A privacy protection method and system for sharing social network environment information

Abstract

This invention discloses a privacy protection method and system for information sharing in a social network environment, comprising: establishing a cascading scale prediction model for privacy information at different time points to obtain the scale of privacy information propagation over time; assessing the influence of privacy recipients on the privacy subject; evaluating privacy risks from both objective and subjectively anticipated perspectives; determining privacy utility based on privacy gains and losses; and summarizing the privacy utility maximization problem in online social networks into a constrained Markov decision process to analyze the decision gains at different time points and formulate appropriate strategies to achieve a balance. By studying the dynamic characteristics of privacy information propagation, the prediction results of privacy information propagation scale are used to assess the influence of the privacy subject and the risks of privacy information sharing. Finally, a privacy protection strategy that maximizes utility is formulated through a Markov decision process to achieve privacy information sharing in online social networks.

Description

Technical Field

[0001] This invention relates to the field of data privacy protection technology, specifically to a method and system for protecting the privacy of information sharing in a social network environment. Background Technology

[0002] Social networks have become a mainstream service used by tens of millions of internet users. They can be used for various activities (such as professional networking, making friends, media sharing, and location services), sharing the common characteristic of enabling users to share information and interact with others. Furthermore, users on social networks typically fill in their personal information on their profiles, such as gender, age, date of birth, and location. Using these social networks brings users many benefits, such as influencing others, enhancing reputation, enjoying brand benefits, and connecting with a large community. However, one of the drawbacks of using social networks is privacy breaches and their potential consequences. Privacy protection technologies, which can provide reliable privacy protection for users, have thus become a key area of ​​research for researchers.

[0003] Past methods for protecting privacy in online social networks can be categorized into three types: 1) Access control-based privacy sharing: This category primarily controls access permissions through information attributes, user roles, dynamic access, and access control lists to achieve privacy protection. 2) Differential privacy-based privacy sharing: This category protects user privacy by adding noise to the information content or graph structure while ensuring the information remains usable. 3) Anonymity-based privacy sharing: This category primarily uses anonymity technology as a key method for privacy protection, aiming to prevent personal data leakage and identity verification by obscuring user identity information.

[0004] The existing solutions provide theoretical basis and methodological support for the research conducted in this invention. However, the traditional access control mechanisms are usually static and difficult to adapt to the dynamic changes in user roles, user numbers and user relationships in social networks. In addition, existing privacy protection settings and tools are often too complicated for ordinary users to operate. Users often use the default settings when setting privacy, which makes it impossible for privacy protection settings to meet users' personalized privacy needs, thus affecting the privacy protection effect. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a privacy protection method and system for information sharing in social network environments. By studying the dynamic characteristics of privacy information dissemination, the prediction results of the scale of privacy information dissemination are used to assess the influence of privacy subjects and the risks of privacy information sharing. Finally, a Markov decision process is used to formulate a privacy protection strategy that maximizes utility in order to achieve privacy information sharing in online social networks.

[0006] This invention discloses a privacy protection method for information sharing in a social network environment, comprising:

[0007] Step 1: Establish a cascading scale prediction model for privacy information at different points in time to obtain the scale of privacy information propagation over time;

[0008] Step 2: Based on the sensitivity of the privacy attributes contained in the privacy information and the dissemination scale obtained in Step 1, assess the influence of the privacy subject's privacy information recipients on the privacy subject.

[0009] Step 3: Assess privacy risks from both objective and subjective perspectives.

[0010] Step 4: Determine privacy benefits based on the influence obtained in Step 2, determine privacy losses based on the privacy risks obtained in Step 3, and determine privacy utility based on privacy benefits and privacy losses.

[0011] Step 5: The problem of maximizing privacy utility in online social networks is summarized as a constrained Markov decision process to analyze the decision payoffs at different time points and formulate appropriate strategies to achieve balance. In the Markov decision process, the reward function at one step is determined by the current state of the system and the selected action. The privacy utility obtained in Step 4 is used as the reward function in the Markov decision process.

[0012] As a further improvement of the present invention, step 1 specifically includes:

[0013] Step 11: Based on graph convolutional networks, capture and model the contextualized structural patterns in the information cascade graph and the implicit relationships of users in social networks to obtain the node feature matrix H. GCN ;

[0014] Step 12: Use the self-attention mechanism in Transformer to model the temporal dependencies in information propagation and obtain the output result H. Trans ;

[0015] Step 13: A Bayesian neural network was used to model the changes and uncertainties in information propagation and information growth, and the output result H was obtained. BNN ;

[0016] Step 14: First, capture structural information through a graph convolutional network, then capture time series dependencies through an LSTM, and finally make predictions through a multilayer perceptron to obtain the scale of privacy information propagation over time.

