A fine-grained privacy information propagation control policy generation method and device

Abstract

The application discloses a fine-grained privacy information propagation control strategy generation method and device, and belongs to the technical field of privacy control, comprising the following steps: obtaining an intimacy heterogeneous graph according to a social graph and a diffusion graph between users; updating the node embedding representation of the user node according to the graph neural network and the intimacy heterogeneous graph to obtain the updated node embedding representation; obtaining an initial user activation probability matrix according to the updated node embedding representation; obtaining a privacy leakage score and a propagation scale according to the initial user activation probability matrix; constructing an optimization problem according to the privacy leakage score and the propagation scale; solving the optimization problem to obtain an optimal solution set, and taking the user in the optimal solution set as the node for next-hop forwarding. The application has higher accuracy in information diffusion prediction and is closer to actual data, and solves the static and rigid problems of the prior art, thereby providing a privacy control scheme with higher quality and more comprehensive selection for users, and realizing a real utility optimal balance.

Description

Technical Field

[0001] This invention belongs to the field of privacy control technology, specifically relating to a method and apparatus for generating fine-grained privacy information dissemination control strategies. Background Technology

[0002] With the rapid development of internet technology and the deep penetration of social media, social networks have become core platforms for information dissemination and interpersonal communication. While users enjoy the convenience, social recognition, and commercial benefits brought by information sharing, they also face unprecedented risks of privacy breaches. A fundamental challenge is that once users publish information, they often lose control over its subsequent multi-hop propagation path. This uncontrolled propagation can lead to a "cascading diffusion" of users' personal data, sensitive opinions, or private life details on the network, ultimately reaching unauthorized individuals or even entities explicitly marked by users as "blacklisted" from whom they should not receive information. Therefore, how to strike a delicate balance between the benefits of information dissemination and the risks of privacy protection has become a key technical challenge in the field of privacy computing.

[0003] Against this backdrop, academia and industry have proposed various privacy control schemes. Existing privacy control strategies lack flexibility, specifically in their simplistic control methods. For example, social media platforms commonly use a "public / friends visible / private" policy, which fails to finely control who sees different content. Furthermore, they neglect the complexity of dissemination processes and the possibility that users can indirectly access sensitive information through social media networks. They also fail to dynamically adjust to users' privacy needs in different scenarios, making it difficult to provide a comprehensive control strategy.

[0004] Besides their lack of flexibility, existing privacy diffusion control strategies often struggle to balance information sharing and privacy protection because the goals of maximizing dissemination and minimizing privacy are in conflict and difficult to optimize simultaneously. Previous studies have overemphasized the scale of dissemination while neglecting the ethical and regulatory requirements for user privacy protection. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention provides a method and apparatus for generating fine-grained privacy information dissemination control strategies.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides a method for generating a fine-grained privacy information dissemination control strategy, the method comprising:

[0008] A heterogeneous graph of intimacy is obtained based on the social graph and diffusion graph between users; wherein, the social graph is obtained based on the social relationships between users; the diffusion graph is obtained based on the information diffusion behavior between users within a certain time interval; user nodes in the social graph represent users, and edges represent the social relationships between users;

[0009] The node embedding representation of the user node is updated based on the graph neural network and the proximity heterogeneous graph to obtain the updated node embedding representation;

[0010] The initial user activation probability matrix is ​​obtained based on the updated node embedding representation;

[0011] The privacy breach score and the scale of spread are obtained based on the initial user activation probability matrix.

[0012] An optimization problem is constructed based on the privacy breach score and the scale of the spread.

[0013] Solve the optimization problem to obtain the optimal solution set, and use the users in the optimal solution set as the nodes for next-hop forwarding.

[0014] Optionally, obtaining the initial user activation probability matrix based on the updated node embedding representation includes:

[0015] The user dynamic embedding is obtained based on the updated node embedding representation;

[0016] Based on the multi-head attention mechanism, user dependency embedding is obtained based on the user dynamic embedding;

[0017] The gating weights are calculated based on the user dynamic embedding and the user dependency embedding.

