performance based on optimizing training sample quality, comprising:
[0007] The original training set is input into the target network, and the corresponding saliency map is output for each original training sample based on the guid-Xmage interpretive method; the original training set includes several original training samples.
[0008] Based on the saliency map, the background pixels of the original training sample are monotonically decreased iteratively modified within different tampering intensity ranges to obtain adversarial samples with the function of weakening the background classification contribution value.
[0009] A new training set is constructed based on the adversarial examples and the original examples, and the performance of the network model is improved based on the training set.
[0010] Optionally, the process of outputting a corresponding saliency map for each original training sample based on the guid-Xmage interpretive method includes:
[0011] A symbolic model is constructed based on the original training samples;
[0012] Extract the weights corresponding to the symbolic variables in the symbolic model;
[0013] Construct a pixel weight matrix based on the weights;
[0014] The pixel classification contribution matrix is obtained by multiplying the pixel weight matrix and the pixel values of the original training image by a dot product.
[0015] The saliency map corresponding to the original training sample is obtained based on the pixel classification contribution value matrix.
[0016] Optionally, the process of constructing a symbolic model based on the original training samples includes:
[0017] The original training samples are input into the target convolutional neural network to obtain the target network model, and the execution trajectory information is recorded;
[0018] Based on the original training samples, a symbolic image is defined, and the symbolic image accurately executes the trajectory information to obtain a symbolic model.
[0019] Optionally, the expression for the symbolic model is:
[0020]
[0021] In the formula, For the symbolic model, Z is the symbolic image. Let i and j be symbolic variables in the symbolized image Z. The corresponding pixel x ij The position ∈X, k is the k-th channel of the input real image X, and ρ is a symbolic variable. The irrelevant constant represents the cumulative value of the bias in the target network model f(X). S is an element in the pixel weight matrix, where S indicates that the current f(Z) is a symbolic model obtained based on the original network model, and t indicates that the t-th classification branch is the target of symbolic execution.
[0022] Optionally, the pixel values at the same position in different channels of the pixel classification contribution value matrix are summed and subjected to a maximum-minimum normalization operation to obtain the corresponding distribution map.
[0023] Optionally, the saliency map corresponding to the original training sample is obtained by processing the pixel classification contribution value matrix based on the segmented mean method.
[0024] Optionally, the process of obtaining the adversarial sample based on the tampering method includes:
[0025] Calculate the category label of the original image and initialize the perturbation intensity range and the number of iterations for each element in the original image;
[0026] The tamperable elements are obtained by determining the contribution of each element based on the classification contribution value matrix;
[0027] The tamperable element is modified to obtain an element value, and the element value is limited to obtain the image of this round of tampering;
[0028] Calculate the category label of the image in this round of tampering;
[0029] The class label of the image being tampered with in this round is compared with the class label of the original image. If the two class labels are different, the tampering is successful and an adversarial example is obtained; if the two class labels are the same, the tampering fails and the next round of tampering is performed.
[0030] Optionally, the calculation formula for modifying the modifiable element is:
[0031]
[0032] In the formula, This represents the altered element value. Let λ represent the elements in the original image, and let λ represent the step size of the iterative modification.
[0033] Compared with the prior art, the present invention has the following advantages and technical effects:
[0034] This invention inputs the original training set into the target network and outputs a corresponding saliency map for each original training sample based on the guid-Xmage interpretive method. The original training set includes several original training samples. The background features of the original training samples are modified based on the saliency map to obtain counterexample samples. A new training set is constructed based on the counterexample samples and the original samples. The network model performance is improved based on the training set, which can improve the accuracy and robustness of the network model. Furthermore, the optimized training samples are transferable.
