New media content privacy protection system based on multi-objective optimization adversarial sample generation
Through multi-objective optimization of the adversarial sample generation system, combined with the target model module of CNN and ViT networks, the problem of insufficient quality and migration of adversarial sample generation in the existing technology is solved, and high-quality and highly migratory adversarial sample generation is achieved, which improves the adversarial sample generation with high quality and strong migration. The effect of image privacy protection.
Patent Information
- Application Number
- CN202411158319.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-22
- Publication Date
- 2024-11-26
AI Technical Summary
Existing new media image privacy protection technology based on adversarial samples is difficult to generate high-quality and highly migratory adversarial samples, and adversarial sample generation may affect image quality, resulting in poor visual effects and poor migration between CNN and ViT architectures.
The multi-objective optimization adversarial sample generation system is adopted to perform image preprocessing through the input module. The multi-objective optimization adversarial perturbation generation module uses evolutionary algorithms and adaptive evaluation functions to generate adversarial samples with deceptive and good visual quality, and input them into the target model. The module conducts attacks, and the target model module is built by a combination of CNN and ViT networks to optimize the anti-perturbation to improve migration.
Generate adversarial samples that can effectively deceive multiple models, maintain image visual quality, improve privacy protection integrity and migration, and ensure the recognition ability of images between different architectures.
Smart Images

Figure CN119026167A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image privacy protection systems, and in particular to a new media content privacy protection system based on multi-objective optimization adversarial sample generation. Background Art
[0002] The research on online user image privacy protection has important value and significance, especially in today's information and digital society. With the popularity of social media, online communities and digital services, users share a lot of personal information and images on the Internet. However, this information often involves private content, and protecting the user's image privacy becomes particularly important. Unauthorized image acquisition and use may lead to the leakage and abuse of personal privacy, such as identity theft, phishing fraud, etc.
[0003] The application of deep learning in privacy protection, especially de-identification technology, has attracted much attention in recent years. It aims to protect the privacy of users while processing personal data. Among them, privacy protection technology based on adversarial samples is a method of processing personal data using adversarial generative models to protect data privacy. This technical background involves the intersection of deep learning, adversarial learning and privacy protection. Adversarial samples are samples that make incorrect predictions by deep learning models by making slight modifications to the original input samples. These modifications usually do not affect human recognition ability, but can deceive machine learning models. Privacy protection technology based on adversarial samples attempts to modify the data covertly so that the original data cannot be accurately identified and inferred by deep learning models. This method can effectively protect data privacy while maintaining data availability. Privacy protection technology based on adversarial samples faces the challenge of difficulty in generating high-quality and highly transferable adversarial samples. In addition, the generation of adversarial samples needs to maintain the validity and availability of the data and ensure that the generated samples remain consistent with the key features of the original data.
[0004] Among the existing new media image privacy protection technologies based on adversarial samples, most are based on a single network model architecture of CNN or ViT. CNN effectively learns local features in images through convolutional layers and pooling layers, and has good spatial invariance. This makes CNN perform well in processing tasks such as image classification, object detection, and semantic segmentation. However, CNN mainly relies on local convolution operations, and is relatively limited in capturing long-distance dependencies of global images, which may lead to poor performance in some tasks that require global information. ViT learns the relationship between different positions in the image through the self-attention mechanism, which can effectively capture the global features of the image and has strong global perception capabilities. Similarly, the ViT model structure also has shortcomings such as high computational cost, reliance on large amounts of data, and sequence length restrictions. The CNN and ViT model architectures are very different, so the research of Mahm et al. shows that adversarial examples are not easy to transfer between CNN and Transformer. Therefore, it is necessary to build a source model that can deceive the two model architectures, and perform white-box attacks on it to generate adversarial samples, improve its transferability between the models of the two architectures, and meet the integrity requirements of privacy protection.
[0005] While paying attention to privacy protection, user experience is also a very important key point to note. Therefore, when adding adversarial perturbations, it is necessary to ensure high image quality to improve the visual effect of the image. At present, de-identification technology and adversarial sample generation technology often process images, such as adding noise, blurring, randomization, etc., to confuse key information in the image to protect privacy. These processes may cause the image quality to deteriorate, making the image blurry or distorted, and reducing the visual quality. The goal of de-identification processing is to blur or hide sensitive information in the image, but sometimes it may affect other important information or features in the image, making the image difficult to identify or understand, which may affect the actual use and usability of the image. The generation of adversarial samples may cause the image quality to deteriorate, causing the image to appear distorted or visually unnatural. This is because the design of adversarial samples usually involves adding small perturbations to the original image to cause the machine learning model to misjudge. Summary of the invention
[0006] The present invention provides a new media content privacy protection system based on multi-objective optimization adversarial sample generation to solve the problems that adversarial samples generated by the prior art affect the main body of the picture, have poor visual effects, and have poor migration between CNN and ViT architectures.