[0018] The user dynamic embedding and the user dependency embedding are fused according to the gating weight to obtain the final user embedding;

[0019] After embedding the end user into the fully connected layer and the activation function, the initial user activation probability matrix is ​​obtained.

[0020] Optionally, updating the node embedding representation of the user node based on the graph neural network and the proximity heterogeneous graph to obtain the updated node embedding representation includes:

[0021] Add self-loops to the adjacency matrix of the proximity heterogeneous graph and perform symmetric normalization to obtain a normalized adjacency matrix;

[0022] The node degree matrix is ​​obtained from the adjacency matrix of the aforementioned proximity heterogeneous graph;

[0023] The node embedding representation of the user node, the normalized adjacency matrix, the node degree matrix, and the weight matrix are multiplied together, and then processed by an activation function and layer normalization to obtain the updated node embedding representation.

[0024] Optionally, the calculation process of the updated node embedding representation includes:

[0025] Linear transformations are performed based on the node embedding representation, adjacency matrix, and node degree matrix of the previous layer. The data is then processed using learnable parameters and activation functions, and finally, layer normalization is performed to obtain the updated node embedding representation of the current layer.

[0026] Optionally, the computation process for the user-dynamic embedding includes:

[0027] The user dynamic embedding is obtained by calculating the node embedding representation and temporal embedding between users, combined with a scaling factor, and then normalizing the result.

[0028] Optionally, the calculation process of the initial user activation probability matrix includes:

[0029] The final user embedding is linearly transformed and then processed by an activation function to obtain the initial user activation probability matrix.

[0030] Optionally, the process of constructing the optimization problem includes:

[0031] The optimization problem is constructed based on the privacy leakage score and propagation scale of the solution set in the iteration.

[0032] Secondly, the present invention provides a fine-grained privacy information dissemination control strategy generation apparatus, the apparatus comprising:

[0033] The heterogeneous graph acquisition module is used to obtain a proximity heterogeneous graph based on the social graph and diffusion graph between users; wherein, the social graph is obtained based on the social relationships between users; the diffusion graph is obtained based on the information diffusion behavior between users within a certain time interval; user nodes in the social graph represent users, and edges represent the social relationships between users;

[0034] The node update module is used to update the node embedding representation of the user node according to the graph neural network and the proximity heterogeneous graph, so as to obtain the updated node embedding representation.

[0035] The probability matrix acquisition module is used to obtain the initial user activation probability matrix based on the updated node embedding representation;

[0036] The privacy parameter acquisition module is used to obtain a privacy leak score and propagation scale based on the initial user activation probability matrix.

[0037] An optimization problem construction module is used to construct optimization problems based on the privacy breach score and the propagation scale.

[0038] The solution set acquisition module is used to solve the optimization problem, obtain the optimal solution set, and use the users in the optimal solution set as the nodes for next-hop forwarding.

[0039] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0040] In the above technical solution, the present invention introduces a proximity heterogeneous graph, which can more realistically simulate the path of information propagation based on proximity in the real world. This significantly improves the accuracy of information diffusion prediction, thereby ensuring that the subsequently generated privacy policy is based on more reliable and realistic data, resulting in better policy execution. Furthermore, when solving optimization problems, it can adaptively adjust according to the scenario and personal preferences, solving the static and rigid problems of existing technologies. Moreover, the solution set of the present invention is closer to the theoretical optimal boundary, providing users with a higher quality and more comprehensive privacy control scheme, achieving a true optimal balance of utility.

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0042] Figure 1 This is a flowchart of a method for generating a fine-grained privacy information propagation control strategy provided in an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the structure of DHGCNUI, a privacy information diffusion prediction model based on intimacy.