[0035] The training set generated by the method for optimizing training samples proposed in this invention has knowledge transferability and can be widely applied in industry. Attached Figure Description
[0036] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:
[0037] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;
[0038] Figure 2 This is a general distribution diagram of the classification values y and y' output by the convolutional neural network model f(X) in an embodiment of the present invention for different input images X and X'. Detailed Implementation
[0039] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0040] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0041] Example 1
[0042] Convolutional neural network (CNN) models, benefiting from their superior classification performance in image recognition, are now widely used in civilian and military fields in my country, such as intelligent manufacturing, disaster identification, aerial image recognition of weaponry, and radar-assisted decision-making. The performance of a CNN model applied in real-world scenarios, such as accuracy and robustness, is often determined by both the quality of the training samples and the network structure. In the research field of optimizing network structure to improve network performance, many excellent optimization methods have been proposed, such as residual modules and attention mechanisms. Based on these optimization methods, high-performance network models such as ResNet and DenseNet have been widely applied to various image classification tasks. Compared to optimizing network structure, this invention focuses on improving the quality of training samples. Based on the proposed interpretive method guid-Xmage, it designs and implements a method to improve the performance of CNN models by optimizing training samples.
[0043] This invention mainly includes a method for generating optimized training samples based on the guid-Xmage interpretive method and using the optimized training samples as a supplement to the model training set, which can effectively improve the model's performance. The method includes the following steps: inputting the original training set into the target network; outputting a corresponding saliency map for each original training sample based on the guid-Xmage interpretive method; the original training set includes several original training samples; performing monotonically decreasing iterative manipulation on the background pixels of the original training samples based on the saliency map within different manipulation intensity ranges to obtain adversarial samples with the function of weakening the background classification contribution value; constructing a new training set based on the adversarial samples and the original samples; and improving the network model performance based on the training set.
[0044] This embodiment discloses a method for improving network model performance based on optimizing training sample quality. This method relies on the weights of the convolutional neural network model, uses symbolic execution technology, and optimizes training sample quality based on a proposed precise interpretability method (guid-Xmage) with pixel-level interpretability to improve network model performance. The computational flow is as follows: Figure 1As shown.
[0045] The original training set is input into the target network, and the guid-Xmage interpretive method outputs a corresponding saliency map for each original training sample, specifically including the following:
[0046] A symbolic model is constructed based on the original training samples. Specifically, the original training samples are input into the target convolutional neural network to obtain the target network model, and the execution trajectory information is recorded. A symbolic image is defined based on the original training samples, and the symbolic image is used to accurately execute the trajectory information to obtain the symbolic model. The specific implementation process includes: First, inputting a real image X to drive the target convolutional neural network model f(X) for execution, while acquiring and recording the execution trajectory information of the input image X at each step in the target network model f(X), such as the magnitude relationship of the ReLU function output values and the position where the maximum value is selected in max-pooling. Second, defining a symbolic image Z of the same size as the input image X, and using the recorded execution trajectory information of the input image X as a guide, accurately executing the target network model f(X) using the symbolic image Z. Third, assuming the t-th classification branch of the target network model f(X) is the target of symbolic execution, based on the result of accurate symbolic execution, obtaining a symbolic image that is identical to the target network model f(X). t (x) Symbolic model of execution trajectory equivalent See expression (1), where: Let i and j be symbolic variables in the symbolized image Z. The corresponding pixel x ij The position ∈X, k is the k-th channel of the input image X, and ρ is a symbolic variable. The irrelevant constants represent the cumulative values of the biases in the target network model f(X).
[0047] like Figure 2 As shown, when training a CNN model using a training dataset, the gradient concept in nonlinear functions is used to provide direction for the gradual refinement of weights and biases within the CNN model, causing the classification values output by the trained CNN model to generally tend towards a nonlinear trend. When the CNN model takes images with different similarities as input, it exhibits a monotonically increasing or decreasing trend in local intervals. Based on the monotonically increasing or decreasing trend in certain local intervals, it can be inferred that if multiple pixel values in the input image X are weakly perturbed at a certain intensity, then the trained CNN model f(X) will take the perturbed image X' as input, and the output classification value y' should match the possible scenarios of the original input image X's classification value y in a certain classification branch. Figure 2The scenario shown in (a) or (b) is where (a) represents the distribution trend of the classification value y' in the increasing scenario when the contribution of the perturbation pixel to the classification value is positive; and (b) represents the distribution trend of the classification value y' in the decreasing scenario when the contribution of the perturbation pixel to the classification value is negative.
