A Label-Only Model Reverse Attack Method Based on Conditional Diffusion Model
By using a conditional diffusion model to generate multiple accurate, realistic, and similar privacy images under label-only access permissions, this technology solves the problem of generating single images with poor results in existing technologies. It achieves powerful model reverse attack capabilities and is suitable for practical application scenarios.
Patent Information
- Application Number
- CN202311391869.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-25
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2043-10-25
AI Technical Summary
Existing model reverse engineering methods are insufficient in terms of the accuracy, realism, and similarity of the generated images, and the number of generated images is limited, which cannot meet the attack requirements of real-world application scenarios. In particular, it is difficult to achieve effective privacy data theft under label-only access permissions.
A conditional diffusion model is adopted. By obtaining label-only access to the target model, selecting an auxiliary dataset, obtaining the target label and the predicted labels of the auxiliary samples, training the U-net model to perform forward diffusion and backward prediction, generating multiple images, performing gamma correction and random transformation, and outputting the most robust generated image.
Images generated under label-only access permissions are more accurate, realistic, and similar, effectively recovering private images. Furthermore, generating multiple images facilitates subsequent research and analysis, surpassing the effectiveness of existing white-box and black-box attacks.
Smart Images

Figure CN117744130B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data privacy and security, and specifically relates to a label-only model reverse attack method based on a conditional diffusion model. Background Technology
[0002] Deep learning has been widely explored and applied in fields such as government affairs, finance, facial and fingerprint recognition, and biomedical diagnostics. Building such deep learning models requires a large amount of training data, which may contain sensitive information, such as facial images. Research shows that models tend to memorize training data, thus highlighting the growing security concerns associated with deep learning.
[0003] Model inversion attack is a significant attack method targeting the data privacy of artificial intelligence algorithms. Attackers utilize APIs provided by machine learning systems (such as model output and parameter updates) to obtain preliminary information about the model. They then use this preliminary information to perform reverse analysis and obtain private data within the model. For example, a facial recognition model can be used to reconstruct the facial image of a target individual in the training set; similarly, the prediction vector generated by the facial recognition model can be used to reconstruct the input facial image. Alternatively, a medical prediction model can be used to infer sensitive attributes of an individual's genome.
[0004] In recent years, some scholars have proposed model inversion attack methods based on deep learning techniques such as Generative Adversarial Networks (GANs) and Auto-Encoders (AEs). However, existing model inversion attack methods still have some shortcomings. For example, the paper "Knowledge-enriched distributional model inversion attacks" by Si. Chen et al. proposes a model inversion attack method that guides GAN training with soft labels under white-box access permissions. This method is time-consuming to train the attack model, difficult to achieve transfer, and makes overly strong assumptions that do not conform to reality. The paper "Neural network inversion inadversarial setting via background knowledge alignment" by Yang Ziqi et al. proposes an inverse model based on the AE framework, which successfully recovers the privacy training set from the prediction vector under black-box access permissions. This method requires multiple accesses to the target model, and the generated image classification accuracy is low, the realism is poor, and the facial contours are unclear. The paper "Label-only Model Inversion Attacks: Attack with the Least" by Tianqing Zhu et al. The Information article proposes a method to achieve model inversion attack using minimal information. The authors recover the confidence vector by training a shadow model under label-only access, and then train an inverse model to carry out the attack. This method has the same attacker access permissions as the present invention, but it requires multiple optimizations and the generated images have poor recognizability. In addition, the above three attack methods can only generate a unique representative image for a specific target, which is not conducive to subsequent research and analysis. Summary of the Invention
[0005] To address the shortcomings of existing methods, this invention provides a label-only model reverse attack method based on a conditional diffusion model. It considers the security issue of attackers stealing user privacy data using model reverse attacks in practical deep learning model application scenarios. This method solves the problems of limited generative model capabilities and optimization algorithms in previous methods, resulting in single generated images and low similarity. It improves the accuracy, realism, and similarity of generated images while increasing the number of generated images. Specifically, the method involves: first, obtaining label-only access permissions to the target model; selecting an auxiliary dataset; obtaining the target label; obtaining the predicted labels of the auxiliary samples; using the predicted labels of the target model as conditions to guide the training process of the diffusion model, instructing the diffusion model to learn the noise distribution added to the predicted target during forward diffusion and backward prediction; inputting the target label, random standard normal distribution noise, and a predefined guidance strength into the trained conditional diffusion model to obtain multiple generated images corresponding to the target label, and performing gamma correction on the generated images; and selecting robust generated images through random transformation. This invention solves the problems of poor similarity and limited number of generated images in previous model reverse attacks, and can even steal training images when only the target model's predicted labels are available.