[0044] Figure 3 This is a schematic diagram of the structure of an iGCN module;

[0045] Figure 4 This is a flowchart of a privacy diffusion multi-objective optimization algorithm PDO-NSGA-Ⅱ provided in an embodiment of the present invention;

[0046] Figure 5 This is a block diagram of a fine-grained privacy information dissemination control strategy generation device provided in an embodiment of the present invention. Detailed Implementation

[0047] To facilitate understanding of the present invention, a brief description of the prior art and the inventive concept of the present invention will be provided first.

[0048] The existing technology has the following problems:

[0049] (1) Existing technical solutions have coarse control granularity and cannot achieve flexible differentiated privacy control. They can only manage information through the binary logic of "allow / deny" or a single propagation blocking method. They cannot display information of different granularities to different users based on content semantics, user relationship closeness, scene context, etc., which makes it difficult to meet the differentiated privacy protection needs in complex social scenarios.

[0050] (2) Existing technical solutions generate static strategies that rely entirely on fixed rules preset by users. They do not have the ability to automatically generate or intelligently adjust privacy strategies based on dynamic changes in social networks (such as updates in user communication behavior and changes in user relationships), and cannot adapt to dynamic and ever-changing information dissemination scenarios.

[0051] (3) Existing technical solutions do not take into account the complexity of the dissemination behavior, ignore the possibility that users can indirectly access sensitive information through social dissemination chains, only carry out privacy control for direct dissemination paths, and do not prevent the risk of privacy leakage in multi-hop indirect dissemination scenarios, which means that sensitive information may still reach unauthorized users through indirect dissemination chains, resulting in privacy protection vulnerabilities.

[0052] (4) Existing technical solutions have a single optimization goal, focusing only on a single dimension of needs (such as only blocking privacy leaks or only ensuring information dissemination), without synergistically weighing the benefits of information dissemination against the risks of privacy leaks. Furthermore, traditional optimization methods can only output one "compromise" strategy at a time, failing to provide users with a diverse range of strategy options.

[0053] (5) The existing technical solutions have insufficient accuracy in predicting information diffusion. They do not fully consider key factors such as user intimacy and propagation behavior characteristics in social networks, resulting in a large deviation between the predicted user activation probability and the actual propagation situation. Consequently, the privacy control strategy generated based on this prediction has low reliability and is difficult to conform to the real information propagation pattern.

[0054] To address the problems of existing technologies, such as coarse control granularity, static strategies, single optimization objectives, low prediction accuracy, failure to consider the complexity of propagation behavior, and neglect of privacy risks in indirect propagation, this invention proposes a method and apparatus for generating fine-grained privacy information propagation control strategies.

[0055] Figure 1 This is a flowchart of a method for generating a fine-grained privacy information propagation control strategy provided in an embodiment of the present invention, such as... Figure 1 As shown, the method may include the following steps:

[0056] S101. Obtain a heterogeneous graph of intimacy based on the social graph and diffusion graph between users; wherein, the social graph is obtained based on the social relationships between users; the diffusion graph is obtained based on the information diffusion behavior between users within a certain time interval; user nodes in the social graph represent users, and edges represent the social relationships between users.

[0057] Understandable Figure 2 This is a schematic diagram of the structure of the DHGCNUI model, a privacy information diffusion prediction model based on intimacy. Figure 2 As shown, based on information diffusion prediction, we propose the DHGCNUI (Dynamic Heterogeneous Graph Convolutional Network with User Intimacy) model. This model combines intimacy and time-social dual-dimensional embeddings, and uses graph neural networks and attention mechanisms to obtain the initial user activation probability matrix. (Social graph) It represents the social relationships between users. Represents a set of user nodes. It represents the social relationships between users, if users and If there is an edge, then Diffusion diagram Indicates the time interval Information dissemination behavior among internal users Indicates the nodes that participate in the diffusion. Represents the edges generated by the diffusion behavior. Each time interval The diffusion maps may differ, therefore this invention introduces the concept of intimacy, which is defined up to the current time interval. The number of diffusion actions among all users is used as the weight of the edge relationship. This can be understood as the number of times diffusion behavior occurs between users, and this number is used as the edge weight in the diffusion graph, which is the intimacy level. Intimacy Heterogeneous Graph Combining social graphs and diffusion diagram Obtain a heterogeneous graph containing user social relationships and diffusion behavior. , It uses the user node set from the social graph, edges That is, within the time interval Internal merging of static social edges and dynamic diffusion edge .