[0007] In order to achieve the above object, the technical solution adopted by the present invention is:
[0008] A new media content privacy protection system based on multi-objective optimization adversarial sample generation is characterized by comprising an input module, a multi-objective optimized adversarial perturbation generation module, and a target model module, wherein:
[0009] The input module converts the original image samples into serialized data after preprocessing, and generates an initial perturbation population based on the serialized data and sends it to the multi-objective optimized adversarial perturbation generation module;
[0010] In the multi-objective optimized adversarial perturbation generation module, the initial perturbation population is randomly added with perturbations for initialization; then, individuals in the initial perturbation population are randomly selected as parent individuals, and the parent individuals are evolved to generate offspring perturbation individuals; then, the generated offspring perturbation individuals are subjected to adaptability evaluation, and the problem is decomposed into three sub-problems during the adaptability evaluation, namely, minimizing the prediction probability of the true class label of the original image, minimizing the perturbation metric value of the adversarial perturbation, and minimizing the number of perturbation pixels. Thus, offspring perturbation individuals that meet the requirements of deceptiveness and visual quality are obtained through adaptability evaluation to constitute adversarial samples, and the adversarial samples are sent to the target model module to attack the target model module;
[0011] The target model module processes the input adversarial samples and generates a predicted probability result, and the target model module feeds the predicted probability result back to the multi-objective optimized adversarial perturbation generation module.
[0012] Furthermore, the preprocessing of the input module includes adjusting the image size and resolution, image denoising or smoothing, image enhancement, and normalizing or standardizing image pixel values.
[0013] Furthermore, when performing adaptability evaluation in the adversarial perturbation generation module of the multi-objective optimization, for the sub-problem of minimizing the predicted probability of the true class label of the original image, the predicted probability result output by the target model is used as the optimization index; for the sub-problem of minimizing the perturbation measure value of the adversarial perturbation, the L2 norm is used as the perturbation measure index; for the sub-problem of minimizing the number of perturbed pixels, the number of pixels with added perturbation is used as the optimization index.
[0014] Furthermore, the target model module is composed of a combination of a CNN network and a ViT network.
[0015] The present invention proposes to construct a novel dual-architecture target model module as a source model to generate adversarial samples, so as to solve the problem of poor migration between different architectures. At the same time, in order to maintain the visual effect of the main part and constrain the attack vision to be insignificant, an evolutionary learning algorithm is added to the generation adversarial sample module, and the idea of a multi-objective optimization algorithm is introduced. Not only the length of the perturbation vector is used to measure the perturbation intensity, but also the number of changed pixels is used to measure the perturbation intensity, forming a visual smoothing effect and optimizing the adversarial perturbation. Through the above method, the adversarial nature of the adversarial samples is enhanced and the visual quality of the new media images is guaranteed.
[0016] The present invention utilizes evolutionary algorithms and multi-objective optimization techniques to search for effective adversarial perturbations to deceive the target model module CTPNet. The attacker optimizes the selected perturbations by designing adaptive evaluation functions and screening strategies to achieve attacks on the target model. The present invention aims to generate adversarial samples that can effectively deceive the target model, and can be adjusted and optimized according to the needs and goals of the attacker. The adversarial samples that introduce adversarial perturbations can successfully deceive the target model while fully maintaining the visual quality of the original image, thereby better protecting privacy. At the same time, since the target model module constructed by the present invention integrates two network architectures, CNN and ViT based on the Transformer architecture, the adversarial perturbations formed by optimization are more likely to deceive the two network architectures, thereby improving the mobility of adversarial samples, enabling images to resist the recognition of multiple models, and better protecting the privacy security of images. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 It is a schematic diagram of the overall structure of the system of an embodiment of the present invention.