[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is: a label-only model reverse attack method based on a conditional diffusion model, comprising the following steps:
[0007] S1. Obtain label-only access to the target model: The target model has been trained using privacy images and has image classification capabilities. The privacy images are kept confidential. The label-only access is limited to querying the target model's classification results for the input image.
[0008] S2. Select auxiliary dataset: The auxiliary dataset refers to the auxiliary dataset related to the target model task. For example, if the target model is a face recognition model, the auxiliary dataset is a face dataset.
[0009] S3. Obtain target labels: The target labels are the classification results of the privacy images by the target model;
[0010] S4. Obtain the predicted labels of the auxiliary samples: The predicted labels of the auxiliary samples are the classification results of the target model on the auxiliary dataset samples;
[0011] S5. Training the Conditional Diffusion Model for Attacks: The conditional diffusion model uses the U-net model, whose training process is divided into forward diffusion and backward prediction. Forward diffusion adds noise, and backward prediction removes noise. During forward diffusion, Gaussian noise is added to the auxiliary samples, repeated T times, resulting in noise with a standard normal distribution. Therefore, the noise data obtained after adding noise T times can be directly derived from the original data. During backward prediction, the U-net model is used to predict the noise added during forward diffusion. Timestamps between 1 and T are randomly selected, and the loss between the real Gaussian noise and the noise added from the original data to the noise data is calculated. To ensure that the training is guided by the prediction labels, coded labels are added to the timestamps of each step, enabling the diffusion model to learn the noise distribution added to the prediction labels.
[0012] S6. Obtain generated images of multiple specific targets: Input multiple images with standard normal distribution noise, target labels, and guidance strength into the trained conditional diffusion model to obtain multiple generated images corresponding to the target labels, and perform gamma correction on the generated images.
[0013] S7. Output Results: Randomly transform the corrected generated image and input it into the target model for prediction. Repeat this process M times and select the top k most robust generated images as the final output.
[0014] Furthermore, in step S5, training the conditional diffusion model consists of two processes: a forward diffusion process, which performs T-step noise addition to transform the original image into noise distributed in a standard normal distribution; and a backward prediction process, which uses the prediction labels to predict the noise added during forward diffusion. The specific steps are as follows:
[0015] S5.1. Set the diffusion step number T and Gaussian noise parameters β0, β1, ..., β T ;
[0016] S5.2 Calculate the diffusion coefficient based on the current diffusion step t.
[0017] S5.3, Input auxiliary sample x0∈D aux Calculate the noisy output of the diffusion at step t. Where ε represents noise, and N(0,I) represents a standard Gaussian distribution;
[0018] S5.4. Use the U-net model for prediction, input x t Predict x0 to x t noise ε θ (x t ,t);
[0019] S5.5 Calculate the mean square error (MSE) loss between the predicted noise and the random Gaussian noise, as shown below:
[0020]
[0021] Where ε represents the Gaussian noise added from 0 to t times, ε θ For predicting noise, θ represents the parameters of the target model;
[0022] S5.6 Calculate the predicted label F embedded in S2 w The MSE loss between the predicted noise and random Gaussian noise at the timestamp t of (x0) is shown below:
[0023]
[0024] The timestamp t is encoded into a preset dimension using the position_encoding() function in the transformer, while the label F is predicted. w (x0) also uses the embedding() function to convert the dimension to a preset dimension;
[0025] S5.7, Iterative optimization of real Gaussian noise and from x0 to x t The loss between predicted noise is used to obtain the trained conditional diffusion model G. θ The optimization objectives are as follows:
[0026]
[0027] The predicted labels are trained with a certain probability p without guiding the diffusion model.
