A social trust-driven method for protecting facial information

Abstract

This invention relates to a social trust-driven method for protecting facial information, belonging to the field of information security technology, and includes the following steps: S1: The system initializes three entities, including a data publisher, a trusted third party, and a data visitor; S2: The trusted third party uses a graph attention mechanism to calculate the real-time social trust value between the data publisher and the data visitor; S3: Based on the social trust value of the data visitor, a privacy level is assigned, and a protected identity vector is generated using a differential privacy mechanism, wherein the privacy level is used to control the size of the privacy budget; S4: The protected identity vector is input into a conditional identity anonymization generative adversarial network to obtain a facial information protected image, which is then sent to the corresponding data visitor.

Description

Technical Field

[0001] This invention belongs to the field of information security technology and relates to a social trust-driven method for protecting facial information. Background Technology

[0002] With the rapid development of internet technology, an increasing amount of multimedia data (images and videos) is being disseminated and shared on social media platforms. However, when sharing image data, social network users often overlook the facial information contained within the massive amounts of data, leading to security risks. Currently, facial biometric technology based on deep learning algorithms can easily identify and steal facial feature information from images, modifying it into 3D masks to deceive facial recognition systems in real life, gaining unauthorized access or failing identity verification. In today's world where soft biometric features such as faces are becoming a new form of password, this seriously threatens users' personal information security.

[0003] Current research on facial information security on social networks has made some progress. Existing studies often employ image editing, image encryption, and deep learning-based face de-identification techniques to protect facial information. More and more researchers hope to ensure facial information security while maintaining the naturalness of de-identified images. For example, Li et al. proposed an SF-GAN method based on generative adversarial networks in "SF-GAN: face de-identification method without losing facial attribute information" [IEEE Signal Processing Letters, 2021, 28:1345-1349.]. This method uses shallow and deep facial attribute information to process non-critical information such as facial expressions and backgrounds differently from critical information related to identity attributes. This further enhances the protection of facial information while generating high-fidelity de-identified images. The face de-identification results achieved by generative adversarial networks are encouraging, but simply replacing or modifying all facial information in an image may reduce its utility, thus affecting users' willingness to share images. Users often want to minimize information security risks while preserving the original image quality as much as possible, allowing trusted users to see more realistic images. Meanwhile, the online social network image sharing scenario is inherently complex. The social attributes, relationships, and trust among users are constantly changing, making it cumbersome for users to manually set privacy policies each time. Some users may be unwilling to invest the extra time and effort to filter who can view their shared images. To reduce the time cost of manually setting privacy policies and improve the efficiency of user image sharing, much research is currently focusing on the automatic setting of privacy policies. For example, Jiao et al. proposed an IEye model for automatically detecting image privacy based on common sense and personal knowledge in "Ieye: Personalized image privacy detection" [2020 6th International Conference on Big Data Computing and Communications (BIGCOM). IEEE, 2020: 91-95]. Specifically, for each user's image, a multi-layer semantic graph is constructed as a feature representation of the image, describing its personalized privacy policy and minimizing image utility loss, but this approach has certain limitations in terms of efficiency.

[0004] Despite the extensive research and application of existing methods for protecting facial information, the following major challenges remain:

[0005] (1) Adaptive privacy protection strategy: It is time-consuming and costly for users to manually set privacy protection strategies, while existing adaptive privacy strategies cannot take into account the frequently changing factors in social networks and are not suitable for dynamic social network scenarios.