[0018] Figure 2 It is a schematic diagram of the target model module structure in an embodiment of the present invention. DETAILED DESCRIPTION
[0019] The present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0020] like Figure 1 As shown, this embodiment discloses a new media content privacy protection system based on multi-objective optimization adversarial sample generation, including an input module, a multi-objective optimized adversarial perturbation generation module, and a target model module. The specific description of each module is as follows:
[0021] (1) The input module converts the original image samples into a serialized data format that can be processed by the computer for subsequent processing and analysis. This process usually involves image preprocessing and feature extraction, converting image information into vectors or other forms of data representation, which facilitates applications in machine learning, computer vision or other fields.
[0022] The original image samples obtained by the input module can come from any source, such as cameras, image databases, the Internet, etc.
[0023] In the input module, the original image is first preprocessed to improve the effect of subsequent processing. Preprocessing includes but is not limited to adjusting the image size and resolution, image denoising or smoothing, image enhancement (such as adjusting contrast, brightness), and normalizing or standardizing image pixel values.
[0024] Then, the input module converts the preprocessed or original image into serialized data, specifically, the preprocessed original image samples are encoded by feature vectorization, scalarization or other encoding methods, so as to convert the image information into a format that can be processed by the computer, and then the encoded feature data is output as serialized data, and the serialized data is saved as a file or passed as input to the subsequent multi-objective optimization adversarial perturbation generation module.
[0025] The original image sample is encoded by the input module into a serialized data suitable for evolutionary algorithm processing and sent to the multi-objective optimized adversarial perturbation generation module. This encoding process can convert the original image sample into a set of numerical vectors or symbol sequences. By using the specified encoding mode, the serialized data is initialized into an initial perturbation population, which contains multiple variants after slight changes to the original sample.
[0026] The serialized data output by the input module should contain enough information to preserve the important features and structure of the original image. It should be noted that it is necessary to ensure that the encoded data has good representation and interpretability so that the subsequent models or algorithms can effectively process and analyze it. During the data processing process, the consistency and accuracy of the data should be maintained to avoid information loss or distortion to ensure the reliability of subsequent analysis.
[0027] (2) The multi-objective optimized adversarial perturbation generation module converts the multi-objective optimization problem of the complex original image samples into multiple single-objective sub-problems and uses an evolutionary algorithm to solve these single-objective sub-problems, thereby generating adversarial perturbations for the original image samples. These adversarial perturbations can make the original image samples after perturbation look similar to the original image samples, but can deceive the target model and lead to incorrect classification or output.
[0028] In the multi-objective optimization adversarial perturbation generation module, the quality of the generated adversarial perturbation is evaluated in two ways: one is whether it can deceive the target model, and the other is whether the visual quality is good, that is, whether the adversarial perturbation has good deception and good visual quality for the target model. These two goals constitute multi-objective optimization. Through the multi-objective optimization evolution direction, adversarial perturbations with high deception and good visual quality are obtained to form adversarial samples.
[0029] The adversarial perturbation generation module of this multi-objective optimization involves sub-modules such as population initialization, sub-problem decomposition, population evolution and final perturbation screening. The entire evolutionary process is controlled by defining and optimizing the objective function and adopting an effective evolutionary strategy to achieve the expected optimization goal.
[0030] The role of the multi-objective optimization adversarial perturbation generation module is to transform the complex multi-objective optimization problem into multiple single-objective sub-problems for solution, so as to balance and optimize the relationship between multiple objectives, mainly the trade-off between the attack success rate and the perturbation intensity and the number of perturbed pixels. It should be noted that the objective function of the single-objective sub-problem should be reasonably designed to ensure that each objective can be effectively measured and optimized while avoiding conflicts between objectives, and adjusting the parameters of the evolutionary algorithm (such as population size, number of iterations, mutation rate, etc.) to balance the optimization effect and computational efficiency.
[0031] The data processing process of the multi-objective optimization adversarial perturbation generation module is as follows:
[0032] Based on the initialization strategy, pixel perturbations are randomly added to the initial perturbation population to complete the initialization of the initial perturbation population, ensuring that the perturbations in the initial perturbation population have a certain diversity and coverage so that the evolutionary algorithm can effectively explore the perturbation space.
[0033] Individuals are randomly selected from the initial perturbation population as parent individuals, and the offspring perturbation individuals are generated by crossover and mutation evolution of the parent individuals. These operations simulate the genetic mechanism in biological evolution and explore new perturbation space through random combination and mutation.
[0034] For each offspring perturbation individual, the fitness evaluation function specified by the attacker is used in combination with the feedback information of the target model classifier to evaluate the fitness of each offspring perturbation individual. The goal of the evaluation is to find the most effective adversarial perturbation that causes the target model to misclassify.