[0048] Extracting the weights corresponding to the symbolic variables in the symbolic model, and constructing a pixel weight matrix based on these weights, the specific implementation process includes the fourth step: extracting the symbolic model. Chinese symbolic variables corresponding weights And with the extracted weights Based on this, construct a pixel weight matrix Δ with the same size and number of channels as the input image X.
[0049] The pixel classification contribution matrix is obtained by multiplying the pixel weight matrix and the pixel values of the original training image. Based on the pixel classification contribution matrix, the saliency map corresponding to the original training samples is obtained. The specific implementation process includes the fifth step, which calculates the pixel classification contribution matrix A by multiplying the pixel weight matrix Δ and the pixel values of the input image X. Then, the pixel values at the same position in different channels of the pixel classification contribution matrix A are summed and subjected to max-min normalization. At the same time, a distribution map is output to visualize the relative relationship of the pixel classification contribution values. Furthermore, to facilitate data observation, a saliency map is output based on the idea of segmented equalization to clearly explain the classification criteria of the input image X.
[0050] Adversarial examples are obtained by modifying the background features of the original training samples based on saliency maps. These examples include the following:
[0051] The sixth step, guided by the saliency map, involves obtaining one or more adversarial samples with the same object features as the original training samples and with the function of weakening the background classification contribution value by modifying the background features of the original training samples. Then, the sample set composed of the obtained adversarial samples and the original training samples is used as a new training set to achieve the goal of optimizing the training samples.
[0052]
[0053] A new training set is constructed based on adversarial examples and original samples, and the performance of the network model is improved based on the training set.
[0054] The main theoretical basis of this invention for improving the performance of convolutional neural network models by optimizing training samples is as follows: A precise interpretation of the relationship between pixel values in an input image and the classification values output by the network model clearly reveals the classification basis of the input image. Statistical analysis of the classification basis of multiple input images reveals the defects and deficiencies in the classification knowledge carried by the network model. Based on this, the performance of the network model, such as accuracy and robustness, can be improved by specifically correcting the quality of the training samples. Analysis of the saliency map of the original training samples shows that the background contributes both positively and negatively to the classification value of the target network model. This result indicates that the calculation basis for the target network model's classification of input images is jointly determined by object features and background features; that is, the target network model learns both object features and background features during training. If a convolutional neural network model focuses too much on background features, its classification behavior is easily interfered with by the background, and its accuracy and robustness are often at a low level.
[0055] This invention, based on the proposed interpretive method guid-Xmage, aims to improve the performance of target convolutional neural network models, such as accuracy and robustness. The core idea is to optimize training samples so that the target network model learns as many object features as possible and as few background features as possible in the optimized training set. In other words, by relearning on the optimized training set, the target network model minimizes the interference of the background with its classification behavior; that is, the positive and negative contributions of background features to the classification value of the target network model are both relatively small compared to object features. To achieve the goal of optimizing training samples, this invention uses samples from the original training set of the target network model as input. The proposed interpretive method guid-Xmage outputs a corresponding saliency map for each original training sample. Guided by the saliency map, one or more adversarial samples with the same object features as the original training samples and weakening the background classification contribution are obtained by modifying the background features of the original training samples. The sample set composed of the obtained adversarial samples and the original training samples is then used as a new training set, thereby achieving the goal of optimizing training samples. The adversarial sample generation algorithm with reduced background classification contribution is shown below.