[0028] Furthermore, in step S6, generated images of multiple specific targets are obtained, and the specific steps are as follows:
[0029] S6.1 Input the target label l and the noisy image z into the trained diffusion model G θ The predicted noise is as follows:
[0030]
[0031] Where ω represents the guidance strength of the target label l, x t The noise image for round t;
[0032] S6.2. Use predicted noise to gradually denoise the noisy image, as shown in the following formula:
[0033]
[0034] S6.3 After T iterations, representative generated images G of multiple target labels are obtained.θ (z), where the number of times noise is added in wheel T is the same as in S5;
[0035] S6.4 Perform gamma correction on the generated image, i.e., G θ (z)→A·G θ (z) γ To adjust its matching to human eye perception, parameter A is typically 1 and γ = 2.2.
[0036] Furthermore, in step S7, the generated images are randomly transformed, and the top k most robust generated images are output as the final result. The specific steps are as follows:
[0037] S7.1, randomly crop and flip vertically or horizontally with a certain probability G θ (z) γ G θ (z) γ →T(G θ (z) γ T(·) is a random transformation function;
[0038] S7.2, Convert the generated image T(G) θ (z) γ Input target model F w The prediction process is repeated M times.
[0039] S7.3 Calculate the representative weight of each generated image, as shown below:
[0040]
[0041] The δ function returns 1 when its inputs are equal and 0 when they are not equal.
[0042] S7.4 Select the generated images G corresponding to the top k largest representative weights. θ (z) γ That is, the final output.
[0043] Compared with the prior art, the advantages of this invention are:
[0044] (1) This invention is the first to develop a practical and powerful attack model for model reverse engineering under label-only access permission. The attack method used in this invention is based on the conditional diffusion model, which can effectively recover privacy images in the training set of the labeled model. At the same time, compared with existing model reverse engineering methods, the images generated by this invention are more accurate, more realistic and more similar.
[0045] (2) The attacker's capability assumptions in this invention are closer to reality and have more practical significance. Compared with previous white-box access control and black-box access control based on confidence vectors, this invention achieves better attack performance than existing black-box attacks, and is comparable to the most advanced white-box attacks, even when the attacker can only obtain the predicted labels of the target model.
[0046] (3) The attack method proposed in this invention can generate multiple images for the target label without requiring further optimization of the generated images. Existing model reverse attack methods require further optimization of the generated images to achieve the attack effect, while this invention only needs to randomly transform and then filter the generated multiple images to achieve a high attack effect. Attached Figure Description
[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0048] Figure 1 This is a flowchart of the label-only model reverse attack method based on the conditional diffusion model proposed in this invention;
[0049] Figure 2 This is a schematic diagram of the structure of the target model to be attacked by the method of the present invention;
[0050] Figure 3 This is a schematic diagram of the auxiliary dataset image in step S2 of embodiment 1 of the present invention;
[0051] Figure 4 This refers to the privacy training data image in Embodiment 1 of the present invention;
[0052] Figure 5 This is a schematic diagram of the training process of the conditional diffusion model of the present invention;
[0053] Figure 6 This is a schematic diagram illustrating the process of generating privacy images using a trained conditional diffusion model according to the present invention.
[0054] Figure 7 The image is recovered using the model reverse attack method used in this invention. Detailed Implementation
[0055] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0056] To address the security issues posed by malicious actors using model reverse engineering attacks to steal privacy in cyberspace, this example provides a label-only model reverse engineering method based on a conditional diffusion model, such as... Figure 1 As shown, it includes the following steps:
[0057] S1. Obtain label-only access to the target model: The target model has been trained using privacy images and has image classification capabilities. The privacy images are kept confidential. The label-only access is limited to querying the target model's classification results for the input image. In this example, the specific structure of the target model is as follows: Figure 2 As shown, it consists of the following two parts:
[0058] (1) A feature extractor consisting of four convolutional groups is used to extract depth features from the input 64×64 image. Each convolutional group includes a convolutional layer, a batch normalization layer, a max pooling layer and a ReLU activation layer. The number of channels in the three convolutional modules are 128, 256, 512 and 1024 respectively.
[0059] (2) A classifier consisting of two fully connected layers is used to classify the extracted deep features. The input feature dimension is 16384, the hidden layer dimension is 2650, and the output dimension is 530.