[0058] S102. Update the node embedding representation of the user node according to the graph neural network and the proximity heterogeneous graph to obtain the updated node embedding representation.

[0059] Understandably, GCN (Graph Convolutional Network) is a type of graph neural network that updates the embedded representation of a node by aggregating the features of the node and its neighboring nodes and performing non-linear transformations. Figure 3 This is a structural diagram of an iGCN module, such as... Figure 3 As shown, iGCN (Intimacy Graph Convolutional Network) employs a two-layer GCN.

[0060] Optionally, S102 may include:

[0061] Add self-loops to the adjacency matrix of the proximity heterogeneous graph and perform symmetric normalization to obtain a normalized adjacency matrix;

[0062] The node degree matrix is ​​obtained from the adjacency matrix of the proximity heterogeneous graph;

[0063] The node embedding representation of the user node, the normalized adjacency matrix, the node degree matrix, and the weight matrix are multiplied together, and then processed by the activation function and layer normalization to obtain the updated node embedding representation.

[0064] Optionally, the computation process for the updated node embedding representation includes:

[0065] Linear transformations are performed based on the node embedding representation, adjacency matrix, and node degree matrix of the previous layer. The data is then processed using learnable parameters and activation functions, and finally, layer normalization is performed to obtain the updated node embedding representation of the current layer.

[0066] The updated node embedding representation is as follows:

[0067] ;

[0068] in, Indicates the first Layer node embedding representation The updated node embedding representation Representation layer normalization, This represents the LeakyReLU activation function. Represents the node degree matrix, This represents the normalized adjacency matrix. , It is the identity matrix. Indicates the number of users. It is a diffusion-intimacy heterogeneous graph The weighted adjacency matrix is ​​constructed from user nodes and two types of edge relationships. yes The degree matrix, , It means diagonal elements in User All adjacent nodes, The user embeddings are initialized using the Xavier normal distribution. It is the dimension of user embedding. Indicates the first Learnable parameters of the layer.

[0069] S103. Obtain the initial user activation probability matrix based on the updated node embedding representation.

[0070] Optionally, S103 may include:

[0071] The user dynamic embedding is obtained based on the updated node embedding representation;

[0072] Based on the multi-head attention mechanism, user dependency embedding is obtained based on user dynamic embedding;

[0073] The gating weights are calculated based on user dynamic embedding and user dependency embedding.

[0074] The user dynamic embedding and user dependency embedding are merged according to the gating weight to obtain the final user embedding;

[0075] The initial user activation probability matrix is ​​obtained by embedding the end user into two fully connected layers and then passing it through Softmax.

[0076] Understandably, to accelerate model training and capture changes in user propagation behavior over time, this module divides user propagation behavior into time blocks. Within a batch, the propagation sequence of all information can be divided into multiple groups according to the length of step_len. The latest timestamp in each group is used as the timestamp of all users in that group. Through this process, the transformed timestamps of all users are obtained. Next, each transformed timestamp is assigned to one of n time intervals and converted into a temporal embedding. (temporal embedding), finally the iGCN generated and Combine, use Perform weighted adjustments to generate dynamic user embeddings. .

[0077] Employing a multi-head attention mechanism based on user dynamic embedding Capture relationships between users and generate user dependency embeddings. This is used to measure the relevance of a user to other users. The formula is as follows:

[0078] ;

[0079] in It is a splicing of all user-embedded elements. It is the concatenation of the output matrices from each head. To output the projection matrix;

[0080] The computation process for user-dynamic embedding includes:

[0081] Based on the node embedding representation and temporal embedding between users, the calculation is performed using a scaling factor and then normalized to obtain the dynamic user embedding.