[0056]
[0057]
[0058] For the original image X, the target convolutional neural network model f(X) is input to calculate the class label of the original image. Define each element in the image The perturbation intensity range is ε∈(l,u], and the step size of each iteration is λ. First, all pixels belonging to the background region of the original image are located. Then, the corresponding position in the pixel classification contribution value matrix is queried to determine whether the element contributes positively or negatively to the classification. The positions of all elements in the background region that contribute positively to the classification are recorded; these are the elements that can be tampered with. Next, all tamperable elements are tampered with using the following method: Next, the altered element values Limited to After the image is processed within the specified range, the modified image X' for this round is obtained. Next, it is checked whether the modified image X' obtained in this round is the same as the input image for this round, and the modified image X' is then input into the target convolutional neural model to calculate the category label. If they are the same, then the two images are identical. This indicates that the tampering failed; if the two images are different and This indicates successful tampering; the current image X' is the final tampered image. If the two images are different and... Then the next round of tampering will continue.
[0059] The method for optimizing training sample quality proposed in this invention differs fundamentally from traditional methods such as sample scaling, sample reversal, and random background replacement. The main difference lies in the following: Traditional methods essentially rely on increasing the number of training samples as the basis for optimization, lacking a clear optimization objective and exhibiting a degree of arbitrariness. The newly generated training samples often fail to accurately reduce the network model's excessive focus on background features in its classification calculations, and a certain number of new training samples are often required to achieve the goal of optimizing network model performance. In contrast, the method proposed in this invention is based on the classification criteria of the input image. Each generated adversarial sample retains the same object features as the original training sample; the only change is the background, and the background's impact on the target network model's output classification value is reduced. Therefore, using the sample set composed of adversarial samples and the original training samples as a new training set can purposefully achieve the goal of reducing the network model's excessive focus on background features in its classification behavior. For example, suppose the expression l = p + q represents a fitting calculation of the target network model, where: l is the label of the training sample, p is the fitted value of the object, and q is the fitted value of the background. Since the new training samples and the original training samples have the same object features and the same labels, when the fitted value q of the background in the new training samples decreases relative to the original training samples, the fitted value p of the objects must increase to maintain the same labels for the new and original training samples. Based on the above conclusions, the method for optimizing training sample quality proposed in this invention is a method with an optimization objective. In practical applications, improving the original training samples while adding an equal amount of adversarial samples can effectively make the target network model more focused on object features during training and learning. Based on this, not only can the accuracy of the network model be improved, but its robustness can also be enhanced, making it a practical method for improving network model performance.
[0060] To objectively and systematically verify the ability of the proposed method for improving the accuracy and robustness of convolutional neural network models based on optimized training samples, this invention conducted extensive experiments using the widely used public dataset CIFAR-10 in computer vision tasks and different convolutional neural network model architectures such as AlexNet, DenseNet, and ResNet. The experiments objectively evaluated the performance, such as availability and transferability, of the designed method for improving the adversarial robustness of convolutional neural network models. The evaluation process was as follows: First, adversarial examples with the function of weakening the contribution value of background classification were used as a supplement to the original training samples, and a top-1 convolutional neural network model f(X) was retrained. Then, representative white-box attack methods with strong attack capabilities, such as C&W and PGD, were selected as attack methods to attack the retrained convolutional neural network model f(X). Finally, the attack success rates of C&W and PGD methods on the retrained convolutional neural network model were statistically analyzed, and the performance of the designed method for improving the adversarial robustness of convolutional neural network models was evaluated based on the statistical attack success rates. In particular, to fully demonstrate the performance of the adversarial robustness improvement method for convolutional neural network models designed in this invention, experiments were conducted to compare and evaluate the adversarial training method designed in this invention with the existing representative method MI-FGSM in terms of robustness, accuracy, and transferability. The convolutional neural network models used in the experiments and the evaluation samples are as follows:
[0061] (1) Original model f 0 (X): A top-1 convolutional neural network model obtained by training for 40 epochs using the original CIFAR-10 training set based on the AlexNet network model structure.