[0017] As a further improvement of the present invention

[0018] In step 11, H GCN The calculation formula is:

[0019]

[0020] In the formula, H GCN This represents the node feature matrix obtained through GCN. This represents the adjacency matrix after adding self-loops. yes degree matrix; H (l) W represents the feature representation of the nodes in the l-th layer. (l) It is a layer-specific trainable weight matrix, and σ is an activation function that introduces non-linearity;

[0021] In step 12, H Trans The calculation formula is:

[0022]

[0023] In the formula, Q = W Q H GCN K = W K H GCN V = W V H GCN These are the query, key, and value matrices derived from the input sequence, W. Q W K W V These are trainable weight matrices for the query, key, and value, respectively; d k It is the dimension of the key;

[0024] In step 13, H BNN The calculation formula is:

[0025]

[0026] In the formula, H BNN H represents the output of the Bayesian neural network. Trans This represents the output of the Transformer. Given data The posterior distribution of the post-weighted distribution, p(y|H) Trans W) is a given input H Trans The likelihood of the output with weight W;

[0027] In step 14, privacy information I S Propagation scale prediction results at time t The calculation formula is:

[0028]

[0029] In the formula, Indicates privacy information IS Based on the propagation scale prediction results at time t, H GCN h LSTM and h BNN These represent the outputs of GCN, LSTM, and BNN, respectively. concat represents the concatenation operation, and MLP is used for the final prediction.

[0030] As a further improvement of the present invention, in step 2, influence The calculation formula is:

[0031]

[0032] In the formula, δ is the natural decay rate of influence. For privacy information I S The scale of its spread on online social networks, It is privacy information I S Privacy attributes included The sensitivity is α and β, which are weighting parameters.

[0033] As a further improvement of the present invention, in step 3, the privacy risk R(I) S ,t) is:

[0034] R(I S ,t)=r(I S ,t)+N(I S ,t) (6)

[0035] In the formula, r(I) S ,t) represents different times t, I S The objective risks faced; N(I) S ,t) represents different times t, I S Risks based on subjective expectations;

[0036]

[0037] In the formula, It is privacy information I S Privacy attributes included Sensitivity It is privacy information I S The future scale of dissemination in the current environment;

[0038]

[0039] In the formula, μ is the membership function. Indicating a desire to conceal, r(I) S ,t) represents objective risk.

[0040] As a further improvement of the present invention, in step 4, the formula for calculating the privacy utility U(S) is:

[0041]

[0042] In the formula, For the sake of privacy loss, at different times, when the privacy information of the privacy subject S is disseminated in OSNs, σ S =1, otherwise σ S =0.

[0043] As a further improvement of the present invention, step 5 specifically includes:

[0044] Step 51: In the Markov decision-making process, the one-step reward function is jointly determined by the current state of the system and the selected action. This invention selects the utility function U(S) as the reward function in the Markov decision-making process. This function simultaneously considers the social benefits and privacy losses when sharing information and not sharing information, and varies with the privacy subject S's choice σ at different decision moments. S And change;

[0045] Step 52: Relative value iteration is used when solving for the optimal policy. First, the Lagrange method is used to transform the CMDP problem into an unconstrained MDP problem. By introducing the Lagrange multiplier β, the utility function U(S) and the risk function R(I) are compared. S A Lagrange reward function l(s,t) is constructed to solve for the optimal policy; the above Lagrange reward function simultaneously considers the balance between privacy utility and risk, and then the optimal policy is solved by the relative value iteration method;

[0046] Step 53: When transforming the CMDP problem into an unconstrained MDP problem, the reward function L(s,t) is used as the objective function for optimization; at this time, the constraints are R(I S It is determined that there is a relationship between the Lagrange multiplier β and the maximum risk; the Q-learning algorithm is used to determine the maximum risk R that the privacy subject can bear. max And the corresponding Lagrange multiplier β * and through β * By controlling the balance between privacy risks and utility, the reward function is optimized under the unconstrained MDP problem, and the optimal strategy is finally solved.

[0047] Step 54: To obtain the optimal strategy, first test the Lagrange multiplier β. * By perturbating, we obtain β - =β * -Δβ and β + =β * +Δβ; then the perturbed β -and β + Substituting the reward function L(S,t), the corresponding two pure policies δ are solved using the relative value iteration method. - and δ + And combine the utility function U(S) and the risk function R(I) S Calculate the average risk R for each of the following: ,t) - and R + Since the optimal strategy is a mixture of two pure strategies, it can be obtained at each decision point through a comprehensive balance of utility and risk.