[0006] (2) Fine-grained facial information protection: Most existing facial information protection methods completely replace or modify the faces in the original image to achieve the purpose of facial information protection. They do not take into account the user's need for information sharing based on trust, resulting in a certain degree of loss of image utility and failing to achieve fine-grained image sharing. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a social trust-driven facial information protection method. Combining social trust assessment and generative adversarial network (GAN) technology, it provides a fine-grained facial information protection scheme based on user social trust values, addressing the problems faced by existing technologies. It leverages the powerful computing capabilities of trusted third parties to achieve dynamic data sharing among users, while balancing the information security and image utility of shared images. Specifically, this method achieves fine-grained image access and sharing by setting different privacy levels for visitors, thereby reducing the threat caused by facial information leakage. First, a social network trust graph is constructed, and a graph attention mechanism is introduced to calculate the real-time social trust between users, serving as the basis for classifying the privacy levels of visitors. Furthermore, to balance the security protection of user facial information with high image fidelity, this invention designs a facial protection algorithm based on a generative adversarial model. This algorithm adjusts the facial features of the face portion of the original image to different degrees according to the visitor's privacy level to achieve different levels of facial information protection.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A social trust-driven method for protecting facial information includes the following steps:

[0010] S1: The system initializes three entities, including the data publisher, the trusted third party, and the data accessor;

[0011] S2: A trusted third party uses a graph attention mechanism to calculate the real-time social trust value between data publishers and data visitors;

[0012] S3: Assign a privacy level to the data visitor based on their social trust value, and generate a protected identity vector using a differential privacy mechanism. The privacy level is used to control the size of the privacy budget.

[0013] S4: Input the protected identity vector into the conditional identity anonymization generative adversarial network to obtain a face information protected image, and send it to the corresponding data accesser.

[0014] Furthermore, the system initialization in step S1 includes:

[0015] Initialize three entities: the data publisher, the trusted third party, and the data accessor.

[0016] The data publisher is the user who generates the image and shares it on the social network;

[0017] The trusted third party is a relay for image uploading on social networks. It constructs a social trust graph based on the basic attributes of data publishers and data visitors, and is responsible for assessing users' social trust and processing images.

[0018] The data accessor is used to send data requests to a trusted third party in order to obtain data published by the data publisher.

[0019] Initialize the user set of online social networks in the social trust graph. Set the set of all user social interactions within the time period [0, T] as follows: in, Let G represent the set of neighboring nodes that have interacted with node i within the time interval [0,T]. Social trust is modeled as a series of timestamped interaction events G = {e(t1), e(t2), ...} in the social trust graph, where e(t1), e(t2), ..., e(t3) = ..., e(t4) = ..., e(t5) = ..., e(t6) = ..., e(t7) = ..., e(t8) = ..., e(t9) m ) represents t m Social interaction between any pair of user nodes at any given time; user nodes i and j at time t m The social trust with a value of c is represented as a directed edge.

[0020] Furthermore, step S2 specifically includes the following steps:

[0021] S21: A trusted third party generates a social trust graph G(T) at a given time T, where the nodes in the graph are social network users and the edges are the social trust values ​​c connecting pairs of users;

[0022] S22: Given a pair of social user nodes i, j, first extract the user's basic attributes s. i User i and its neighboring nodes k∈N i (T) historical interactions and related timestamps Ψ t ,according to Generate joint embedding vector n ik To quantify the message signals from user i to neighboring user k;

[0023] S23: Weighted calculation and aggregation of neighboring users k∈N i The joint embedding vector n of (T) ikDuring this process, a graph attention mechanism is executed to assign higher weights to the most recent interactions, and the weight assigned to each user for each neighboring user k is denoted as a. ik Finally, the first-order feature representation of user i is obtained.

[0024] S25: By stacking L trust propagation layers, the influence of multi-hop user nodes on user i is aggregated to obtain a higher-order feature representation.

[0025] S26: Calculate the higher-order feature representation of user node pair i,j from step S25. and The result is fed into an MLP, processed through a softmax layer, and outputs the probability h of predicting social trust between any user i and j. ij Ultimately, through c = argmax(h) ij The social trust c between any two users is calculated.