[0062] (2) Attack Sample: Using the original network model f 0 (X) is used as the target convolutional neural network model, and the original test sample X from CIFAR-10 is used as the input to the attack method. Existing attack methods C&W and PGD are used to perform a non-targeted attack on ||X'-X||. ∞ The adversarial sample X' generated within the norm that has the smallest difference from the original test sample X.
[0063] (3) Accuracy test sample: the original test set in CIFAR-10.
[0064] (4) Adversarial examples with the function of weakening the background classification contribution value: using the original model f 0(X) is the target network model. Following the method of generating adversarial samples with the function of weakening the background classification contribution value, a series of adversarial samples X' are generated with each original training sample in CIFAR-10 as input and under different tampering intensities ε∈(0,1 / 3], ε∈(1 / 3,2 / 3] and ε∈(2 / 3,1], where the step size λ is 0.01.
[0065] The comparative evaluation results of the adversarial robustness, accuracy, and transferability of the adversarial training samples of the adversarial-trained convolutional neural network model are shown in Tables 1, 2, and 3. Table 1 shows the comparative evaluation results of the adversarial robustness of the AlexNet model retrained by the adversarial training method designed in this invention and the existing MI-FGSM adversarial training method in the attack methods C&W and PGD, where: f 0 (X) represents the original model, f 2500 (X), f 5000 (X), f 7500 (X) and f 10000 (X) represents 2500, 5000, 75000, and 10000 adversarial examples with the function of weakening the background classification contribution value or adversarial examples generated by the existing method MI-FGSM, randomly added to each class in the original CIFAR-10 training set, and an optimal (top-1) model is obtained after more than 40 epochs of training; Tables 2 and 3 are respectively based on the original model f 0 Based on (X), the transferability evaluation results of the adversarial training method designed in this invention and the adversarial samples generated by the existing MI-FGSM adversarial training method in ResNet and DenseNet network structures are compared. Among them, the ResNet model and DenseNet model in the experiment are the best (top-1) models obtained after more than 40 epochs of training; in the experiment, the number of attack samples for the attack methods C&W and PGD is 500 in each category, and the test samples for accuracy are the same as the original test samples of CIFAR-10.
[0066] Table 1
[0067]
[0068] Table 2
[0069]
[0070]
[0071] Table 3
[0072]
[0073] Observing the experimental results shown in Table 1, it can be seen that as more adversarial examples are added to the original training samples, the attack success rates of attack methods C&W and PGD significantly decrease in the different AlexNet models obtained through retraining. This experimental result indicates that both the adversarial training method designed in this invention and the existing adversarial training method IM-FGSM can improve the adversarial robustness of convolutional neural network models. In particular, compared to the existing adversarial training method IM-FGSM, the adversarial training method designed in this invention not only makes the retrained different AlexNet models more robust, but also, as shown in Table 1, makes the retrained models have more robust accuracy. For example, as more adversarial examples are added to the original training samples, the adversarial training method IM-FGSM causes the attack success rates of attack methods C&W and PGD to decrease significantly from the original AlexNet model f... 0 In (X), the success rate was reduced from over 80% to an average of 24% and 17%, respectively. Compared to the existing method MI-FGSM, the adversarial training method designed in this invention reduces the average attack success rate of attack methods C&W and PGD to 20% and 12%, respectively, and the accuracy of different retrained AlexNet models remains almost unchanged, all around 88%. Furthermore, observing the experimental results shown in Tables 2 and 3, it can be seen that with the original model f 0 Based on (X), the adversarial training method designed in this invention and the adversarial examples generated by the existing MI-FGSM adversarial training method both exhibit transferability in ResNet and DenseNet network structures. However, compared to the MI-FGSM adversarial training method, the adversarial training method designed in this invention enables convolutional neural networks with different structures to achieve stronger adversarial robustness and more robust accuracy. Overall, compared to existing adversarial training methods, the method designed in this invention benefits from the generated adversarial examples that weaken the background classification contribution value, making it a practical adversarial training technique that balances robustness and accuracy. Experimental results fully demonstrate that the training set generated using the optimized training sample method proposed in this invention has knowledge transferability and is a method that can be widely applied in industry.