[0035] In the process of adaptability evaluation, it is necessary to adopt a suitable problem decomposition method, simplify the problem and make full use of the feedback information of the target model classifier to guide the evolutionary algorithm to move in the direction of generating misclassification. That is, in the process of adaptability evaluation, this embodiment decomposes the optimization problem into three sub-problems, namely, the sub-problem of minimizing the predicted probability of the true class label of the original image, the sub-problem of minimizing the perturbation metric value of the adversarial perturbation, and the sub-problem of minimizing the number of perturbed pixels.
[0036] In each evolutionary process, according to the optimization goal, the most suitable offspring perturbation individuals are selected and saved in the current population. Due to the introduction of the concept of multi-objective optimization, the non-dominated sorting and Pareto frontier between individuals need to be continuously updated to retain the diversity and superior performance of the perturbation set.
[0037] When the maximum number of iterations is reached or the termination condition is met, the result finally retained in the population is a group of diverse optimal adversarial perturbation sets (Pareto set).
[0038] The attacker designs a perturbation screening strategy that satisfies subjective preferences and selects the best adversarial perturbation from the final perturbation set to form an adversarial sample for the subsequent target model module. This strategy may be based on the attacker's goals or requirements, such as selecting the smallest perturbation amount or the highest attack success rate.
[0039] Specifically, in this embodiment, after the evolution is completed and a disturbance population of size nn is finally generated, the Pareto disturbance set is determined by comparing the dominance relationship between the disturbance individuals of each offspring in the population. Due to the diversity of the population, the final Pareto set includes both individuals with strong attack capabilities but high disturbance metrics and individuals with weak attack capabilities but low disturbance metrics. In this embodiment, the perturbation screening strategy design process is to retain only the adversarial perturbations that can successfully attack the samples in the Pareto set, and arrange them in ascending order according to the metrics of these perturbations. Among them, the perturbation with the smallest perturbation metric will be selected as the adversarial perturbation that is ultimately used to modify the original sample. In addition, the perturbation screening strategy can also be adjusted according to the specific requirements of the attacker, such as ensuring that the attack is successful while making the predicted probability of the specified target category reach 90%.
[0040] The best adversarial perturbation individuals are finally selected to form adversarial samples. This adversarial sample may be almost identical to the original sample in vision or other fields, but can cause the target model to misclassify, thereby achieving the attacker's purpose.
[0041] The generation of adversarial samples against perturbed individuals can be formulated as a constrained optimization problem. This example assumes that the input image can be represented by a vector, where each scalar element represents a pixel. Let f be the image classifier of the target model that receives n-dimensional input, x = (x1,…,x n ) is the original natural image correctly classified as class t, where x1,…,x n are the pixels in the image respectively. Therefore, the probability that x belongs to class t is f t (x). The vector e(x)=(e1,…,e n ) is an additional adversarial perturbation, the target category is adv, the maximum change range is L, where e1,…,e n is the perturbation added to each pixel. Note that L is generally measured by its length. So the goal of the targeted adversarial attack is to find an optimal solution e(x) * :
[0042] maximize e(x)* f adv (x+e(x))
[0043] subject to∥e(x)∥≤L
[0044] Where e(x)* is the optimal perturbation, subject to represents the constraint condition, and f adv () represents the function of generating adversarial samples, and maximize indicates that the problem is a maximum minimization problem.
[0045] The problem involves finding two values: which pixels need to be disturbed; and the modification range of each pixel. In this embodiment, there are some changes to the above formula:
[0046] maximize e(x)* f adv (x+e(x))
[0047] subject to∥e(x)∥0≤d,∥e(x)∥≤L
[0048] Where d is a very small number. In this embodiment, only the restricted dimensions are modified, while the other dimensions of e(x) are kept as zero. The modification of one pixel can be regarded as perturbing the data point in the direction parallel to the axis of one of the n dimensions. Similarly, limited multi-pixel modifications move the data points into a multidimensional dataset. Overall, attacks with a few pixels produce perturbations on a low-dimensional slice of the input space. In fact, the perturbation of one pixel allows the image to be modified with arbitrary intensity in a selected direction among the n possible directions. At the same time, in order to ensure the imperceptibility of the perturbation, while limiting the number of perturbed pixels, the intensity of the perturbation should also be limited, that is, both focusing on a few pixels and limiting the intensity of the modification.