[0060] S2. Selecting an Auxiliary Dataset: The auxiliary dataset refers to a dataset related to the target model task. Compared to previous model reverse engineering methods that required auxiliary datasets with the same distribution, this invention has more lenient requirements for the auxiliary dataset, only requiring data from the same domain. For example, if the target model is a face recognition model, the auxiliary dataset is a dataset containing faces. In this example, the privacy dataset uses the FaceScrub dataset, and the auxiliary dataset uses the CelebA dataset. Auxiliary image samples are as follows... Figure 3 As shown, the size is 64×64.
[0061] S3. Obtain Target Labels: The target labels are the classification results of the privacy images by the target model. In this example, the target model is a face recognition model, trained using 80% of the privacy dataset and tested using 20% of the privacy dataset. The privacy images used are as follows: Figure 4 As shown, the labels are the names of the people corresponding to the privacy images.
[0062] S4. Obtain the predicted labels of the auxiliary samples: The predicted labels of the auxiliary samples are the classification results of the target model on the auxiliary dataset samples. In this example, they will be... Figure 3 The samples from the auxiliary dataset shown are input into the target face recognition model, and the resulting classification is the predicted label.
[0063] S5. Training the Conditional Diffusion Model for Attacks: The conditional diffusion model adopts the U-net model, which consists of dual convolutional blocks, 3 downsampling blocks, 2 bottom dual convolutional blocks, 3 upsampling blocks, and one convolutional layer. The training of the conditional diffusion model is divided into two processes: a forward diffusion process, which performs T-step noise addition to transform the original image into standard normally distributed noise; and a backward prediction process, which uses prediction labels to predict the noise added by the forward diffusion.
[0064] The training process of the conditional diffusion model in this example is as follows: Figure 5 As shown in the figure, the forward diffusion process from left to right at the top involves adding Gaussian noise to the auxiliary sample x0, performing this process T times, transforming x0 into noise x0 with a standard normal distribution. T Backward prediction primarily predicts the noise added by forward diffusion, utilizing the prediction label F. w (x0) guides the prediction process, such as Figure 5 The lower part is shown, and the MSE loss between the predicted noise and random Gaussian noise ε ~ N(0,I) is calculated. The model is continuously optimized. The specific steps for training the conditional diffusion model are as follows:
[0065] S5.1. Set the diffusion step number T and Gaussian noise parameters β0, β1, ..., β T This example uses the linspace() function to create a uniformly spaced parameter sequence, where β0 = 1e-4, β T =0.02, T=1500;
[0066] S5.2 Calculate the diffusion coefficient based on the current diffusion step t.
[0067] S5.3, Input auxiliary sample x0∈D aux Calculate the noisy output of the diffusion at step t. Where ε represents noise, and N(0,I) represents a standard Gaussian distribution, this formula shows the noise data x after adding t noises. t It can be directly derived from the original data x0, as long as t is large enough. The limit of x will approach 0, causing x to... t Approximates a standard normal noise distribution;
[0068] S5.4 This example uses the U-net model for prediction, with the input x t Predict x0 to x t noise ε θ (x t ,t), where the U-net model iterates 300 times with a learning rate of 3e-4, and the learning rate is reduced to 1e-4 in the last 50 iterations;
[0069] S5.5 Calculate the loss between the predicted noise and the random Gaussian noise. This example uses the MSE loss function, as shown below:
[0070]
[0071] Where ε represents the Gaussian noise added from 0 to t times, ε θ For predicting noise, θ represents the parameters of the target model;
[0072] S5.6 Calculate the predicted label F embedded in S2 w The MSE loss between the predicted noise and random Gaussian noise at the timestamp t of (x0) is shown below:
[0073]
[0074] The timestamp t directly affects the distribution of noise added at each step. To ensure that the training process is guided by the predicted labels, this example encodes the predicted labels and timestamps into the same dimension and adds them together, setting the dimension to 256.
[0075] S5.7, Iterative optimization of real Gaussian noise and from x0 to x t The loss between predicted noise is used to obtain the trained conditional diffusion model G. θ The optimization objectives are as follows:
[0076]
[0077] The predicted labels are trained with a certain probability p, which does not guide the diffusion model. In this example, p = 0.1.
[0078] S6. Obtain generated images of multiple specific targets: Input multiple images with standard normal distribution noise, target labels, and guidance strength into the trained conditional diffusion model to obtain multiple generated images corresponding to the target labels. In order to make the generated images more robust, gamma correction is performed on the generated images.