[0074] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for improving network model performance based on optimizing training sample quality, characterized in that: include: Inputting an original training set into a target network, and outputting a corresponding saliency map for each original training sample based on a guid-Xmage interpretative method; the original training set includes a plurality of original training samples; Based on the saliency map, background pixels of the original training sample are iteratively tampered with in different tampering intensity ranges in a monotonically decreasing manner to obtain adversarial samples with a function of weakening background classification contribution values; A new training set is constructed based on the adversarial samples and the original samples, and the performance of the network model is improved based on the training set.
2. The method for improving network model performance based on optimizing training sample quality according to claim 1, characterized in that: The process of outputting the corresponding saliency map for each original training sample based on the guid-Xmage interpretability method includes: Building a symbolic model based on the original training samples; Extracting weights corresponding to symbolic variables in the symbolic model; Constructing a pixel weight matrix based on the weights; Performing a dot multiplication on the pixel weight matrix and the pixel value of the original training image to obtain a pixel classification contribution value matrix; A saliency map corresponding to the original training sample is obtained based on the pixel classification contribution value matrix.
3. The method for improving network model performance based on optimizing training sample quality according to claim 2, characterized in that: The process of constructing a symbolic model based on the original training samples includes: Inputting the original training sample into the target convolutional neural network to obtain a target network model, and recording execution trajectory information; A symbolized image is defined based on the original training sample, and the symbolized image accurately executes the trajectory information to obtain a symbolized model.
4. The method for improving network model performance based on optimizing training sample quality according to claim 3, characterized in that: The expression of the symbolic model is: In the formula, is a symbolic model, Z is a symbolic image, is a symbolic variable in the symbolic image Z, i and j represent symbolic variables The corresponding pixel x ij ∈X, k is the kth channel of the input real image X, and ρ is an AND symbol variable An irrelevant constant representing the cumulative value of the bias in the target network model f(X), is an element in the pixel weight matrix, S indicates that the current f(Z) is a symbolic model obtained based on the original network model, and t indicates that the t-th classification branch is the target of symbolic execution.
5. The method for improving network model performance based on optimizing training sample quality according to claim 2, characterized in that: The pixel values at the same position of different channels in the pixel classification contribution value matrix are summed and normalized to the maximum and minimum values to obtain the corresponding distribution map.
6. The method for improving network model performance based on optimizing training sample quality according to claim 2, characterized in that: The pixel classification contribution value matrix is processed based on a piecewise averaging method to obtain a saliency map corresponding to the original training sample.
7. The method for improving network model performance based on optimizing training sample quality according to claim 6, characterized in that: The process of obtaining the adversarial sample based on tampering includes: Calculate the category label of the original image and initialize the perturbation intensity range and iterative tampering steps for each element in the original image; Determine the contribution of each element based on the classification contribution value matrix to obtain the tamperable element; The tamperable element is tampered to obtain an element value, and the element value is limited to obtain a tampered image of this round; Calculate the category label of the tampered image in this round; Compare the category label of the tampered image in this round with the category label of the original image. When the two category labels are different, the tampering is successful and the adversarial sample is obtained; when the two category labels are the same, the tampering fails and the next round of tampering is performed.
8. The method for improving network model performance based on optimizing training sample quality according to claim 7, characterized in that: The calculation formula for tampering the tamperable element is: In the formula, Indicates the tampered element value. represents the element in the original image, and λ represents the step size of iterative tampering.
Citation Information
Patent Citations
Image classification method based on adversarial sample training
CN115908898A
Unsupervised variable level adversarial representation learning method for multivariable time series
CN118332340A
Methods and systems to train neural networks
WO2021173715A1