[0049] In this embodiment, the multi-objective optimization problem in the adversarial perturbation generation module includes three problems that need to be optimized, all of which can be defined as minimization problems, namely: minimizing the predicted probability sub-problem of the true class label of the original image, minimizing the perturbation metric value sub-problem of the adversarial perturbation, and minimizing the number of perturbed pixels sub-problem. Assuming that there are m optimization objectives, according to the multi-objective optimization definition, the final optimization objective can be determined as:
[0050] minimizeF(x)=(f1(x),…,f m (x) T subject to x∈Ω
[0051] Among them, F(x) represents the overall objective function to be optimized, f1(x),…,f m(x) represents each single objective function, subject to represents the constraint condition, and Ω represents the domain of definition
[0052] For the sub-problem of minimizing the predicted probability of the true class label of the original image, this embodiment uses the predicted probability result output by the target model as the optimization index; for the sub-problem of minimizing the perturbation measure value of the adversarial perturbation, this embodiment uses the L2 norm as the perturbation measure index; for the sub-problem of minimizing the number of perturbed pixels, this embodiment uses the number of pixels with added perturbation as the optimization index.
[0053] For multi-objective optimization problems, this embodiment uses the multi-optimization objective algorithm MOEA / D to solve them. The MOEA / D algorithm is based on the idea of decomposition optimization, converting the multi-objective optimization problem into the optimization of a set of single-objective sub-problems, and then using an evolutionary algorithm to solve these sub-problems. The core idea is to decompose the multi-objective problem into a set of independent single-objective sub-problems, each sub-problem only focuses on one or a few objectives in the objective function, and then coordinate the solutions of these sub-problems to obtain the global optimal Pareto optimal solution set. The solution process is as follows:
[0054] An original image is selected, serialized and subjected to multiple random perturbations to generate a set of initial solutions as the initial set of the population.
[0055] The multi-objective optimization problem is decomposed into three single-objective sub-problems: minimizing the predicted probability of the true class label of the original image, minimizing the perturbation metric of the adversarial perturbation, and minimizing the number of perturbed pixels. Each sub-problem corresponds to a weight vector, which is used to map the target vector (multiple objective function values) to the space of a single target value.
[0056] An evolutionary algorithm is applied to each subproblem for optimization. The optimization process of each subproblem is performed independently, using only the objective function value and weight vector related to the subproblem. For the subproblem of minimizing the predicted probability of the true class label of the original image, the predicted probability result of the output layer of the target model is used as the optimization indicator. For the subproblem of minimizing the perturbation metric value of the adversarial perturbation, the L2 norm is used as the perturbation metric. For the subproblem of minimizing the number of perturbed pixels, the number of pixels with added perturbation is used as the optimization indicator.
[0057] The differential evolution algorithm is used to optimize the solution space of each subproblem and generate a set of Pareto optimal solutions for the subproblems.
[0058] For each subproblem, the solution to the current subproblem is updated by considering the solutions to its neighbor subproblems. The neighbor solutions are usually a set of subproblems selected according to some topological structure or randomly.
[0059] By comparing the performance of the solution to the current subproblem with that of its neighbors, the solution set of the current subproblem is updated.
[0060] Subproblem optimization and neighbor solution update are performed repeatedly and iteratively until the predetermined stopping condition is reached (e.g., reaching the maximum number of iterations, converging to a stable Pareto frontier, etc.). The final Pareto set contains both adversarial sample individuals with strong attack capabilities but high perturbation metrics and adversarial sample individuals with weak attack capabilities but low perturbation metrics, thereby screening out suitable adversarial samples.
[0061] (3) The adversarial sample obtained by the multi-objective optimized adversarial perturbation generation module is sent to the target model module to attack the target model module. The target model module processes the input adversarial sample and generates a prediction probability result, and the target model module feeds the prediction probability result back to the multi-objective optimized adversarial perturbation generation module.
[0062] like Figure 2 As shown, the target model module in this embodiment (named CTPNet in this embodiment) adopts the idea of combining CNN and ViT networks, so that after the adversarial sample is input into the target model module, an adversarial sample that can be migrated between the two architecture network models can be obtained. Although the methods based on convolutional neural networks have excellent performance, due to the internal limitations of convolution operations (there are still limitations in modeling and displaying long-distance relationships), these architectures usually produce weaker performance. In order to overcome this limitation, in the target model module of this embodiment, a self-attention mechanism is established based on CNN features.