[0026] Furthermore, in step S22, a mapping function is used. Get timestamp Ψ t The historical interaction trust value between user i and its neighboring nodes will be used. Convert to dense vector embedding; then embed the user's attributes into vectors s i Timestamp embedding vector Ψ t and interactive embedding vectors according to Generate a joint embedding vector as input to the trust propagation layer model.

[0027] Furthermore, in step S23, a graph attention mechanism is used to obtain the interaction weights a assigned to user i by user k. ik ←attn(n ik W ik ,p i W i ) = n ik W ik ·(p i W i ) T ,in, and For a trainable matrix, d s d t d c For the dimension of the embedding vector, then use The normalized weights are obtained; the joint embedding vector n of neighboring users is calculated and aggregated through weighted summation. ik To obtain the first-order feature representation of user i Where W and b are the parameters that need to be trained, k∈N i(T) is the set of neighboring nodes of user i.

[0028] Furthermore, in step S26, the user's higher-order feature representations of i and j are... and The data is passed to the MLP, which outputs the social trust prediction probability h. ij :

[0029]

[0030] Among them, W f b f This is the trainable weight matrix defined in the multilayer perceptron;

[0031] The objective function of the entire model is defined as the cross-entropy loss between the predicted values ​​and the true values ​​of trust in graph G:

[0032]

[0033] in This represents all trainable model parameters.

[0034] Furthermore, step S3 specifically includes the following steps:

[0035] S31: Assign trust level d to the accessing user based on the social trust value c calculated in step S2;

[0036] S32: Extract features from the original image and integrate them to obtain the identity vector I. id ;

[0037] S33: Use differential privacy mechanism based on the visitor's trust level d. Obtain the protected identity vector I i ' d ,in, This indicates a mean of 0 and a scale of 0. The random noise is a Laplace distribution; Δ is a parameter used in differential privacy to control the strength of privacy protection and is related to the sensitivity of the data; d is the visitor's trust level, which determines I id The degree of disturbance, the smaller the d value, the greater the added noise, I i ' d with I id The larger the difference, the lower the noise; conversely, the larger the d value, the lower the I. i ' d with I id The smaller the gap.

[0038] Furthermore, step S4 specifically includes the following steps:

[0039] The perturbation identity vector obtained in step S33, the non-facial region image of the original image, and the target identity image are used as inputs to the generator in the generative adversarial network.

[0040] The generative adversarial network uses an encoder-decoder architecture. The encoder combines the inputs in step S41 into a low-dimensional vector, and the decoder combines the vectors to generate a face information protection image through upsampling.

[0041] The loss function is set as follows:

[0042] L total =L G +L D +L Rec

[0043] Where L G L D Let L be the loss function for the generator and discriminator. Rec Image reconstruction loss;

[0044]

[0045] Where a represents the label of the generated data, b represents the label of the real data; D(x) is the output of the discriminator D on the real data x; (D(x)-b) 2 x represents the squared error between the discriminator's output of the real data and the real label; p data (x) represents the distribution of the true data p. data The data obtained by sampling; D(G(z)) represents the output of the discriminator D on the data G(z) generated by the generator G; (D(G(z))-a) 2 The discriminator's squared error between the output of the generated data and the generated labels represents the squared error from the prior distribution z ~ p. z The noise z sampled from (z);

[0046]

[0047] L Rec =||x raw -x adv ||.

[0048] The beneficial effects of this invention are as follows: This invention relies on the trust relationship between users of social networks, thereby effectively addressing the potential security risks of facial information in image sharing on social networks, while ensuring the utility of shared data.