[0082] User-dynamic embedding is as follows:

[0083] ;

[0084] in, Indicates user and users Dynamic embedding of users between them The node embedding representation of the user node The node embedding representation, This indicates the transpose operation. Indicates time embedding, This represents the scaling factor.

[0085] By gating weights Controlling user dynamic embedding With user-dependent embedding The integration of these elements results in end-user embedding. It is expressed as follows:

[0086] ;

[0087] ;

[0088] in, Represents the dot product. These are all learnable parameters. This represents the Sigmoid function.

[0089] The calculation process for the initial user activation probability matrix includes:

[0090] The final user embedding is linearly transformed and then processed by an activation function to obtain the initial user activation probability matrix.

[0091] Embedding end users After feeding in two fully connected layers, then through The initial user activation probability matrix is ​​represented as follows:

[0092] ;

[0093] in, This represents the initial user activation probability matrix. , For the set of real numbers, For the number of users, Dimensions embedded for users , Indicates the number of messages between users. Indicates end-user embedding, , This indicates the maximum number of users a message passes through when it is forwarded. .

[0094] S104. Obtain the privacy leak score and the scale of spread based on the initial user activation probability matrix.

[0095] Understandably, using The benefits of information dissemination and the risks of privacy leakage during the information dissemination process are quantified and transformed into a multi-objective optimization problem for weighing. In view of the problem that traditional methods generate a single solution and are prone to getting trapped in local optimization, this invention designs a privacy diffusion multi-objective optimization algorithm PDO-NSGA-II based on the evolutionary algorithm NSGA-II to generate seed nodes. It is flexible and has a more comprehensive solution set, thereby obtaining the optimal fine-grained privacy control strategy.

[0096] Privacy Breach Rating ( This is used to quantify the risk of privacy breaches and consists of two components:

[0097] Predicting leaks: Assign differentiated "deviation cost" weights to blacklisted and non-blacklisted users' activations.

[0098] Missed activation penalty: Quantify the risk of activation behavior that did not occur but was not predicted by the model (the risk is particularly serious when the user who missed activation is on the blacklist).

[0099] ;

[0100] ;

[0101] in, It is a collection of seed users. It is the set of all users. Derived from the initial user activation probability matrix It is a seed set Under the influence, users Predict the probability of activation. It is the set of users activated in the actual spread. It is the set of users that the model predicts will be "activated", i.e. , It is a collection of blacklisted users. User Predict partial weights, User Missing activation penalty item.

[0102] scale of dissemination ( By accumulating seed sets The propagation utility is quantified by the activation probability of all non-seed users triggered.

[0103] .

[0104] S105. Construct an optimization problem based on the privacy breach score and the scale of the spread.

[0105] To achieve a balance between information utility and privacy protection, a multi-objective optimization problem can be constructed. Its core lies in finding the optimal solution set that minimizes the privacy leakage score while maximizing the propagation scale. This optimization problem is formally defined as follows:

[0106] ;

[0107] in, Represents the solution set in the iteration Privacy breach rating Represents the solution set in the iteration The scale of the spread Represents the solution set in the iteration The number of users.

[0108] S106. Solve the optimization problem to obtain the optimal solution set, and use the users in the optimal solution set as the nodes for next-hop forwarding.

[0109] Understandable Figure 4 This is a flowchart of a privacy diffusion multi-objective optimization algorithm PDO-NSGA-II provided in an embodiment of the present invention, as shown below. Figure 4 As shown, the PDO-NSGA-II (NSGA-II for Privacy-Diffusion Optimization) proposed in this invention balances the scale of information propagation with the risk of privacy leakage, and globally searches for multiple possible seed combinations, offering greater flexibility compared to existing algorithms. The algorithm steps are as follows: First, initialize the population. Randomly generate N seed sets S, and calculate the two objective values ​​for each solution: (Privacy Breach Score) and (Scale of dissemination). Next iteration Generations are generated through selection, crossover, and mutation to produce offspring populations. Evaluate the offspring population and merge the parent and offspring populations to obtain a temporary population. The new parent population is selected based on the non-dominated sorting and crowding distance, and the process is iterated until convergence, outputting the final Pareto optimal solution set.