[0079] In this example, the target label l and the noise image z following a standard normal distribution are input into the conditional diffusion model G trained by S5. θ ,like Figure 6 As shown, the specific steps are as follows:
[0080] S6.1 Input the target label l and the noisy image z into the trained diffusion model G θ The predicted noise is as follows:
[0081]
[0082] Where x tLet ω represent the noise image of round t, and let ω represent the guidance strength of the target label l. As can be seen from the formula, the larger ω is, the larger the proportion of conditional prediction noise, and the closer the generated image features are to the target. However, this comes at the cost of the quality of the generated image. In this example, we analyze the value of ω. Too high a value of ω will lead to a decrease in the quality of the generated image, but it is beneficial to the generation of target features. Therefore, after weighing the accuracy, realism and similarity, this example sets ω = 4.
[0083] S6.2. Use predicted noise to gradually denoise the noisy image, as shown in the following formula:
[0084]
[0085] S6.3 After T iterations, a representative generated image G of the target label is obtained. θ (z);
[0086] S6.4. Since the diffusion model trained with the auxiliary dataset performs differently from the model trained with the privacy dataset, this invention performs secondary gamma correction on the generated images, i.e., G... θ (z)→A·G θ (z) γ This example analyzes the impact of γ on the generated image and finds that increasing γ has a positive impact on attack accuracy and K-Nearest Neighbor distance, but has a positive impact on human visual perception similarity that is first positive and then negative. Therefore, it is necessary to weigh the γ value. In this example, A=1 and γ=2.2 are set.
[0087] S7. Result Output: Due to the large number of images generated by this invention, the corrected generated images can be randomly transformed and input into the target model for prediction. This process is repeated M times, and the top k most robust generated images are selected as the final output. The random transformation and result output process in this example is as follows: Figure 6 The right half shows the specific steps as follows:
[0088] S7.1, random cropping and vertical or horizontal flipping with a certain probability G θ (z) γ G θ (z) γ →T(G θ (z) γ T(·) is a random transformation function. In this example, the horizontal flip probability is set to 0.5 and the vertical flip probability is set to 0.1.
[0089] S7.2, Convert the generated image T(G) θ (z) γ Input target model F w The prediction process is repeated M times.
[0090] S7.3 Calculate the representative weight of each generated image, as shown below:
[0091]
[0092] The δ function returns 1 when its inputs are equal and 0 when they are not equal.
[0093] S7.4 Select the generated images G corresponding to the top k largest representative weights. θ (z) γ The final output, in this example, is the restored image as follows: Figure 7 As shown.
[0094] Specifically, a quantitative analysis was performed on the recovered image in this example, and the results are shown in Table 1. The Top1 accuracy represents the proportion of all samples in which the evaluation model can classify the recovered sample into the target class, the Top5 accuracy represents the proportion of all samples in which the top 5 confidence scores contain the target class, the K-nearest neighbor distance represents the shortest feature distance between the recovered sample and the real target sample, and the Fraser distance represents the feature similarity between the recovered sample and the real sample.
[0095] Table 1: Quantitative Assessment Table of Recovery Samples
[0096] Top 1 accuracy Top 5 accuracy K-nearest neighbor distance Fraser distance 56.13% 84.54% 959.3208 109.5714
[0097] In summary, this invention addresses the problem of model reverse engineering attacks. Existing methods cannot simultaneously meet the following requirements: 1. The method has good attack effectiveness, ensuring that the generated images have a certain degree of realism, accuracy, and similarity. 2. The method is practical and closely resembles real-world application scenarios, as attackers typically only have access to deep learning models and only know the model's predicted labels. 3. The method can generate multiple images for a single target label, facilitating subsequent sampling, evaluation, and analysis.
[0098] To meet the above three requirements, this invention proposes a label-only model reverse attack method based on a conditional diffusion model, addressing the issue from two aspects. First, to closely resemble real-world application scenarios, this invention assumes that the attacker can only obtain the predicted labels of the target model to launch the attack. Second, to ensure that the generated images are more accurate, realistic, and similar, and facilitate subsequent analysis and research, this invention uses a conditional diffusion model guided by the target label to generate multiple privacy-preserving images. This is the first powerful and practical attack model trained under the scenario where the attacker only knows the labels. The generated multiple images have high accuracy, facilitating subsequent sampling, evaluation, and analysis.