[0049] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0051] Figure 1 This is a system architecture diagram as described in this invention;

[0052] Figure 2 This is a flowchart illustrating the workflow of the present invention. Detailed Implementation

[0053] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0054] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0055] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0056] Please refer to the present invention. Figures 1-2This invention discloses a social trust-driven method for protecting facial information. It combines real-time social trust assessment and generative adversarial network (GAN) technology, aiming to address the challenges faced by existing technologies and provide a flexible solution to potential facial information security issues in image sharing on social networks. To achieve the above objectives, the invention provides the following technical solution:

[0057] System S1 Initialization: The system initializes three entities: data publishers, trusted third parties, and data visitors. Data publishers are users who generate images and share them on the social network. Before sharing the data, they upload the images to a trusted third party for processing. The trusted third party extracts social interaction information between data publishers and visitors to construct the social network graph G. Graph G initializes the user set of the online social network. Set the set of all user social interactions within the time period [0, T] as follows: in, Let G represent the set of neighboring nodes that have interacted with node i within the time interval [0, T]. Furthermore, social trust is modeled as a series of timestamped interaction events G = {e(t1), e(t2), ...} in the social graph, where e(t1), e(t2), ..., e(t3) = ..., e(t4) = ..., e(t5) = ..., e(t6) = ..., e(t7) = ..., e(t8) = ..., e(t9) = ..., e(t1) = ..., e(t1) = ..., e(t2) = ..., e(t9 ... m ) represents t m Social interaction between any pair of user nodes at any given time; user nodes i and j at time t m The social trust with a value of c is represented as a directed edge.

[0058] S2 Social Network User Dynamic Trust Assessment: This part mainly consists of three modules: the embedding layer, the trust propagation layer, and the trust prediction layer. In the embedding layer, firstly, the basic information and social attributes of user i are embedded into the attribute embedding vector s. i To describe it. Then, use the mapping function. Obtain the timestamp embedding vector Ψ t Finally, the historical interaction trust values ​​between user i and its neighboring nodes will be used. Convert to dense vector embedding. Further, embed the user's attributes into vectors s. i Timestamp embedding vector Ψ t and interactive embedding vectors according to Generate a joint embedding vector as input to the trust propagation layer model.

[0059] In the single-layer trust propagation layer, a graph attention mechanism is used to obtain the interaction weight 'a' assigned to user i by user k. ik ←attn(n ik W ik ,p i W i ) = n ik Wik ·(p i W i ) T .in, and For a trainable matrix, d s d t d c Let this be the dimension of the embedding vector. Then use... Obtain normalized weights The joint embedding vector n of neighboring users is calculated and aggregated through weighted calculation. ik To obtain the first-order feature representation of user i Where W and b are the parameters that need to be trained, k∈N i (T) is the set of neighboring nodes of user i.

[0060] By stacking L trust propagation layers, the influence of multi-hop user nodes on user i is aggregated to obtain a higher-order feature representation.

[0061]

[0062] in, These are trainable parameters. The higher-order feature representations of user pairs i and j are used. and The data is passed to the MLP, which outputs the social trust prediction probability h. ij :

[0063]

[0064] Among them, W f b f This is the trainable weight matrix defined in the multilayer perceptron. Finally, the social trust values ​​of users i and j are calculated as c = argmax. i (h ij ).

[0065] The objective function of the entire model is defined as the cross-entropy loss between the predicted values ​​and the true values ​​of trust in graph G:

[0066]

[0067] in This represents all trainable model parameters.

[0068] S3: Trust-driven facial feature point perturbation:

[0069] The trusted third party uses the social trust c calculated in step S3 to determine the visitor's trust level d through a simple linear mapping function f(c) → d. This linear mapping function f() can be a function that maps trust values ​​to a predetermined range of levels, for example, mapping trust values ​​to between 0 and 1, and then dividing different trust levels according to this range.

[0070] Next, a deep learning model is used to extract features from the original image to obtain the identity vector I of the face information. id Then, based on the visitor's trust level d, a differential privacy mechanism is used to compute the protected identity vector I. i ' d ,Right now in, This indicates a mean of 0 and a scale of 0. Random noise with a Laplace distribution. Here, Δ is a parameter used in differential privacy to control the strength of privacy protection and is related to the sensitivity of the data; while d is the visitor's trust level, which determines I. id The degree of disturbance. The smaller the d value, the greater the added noise, I i ' d with I id The larger the difference, the lower the noise; conversely, the larger the d value, the lower the I. i ' d with I id The smaller the gap.