[0110] Specifically, the goal of each iteration (each generation) is to generate a better next generation from the current generation (parent generation) through a series of operations. Therefore, the main body of this cycle (iteration) is: parent generation → [selection, crossover, mutation] → offspring generation → [merging, evaluation, selection] → new parent generation. This process will be repeated. The iteration process continues until the quality of the solution no longer shows a significant improvement (convergence). This entire iterative process is the core of the NSGA-II algorithm for solving multi-objective optimization problems.

[0111] Specifically, the selection of new parent populations based on non-dominated sorting and crowding distance includes: Non-dominated sorting: This is the primary sorting criterion used for temporary populations. All solutions are hierarchically classified. 1. Define the "dominance" relation: For two solutions (seed sets) A and B: If both objective values ​​of A are better than or equal to those of B (i.e., ... ≤ and ≥ ), and at least one objective value is strictly better than B (i.e. or ), then "A dominates B". 2. Hierarchy (finding Pareto fronts): Front1 (F1): in the temporary population Find all solutions that are not dominated by any other solution. These are the best solutions among the current 2N solutions, forming the first-order Pareto front. Front2 (F2): Temporarily remove the solutions from F1 from consideration. Find all solutions that are not dominated by any other remaining solutions among the remaining solutions. These solutions form the second-order F2. Continue in this manner until a temporary population is reached. All solutions are assigned to a Front(F1,F2,F3,...). Obviously, the solutions in F1 are better than those in F2, F2 is better than F3, and so on.

[0112] Step 2: Crowding Distance Selection: After completing the non-dominated sorting, the next generation of the population begins to be constructed. 1. Fill by level: First, add all solutions for F1. .if If there aren't enough N solutions yet, add all solutions from F2. Continue this process until a certain level Fk is added. The size will exceed N. Using crowding distance as the "draw-win" criterion: Now we face a situation: All solutions from F1 to F(k-1) have been loaded, but we still need to select a few solutions from Fk to make N solutions. All solutions in Fk are of the same level (non-dominant). How do we select them? This is where the crowding distance comes in. Calculating the crowding distance: For each solution in Fk, calculate its "distance" to its siblings in the objective function space. The larger this distance, the more "empty" the solution's surroundings; it's located in a relatively sparse position on the Pareto front. The smaller this distance, the more crowded it is with other similar solutions. Selection rule: To maintain solution diversity, we prioritize solutions with large crowding distances. Specific operation: Sort the solutions in Fk according to their crowding distance from largest to smallest, and then select them sequentially until... The size is exactly N.

[0113] To summarize the selection process: First, compare the levels (Front's index), with the higher level (lower index) winning unconditionally. If the levels are the same, then compare the crowding distance, with the larger distance winning, to ensure solution diversity.

[0114] Compared to single-objective optimization algorithms that use weighted summation, the PDO-NSGA-Ⅱ algorithm generates multiple seed sets at once. These seed sets are non-dominated, allowing users to select seed sets based on their specific needs and generate privacy control strategies that balance privacy and propagation.

[0115] Alternative solutions for the first stage (information diffusion prediction model):

[0116] Alternative Model Architectures: Other Graph Neural Networks: Besides GCNs, graph attention networks, graph isomorphic networks, or other more advanced variants can be used to capture complex dependencies between nodes. Sequence-Based Models: Recurrent neural networks or their variants, such as long short-term memory networks and gated recurrent units, can be used to model the temporal characteristics of user interactions and predict future activation probabilities. Traditional Probabilistic Models: When computational resources are limited, classic independent cascade models or linear threshold models can be used, and their propagation probability parameters can be learned from historical data through machine learning methods (such as maximum likelihood estimation).