[0063] On the other hand, the Transformer designed for sequence-to-sequence prediction has emerged as an alternative architecture that fully adopts distributed convolution operations and relies only on attention mechanisms. Unlike previous CNN-based methods, the Transformer is not only powerful in modeling global context, but also shows excellent transferability to downstream tasks under large-scale pre-training. However, since the Transformer treats the input as a one-dimensional sequence and only focuses on modeling global context at all stages, low-resolution features that lack detailed positioning information are generated. Moreover, this information cannot be effectively restored by directly upsampling to full resolution, resulting in rough segmentation results. Therefore, in the target model module of this embodiment, the CNN architecture provides a way to extract low-level visual cues, which can well compensate for such fine spatial details.
[0064] In summary, the target model module CTPNet of this embodiment intends to build a CNN-Transformer hybrid structure similar to TransUNet. CNN is used as an encoder to extract local features of input data. It contains convolutional layers, activation functions, pooling layers and fully connected layers. Transformer is used as a decoder to utilize these features and generate outputs. Transformer Layer consists of multi-head self-attention and multi-layer perceptron blocks. These layers process input sequences and capture global context information. The decoder can use the self-attention mechanism of Transformer to model long-distance dependencies and generate task-related outputs. A patch-wise data enhancement strategy for ViTs is added, which applies different enhancements to each input image patch during training to make up for the diversity of enhancements between patches. The target model module is used as the source model for adversarial attacks. After the adversarial sample is input, it is first subjected to local feature extraction by CNN. The extracted feature map is then passed through the Transformer Layer to establish the global context information in the feature representation. Finally, the classifier prediction result is obtained through the fully connected layer, providing the classification probability query result for the evolution process, affecting the evolution direction of the adversarial sample, and then evolving adversarial samples that adapt to the two architectures.
[0065] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are merely descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. The various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction, and such combinations should also be regarded as the contents disclosed in the present disclosure as long as they do not violate the concept of the present invention. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.
[0066] The present invention is not limited to the specific details of the above-mentioned embodiments. Within the technical concept of the present invention and without departing from the design concept of the present invention, various modifications and improvements made to the technical solution of the present invention by technical personnel in this field should fall within the protection scope of the present invention. The technical contents for which protection is sought in the present invention have been fully recorded in the claims.
Claims
1. A new media content privacy protection system based on multi-objective optimization adversarial sample generation, characterized by: It includes an input module, a multi-objective optimized adversarial disturbance generation module, and a target model module, among which: The input module converts the original image samples into serialized data after preprocessing, and generates an initial perturbation population based on the serialized data and sends it to the multi-objective optimized adversarial perturbation generation module; In the multi-objective optimized adversarial perturbation generation module, the initial perturbation population is randomly added with perturbations for initialization; then, individuals in the initial perturbation population are randomly selected as parent individuals, and the parent individuals are evolved to generate offspring perturbation individuals; then, the generated offspring perturbation individuals are subjected to adaptability evaluation, and the problem is decomposed into three sub-problems during the adaptability evaluation, namely, minimizing the prediction probability of the true class label of the original image, minimizing the perturbation metric value of the adversarial perturbation, and minimizing the number of perturbation pixels. Thus, offspring perturbation individuals that meet the requirements of deceptiveness and visual quality are obtained through adaptability evaluation to constitute adversarial samples, and the adversarial samples are sent to the target model module to attack the target model module; The target model module processes the input adversarial samples and generates a predicted probability result, and the target model module feeds the predicted probability result back to the multi-objective optimized adversarial perturbation generation module.
2. The new media content privacy protection system based on multi-objective optimization adversarial sample generation according to claim 1 is characterized in that: The preprocessing of the input module includes adjusting image size and resolution, image denoising or smoothing, image enhancement, and normalizing or standardizing image pixel values.
3. The new media content privacy protection system based on multi-objective optimization adversarial sample generation according to claim 1 is characterized in that: When performing adaptability evaluation in the adversarial perturbation generation module of the multi-objective optimization, for the sub-problem of minimizing the predicted probability of the true class label of the original image, the predicted probability result output by the target model is used as the optimization index; for the sub-problem of minimizing the perturbation metric value of the adversarial perturbation, the L2 norm is used as the perturbation metric index; for the sub-problem of minimizing the number of perturbed pixels, the number of pixels with added perturbation is used as the optimization index.
4. The new media content privacy protection system based on multi-objective optimization adversarial sample generation according to claim 1 is characterized in that: The target model module is composed of a CNN network and a ViT network.