[0071] S4: Face de-identification technique based on conditional identity anonymization generative adversarial networks:

[0072] S41: Generator input: The perturbation identity vector obtained in step S33, the non-face region image of the original image, and the target identity image are used as the generator input;

[0073] (1) Perturbation identity vector: Use the perturbation identity vector obtained by S33 as the generator input;

[0074] (2) Non-facial region image: The goal of this method is to protect information only in the facial region and not to modify the non-facial regions of the original image. Therefore, this method extracts the non-facial region from the original image and uses it as input to the generator to ensure that the non-facial region of the generated face information-protected image is as close as possible to the original image.

[0075] (3) Target Identity: To prevent the generator from overfitting and causing the generated images and original images to share the same identity information, this method introduces a set of target images with different identities to guide the generator's image generation. The target identity information is represented as a one-hot vector through an MLP, which is then used as the input to a transposed convolutional neural network and fed into the generator.

[0076] S42: Generative Adversarial Network (GAN) Module: The GAN uses an encoder-decoder architecture. The encoder combines the inputs from step S41 into a low-dimensional vector, and the decoder combines these vectors to generate an image for recognition through upsampling. The target loss of the discriminator is defined as follows:

[0077]

[0078] Where a represents the label of the generated data, b represents the label of the real data; D(x) is the output of the discriminator D on the real data x; (D(x)-b) 2 x represents the squared error between the discriminator's output of the real data and the real label; p data (x) represents the distribution of the true data p. data The data obtained by sampling; D(G(z)) represents the output of the discriminator D on the data G(z) generated by the generator G; (D(G(z))-a) 2 The discriminator's squared error between the output of the generated data and the generated labels represents the squared error from the prior distribution z ~ p. z The noise z sampled from (z) is used as the generator loss definition.

[0079]

[0080] To ensure the similarity between the generated image and the original image, the mean absolute error between the generated image and the original image is used to construct the reconstruction loss L. Rec This is used to preserve more details of the original image.

[0081] L Rec =‖x raw -x adv ||

[0082] The model loss function is set as least squares loss and reconstruction loss L. Rec The sum. The total loss function can be expressed as:

[0083] L total =L G +L D +L Rec

[0084] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily refer to the same embodiment.

[0085] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0086] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods in this embodiment.

[0087] This embodiment also provides an electronic terminal, including: a processor and a memory;

[0088] The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to cause the terminal to perform any of the methods in this embodiment.