[0117] Alternatives to the second stage (multi-objective optimization algorithm):

[0118] Other multi-objective evolutionary algorithms: In addition to NSGA-II, other classic MOEA algorithms can be used, such as SPEA2 (Strength Pareto Evolutionary Algorithm 2) and MOEA / D (Multi-Objective Evolutionary Algorithm based on Decomposition). These algorithms each have their advantages in dealing with Pareto fronts with different characteristics.

[0119] Alternatives to the objective function: Privacy breach scoring ( An alternative definition: Penalty weights can be set non-linearly and more complexly based on the "danger level" of blacklisted users or the "sensitivity level" of sensitive information. Scale of information dissemination ( An alternative definition of ) could be defined as the depth or breadth of the spread, rather than just the total number of activated users.

[0120] Figure 5 This is a block diagram of a fine-grained privacy information propagation control strategy generation device provided in an embodiment of the present invention, such as... Figure 5 As shown, the device 500 includes:

[0121] The heterogeneous graph acquisition module 501 is used to obtain a closeness heterogeneous graph based on the social graph and diffusion graph between users; wherein, the social graph is obtained based on the social relationships between users; the diffusion graph is obtained based on the information diffusion behavior between users within a certain time interval; user nodes in the social graph represent users, and edges represent the social relationships between users;

[0122] The node update module 502 is used to update the node embedding representation of the user node according to the graph neural network and the proximity heterogeneous graph, so as to obtain the updated node embedding representation.

[0123] The probability matrix acquisition module 503 is used to obtain the initial user activation probability matrix based on the updated node embedding representation;

[0124] The privacy parameter acquisition module 504 is used to obtain a privacy leak score and the scale of propagation based on the initial user activation probability matrix.

[0125] The optimization question building module 505 is used to build optimization questions based on privacy breach scores and the scale of dissemination.

[0126] The solution set acquisition module 506 is used to solve the optimization problem, obtain the optimal solution set, and use the users in the optimal solution set as the nodes for next-hop forwarding.

[0127] This invention, by introducing a proximity heterogeneous graph, can more realistically simulate the path of information propagation based on proximity relationships in the real world. This significantly improves the accuracy of information diffusion prediction, ensuring that the subsequently generated privacy policies are based on more reliable and realistic data, resulting in better policy execution. Furthermore, when solving optimization problems, it can adaptively adjust according to the scenario and personal preferences, solving the static and rigid problems of existing technologies. Moreover, the solution set of this invention is closer to the theoretical optimal boundary, providing users with a higher-quality and more comprehensive privacy control scheme, achieving a true balance of optimal utility.

[0128] It should be noted that the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention.

[0129] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0130] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0131] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for generating a fine-grained privacy information dissemination control strategy, characterized in that, The method includes: A heterogeneous graph of intimacy is obtained based on the social graph and diffusion graph between users; wherein, the social graph is obtained based on the social relationships between users; the diffusion graph is obtained based on the information diffusion behavior between users within a certain time interval; user nodes in the social graph represent users, and edges represent the social relationships between users; The node embedding representation of the user node is updated based on the graph neural network and the proximity heterogeneous graph to obtain the updated node embedding representation; The initial user activation probability matrix is ​​obtained based on the updated node embedding representation; The privacy breach score and the scale of spread are obtained based on the initial user activation probability matrix. An optimization problem is constructed based on the privacy breach score and the scale of the spread. Solve the optimization problem to obtain the optimal solution set, and use the users in the optimal solution set as the nodes for next-hop forwarding; The step of obtaining the initial user activation probability matrix based on the updated node embedding representation includes: The user dynamic embedding is obtained based on the updated node embedding representation; Based on the multi-head attention mechanism, user dependency embedding is obtained based on the user dynamic embedding; The gating weights are calculated based on the user dynamic embedding and the user dependency embedding. The user dynamic embedding and the user dependency embedding are fused according to the gating weight to obtain the final user embedding; The final user is embedded into the input and processed through a fully connected layer and an activation function to obtain the initial user activation probability matrix; The step of updating the node embedding representation of user nodes based on the graph neural network and the proximity heterogeneous graph to obtain the updated node embedding representation includes: Add self-loops to the adjacency matrix of the proximity heterogeneous graph and perform symmetric normalization to obtain a normalized adjacency matrix; The node degree matrix is ​​obtained from the adjacency matrix of the aforementioned proximity heterogeneous graph; The node embedding representation of the user node, the normalized adjacency matrix, the node degree matrix, and the weight matrix are multiplied together, and then processed by an activation function and layer normalization to obtain the updated node embedding representation.