[0099] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should be protected by the present invention.
Claims
1. A label-only model reverse attack method based on a conditional diffusion model, characterized in that, Includes the following steps: S1. Obtain tag-only access permissions for the target model to be attacked; S2. Select an auxiliary dataset; S3. Obtain the target label, which is the classification result of the target model on the privacy image; S4. Obtain the predicted labels of the auxiliary samples, wherein the predicted labels of the auxiliary samples are the classification results of the target model on the auxiliary dataset samples; S5. Training the conditional diffusion model for the attack, wherein the conditional diffusion model adopts the U-net model, and its training process is divided into forward diffusion and backward prediction. Forward diffusion performs noise addition, and backward prediction performs noise removal. During forward diffusion, Gaussian noise is added to the auxiliary samples, repeated T times to obtain noise with a standard normal distribution. During backward prediction, the U-net model is used to predict the noise added during forward diffusion. Timestamps between 1 and T are randomly selected to calculate the loss between the real Gaussian noise and the noise added from the original data to the noisy data. Encoded labels are added to the timestamps of each step. The specific process of S5 is as follows: S5.1, Set the diffusion step number T and Gaussian noise parameters ; S5.2 Calculate the diffusion coefficient based on the current diffusion step t. ; S5.3 Input auxiliary samples Calculate the noisy output of the diffusion at step t. ,in Indicates Gaussian noise. Indicates a standard Gaussian distribution; S5.
4. Use the U-net model for prediction, input... ,predict arrive noise ; S5.5 Calculate the MSE loss between the predicted noise and the random Gaussian noise with respect to the parameters. The gradient is shown below: in To predict noise, Represents the parameters of the target model; S5.6 Calculate the predicted label embedded in S2 The MSE loss between prediction noise and random Gaussian noise at time stamp t with respect to parameters The gradient is shown below: The timestamp t is used In The function encodes time into a preset dimension while predicting labels. Also used The function converts the dimension to a preset dimension; S5.7, Iterative optimization of real Gaussian noise and from arrive The loss between predicted noise is used to obtain the trained conditional diffusion model. The optimization objectives are as follows: The predicted label is based on probability. Unguided diffusion model training; S6. Using the trained conditional diffusion model, obtain the generated image corresponding to the target label, and perform gamma correction on the generated image; S7. Perform random transformations on the corrected generated image and input it into the target model for prediction to obtain the final output result.
2. The label-only model reverse attack method based on the conditional diffusion model according to claim 1, characterized in that, The target model described in S1 has been trained using privacy images and has image classification capabilities. The privacy images are kept confidential. The label-only access permission is limited to querying the target model's classification results for the input image.
3. The label-only model reverse attack method based on the conditional diffusion model according to claim 2, characterized in that, The auxiliary dataset mentioned in S2 refers to the auxiliary dataset related to the target model task.
4. The label-only model reverse attack method based on the conditional diffusion model according to claim 3, characterized in that, The specific process of S6 is as follows: S6.1, Target Label and noisy images Input into the trained diffusion model The predicted noise is as follows: in Indicates target label The intensity of guidance, The noise image for round t; S6.
2. Use predicted noise to gradually denoise the noisy image, as shown in the following formula: S6.3 After T iterations, representative generated images of multiple target labels are obtained. The number of times noise is added to the T wheel is the same as that in S5; S6.4 Perform gamma correction on the generated image, i.e. , where parameters =1, =2.
2.
5. The label-only model reverse attack method based on the conditional diffusion model according to claim 4, characterized in that, The specific process for S7 is as follows: S7.1, random cropping and vertical or horizontal flipping ,Right now It is a random transformation function; S7.2, Convert the generated image Input target model Make predictions, repeating M times; S7.3 Calculate the representative weight of each generated image, as shown below: in, The function returns 1 if its inputs are equal, and 0 if they are not. S7.4 Select the generated images corresponding to the top k largest representative weights. That is, the final output.
Citation Information
Patent Citations
Cross-domain model reverse attack method
CN115718929A
Cross-domain black box model reverse attack method
CN115759190A