[0089] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0090] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0091] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0092] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0093] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0094] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A social trust-driven method for protecting facial information, characterized in that: Including the following step: S1: The system initializes three entities, including the data publisher, the trusted third party, and the data accessor; the system initialization includes: Initialize three entities: the data publisher, the trusted third party, and the data accessor. The data publisher is the user who generates the image and shares it on the social network; The trusted third party is a relay for image uploading on social networks. It constructs a social trust graph based on the basic attributes of data publishers and data visitors, and is responsible for assessing users' social trust and processing images. The data accessor is used to send data requests to a trusted third party in order to obtain data published by the data publisher. Initialize the user set of online social networks in the social trust graph. Set time period The collection of all user social interactions within the system is ,in, Indicates time Inner nodes The set of neighboring nodes that have interacted; modeling social trust as a series of timestamped interaction events in the social trust graph. ,in, express Social interaction between any pair of user nodes at any time; user nodes and In time The value formed Social trust is represented as directed edge ; S2: A trusted third party uses a graph attention mechanism to calculate the real-time social trust value between data publishers and data users; this includes the following steps: S21: A trusted third party generates a given time. A social trust graph In this graph, the nodes represent social network users, and the edges represent the social trust values ​​connecting user pairs. ; S22: Given a pair of social user nodes , First, extract the user's basic attributes. ,user With neighboring nodes Historical Interaction and related timestamps ,according to Generate joint embedding vectors To quantify users To neighboring users Message signal; S23: Weighted calculation and aggregation of adjacent users joint embedding vector During this process, a graph attention mechanism is implemented to assign higher weights to the most recent interactions, allowing each user to interact with each of their neighboring users. The assigned weights are represented as Ultimately, users First-order feature representation ; S25: By stacking A trust propagation layer aggregates multi-hop user nodes for user... The influence of obtaining higher-order feature representations ; S26: The user node pair is calculated from step S25. , Higher-order feature representation and The input is then fed into an MLP, processed through a softmax layer, and outputs any user's input. , Social trust prediction probability between Finally passed Calculate the social trust between any two users. ; S3: Assign a privacy level to the data visitor based on their social trust value, and generate a protected identity vector using a differential privacy mechanism. The privacy level controls the privacy budget. This includes the following steps: S31: Based on the social trust value calculated in step S2 Grant trust levels to users who access the site ; S32: Extract features from the original image and integrate them to obtain an identity vector. ; S33: Based on the visitor's trust level Using differential privacy mechanism Obtain the protected identity vector ,in, This indicates a mean of 0 and a scale of 0. The random noise is distributed by the Laplace distribution, and Δ is a parameter used in differential privacy to control the strength of privacy protection, which is related to the sensitivity of the data; It is the visitor's trust level, which determines... The degree of disturbance, The smaller the value, the greater the added noise. and The greater the gap, the better; conversely, The larger the value, the lower the noise. and The smaller the gap; S4: Input the protected identity vector into the conditional identity anonymization generative adversarial network to obtain a face-protected image, and send it to the corresponding data accessor; specifically including the following steps: The perturbation identity vector obtained in step S33, the non-facial region image of the original image, and the target identity image are used as inputs to the generator in the generative adversarial network. The generative adversarial network uses an encoder-decoder architecture. The encoder combines the inputs in step S41 into a low-dimensional vector, and the decoder combines the vectors to generate a face information protection image through upsampling. The loss function is set as follows: in , Let the loss function be the generator and the discriminator. Image reconstruction loss; in, Labels representing the generated data, Labels representing real data; For the discriminator For real data The output; This represents the squared error between the discriminator's output of the real data and the real label; Indicates the distribution of real data Data obtained from sampling; Representative Discriminator For generator The generated data The output; The discriminator's output on the generated data is the squared error between the generated labels and the output of the discriminator; representing the error from the prior distribution. noise obtained from mid-sample ; 。 2. The social trust-driven facial information protection method according to claim 1, characterized in that: In step S22, a mapping function is used. Get timestamp ; will work with users Trust value of historical interactions with neighboring nodes Convert to dense vector embedding; then embed the user's attributes into the vector. Timestamp embedding vector and interactive embedding vectors according to Generate a joint embedding vector as input to the trust propagation layer model.

3. The social trust-driven facial information protection method according to claim 1, characterized in that: In step S23, a graph attention mechanism is used to acquire user information. For users Assigned interaction weights ,in, and For a trainable matrix, , , For the dimension of the embedding vector, then use Normalized weights are obtained; joint embedding vectors of neighboring users are calculated and aggregated through weighted summation. to gain users First-order feature representation ,in, , These are the parameters that need to be trained. For users The set of neighboring nodes.

4. The social trust-driven facial information protection method according to claim 1, characterized in that: In step S26, the user's... , Higher-order feature representation and The data is passed to an MLP (Multi-Level Processing) to output the social trust prediction probability. : in, ,for The trainable weight matrix defined in a multilayer perceptron; The objective function of the entire model is defined as the sum of the predicted values ​​and the graph. Cross-entropy loss between trust truth values: in This represents all trainable model parameters.

Images