2. The method for generating fine-grained privacy information dissemination control strategies according to claim 1, characterized in that, The calculation process for the updated node embedding representation includes: Linear transformations are performed based on the node embedding representation, adjacency matrix, and node degree matrix of the previous layer. The data is then processed using learnable parameters and activation functions, and finally, layer normalization is performed to obtain the updated node embedding representation of the current layer.

3. The method for generating fine-grained privacy information dissemination control strategies according to claim 1, characterized in that, The calculation process for the user-dynamic embedding includes: The user dynamic embedding is obtained by calculating the node embedding representation and temporal embedding between users, combined with a scaling factor, and then normalizing the result.

4. The method for generating fine-grained privacy information dissemination control strategies according to claim 3, characterized in that, The calculation process of the initial user activation probability matrix includes: The final user embedding is linearly transformed and then processed by an activation function to obtain the initial user activation probability matrix.

5. The method for generating fine-grained privacy information propagation control strategies according to claim 4, characterized in that, The process of constructing the optimization problem includes: The optimization problem is constructed based on the privacy leakage score and propagation scale of the solution set in the iteration.

6. A device for generating a fine-grained privacy information dissemination control strategy, characterized in that, The device includes: The heterogeneous graph acquisition module is used to obtain a proximity heterogeneous graph based on the social graph and diffusion graph between users; wherein, the social graph is obtained based on the social relationships between users; the diffusion graph is obtained based on the information diffusion behavior between users within a certain time interval; user nodes in the social graph represent users, and edges represent the social relationships between users; The node update module is used to update the node embedding representation of the user node according to the graph neural network and the proximity heterogeneous graph, so as to obtain the updated node embedding representation. The probability matrix acquisition module is used to obtain the initial user activation probability matrix based on the updated node embedding representation; The privacy parameter acquisition module is used to obtain a privacy leak score and propagation scale based on the initial user activation probability matrix. An optimization problem construction module is used to construct optimization problems based on the privacy breach score and the propagation scale. The solution set acquisition module is used to solve the optimization problem, obtain the optimal solution set, and use the users in the optimal solution set as the nodes for next-hop forwarding; The probability matrix acquisition module is used to obtain the user dynamic embedding based on the updated node embedding representation. Based on the multi-head attention mechanism, user dependency embedding is obtained based on the user dynamic embedding; The gating weights are calculated based on the user dynamic embedding and the user dependency embedding. The user dynamic embedding and the user dependency embedding are fused according to the gating weight to obtain the final user embedding; The final user is embedded into the input and processed through a fully connected layer and an activation function to obtain the initial user activation probability matrix. The node update module is used to add self-loops to the adjacency matrix of the proximity heterogeneous graph and perform symmetric normalization to obtain a normalized adjacency matrix. The node degree matrix is ​​obtained from the adjacency matrix of the aforementioned proximity heterogeneous graph; The node embedding representation of the user node, the normalized adjacency matrix, the node degree matrix, and the weight matrix are multiplied together, and then processed by an activation function and layer normalization to obtain the updated node embedding representation.

Images