[0048]

[0049] In the formula,

[0050] This invention also discloses a privacy protection system for information sharing in a social network environment, comprising:

[0051] The privacy information dissemination scale prediction module is used to implement step 1 of the above privacy protection method;

[0052] The influence measurement and update module is used to implement step 2 of the above privacy protection method;

[0053] The privacy risk measurement module is used to implement step 3 of the above privacy protection method;

[0054] The privacy utility measurement module is used to implement step 4 of the above privacy protection method;

[0055] The privacy policy decision module is used to implement step 5 of the above privacy protection method.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] This invention effectively formulates a privacy protection strategy that conforms to the dynamic dissemination of privacy information by introducing the characteristics of privacy information dissemination in online social networks.

[0058] This invention assesses the influence of privacy subjects and update methods and privacy sharing risks by evaluating the scale of privacy information dissemination, and effectively evaluates the privacy utility that meets users' personalized needs.

[0059] This invention maximizes the privacy utility obtained by constraining the Markov decision process and balancing the gains and losses of privacy information sharing, thereby formulating a privacy protection strategy that meets the personalized needs of users and achieving the goal of balancing privacy protection and privacy sharing. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the privacy protection system for information sharing in a social network environment disclosed in this invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] This invention provides a privacy protection method and system for information sharing in a social network environment. It is a privacy-protected information sharing scheme that studies the dynamic propagation of privacy information in online social networks to promote privacy protection during information sharing. In this scheme, the privacy subject evaluates different future moments and implements the strategy with the greatest utility. Furthermore, to achieve a balance between privacy protection and information sharing, this invention applies reinforcement learning technology to transform the problem into a Markov chain decision process, analyzing the decision benefits at different time points to formulate appropriate strategies to achieve this balance.

[0063] The present invention will now be described in further detail with reference to the accompanying drawings:

[0064] This invention provides a privacy protection method for information sharing in a social network environment, characterized by comprising:

[0065] Step 1: Establish a cascading scale prediction model for privacy information at different points in time to obtain the scale of privacy information propagation over time;

[0066] Specifically, it includes:

[0067] Step 11, Structure Learning: Based on graph convolutional networks, the contextualized structural patterns in the information cascade graph and the implicit relationships of users in social networks are captured and modeled to obtain the node feature matrix H. GCN The layer-by-layer propagation rule of GCNs is defined as follows:

[0068]

[0069] In the formula, H GCN This represents the node feature matrix obtained through GCN. This represents the adjacency matrix after adding self-loops to ensure that each node contains its own characteristics during the aggregation process, thereby capturing the influence of its direct neighbors and the intrinsic properties of the node; yes The degree matrix is ​​used for normalization to avoid the degree matrix at the lth degree. thLayers distort node features, so this method captures the attributes of nodes and their neighborhood structures within that layer; H (l) W represents the feature representation of the nodes in the l-th layer. (l) It is a layer-specific trainable weight matrix used to transform the aggregated features, and σ is an activation function that introduces non-linearity;

[0070] Step 12, Time Propagation Modeling: The self-attention mechanism in Transformer is used to model the temporal dependencies in information propagation. This attention mechanism captures long-term dependencies, improving parallel computation efficiency and yielding the output result H. Trans ;

[0071]

[0072] In the formula, Q = W Q H GCN K = W K H GCN V = W V H GCN These are the query, key, and value matrices derived from the input sequence, W. Q W K W V These are trainable weight matrices for the query, key, and value, respectively. These matrices are used to calculate the correlation (attention score) between different positions in the sequence; d k It is the dimension of the key, used to scale the dot product to maintain the stability of the gradient;

[0073] Step 13, Modeling Propagation Uncertainty: A Bayesian neural network was used to model the changes and uncertainties in information propagation and growth, obtaining more robust prediction results. By treating network weights as random variables, the Bayesian neural network provides a principled method for handling uncertainty, thereby capturing the variability of user behavior and cascading growth. The prediction distribution of the Bayesian neural network is as follows:

[0074]

[0075] In the formula, H BNN H represents the output of the Bayesian neural network. Trans This represents the output of the Transformer. Given data The posterior distribution of the post-weights represents the weight beliefs after observing the data; p(y|H Trans W) is a given input H Trans The likelihood of the output with weight W represents the probability of the output under a specific weight.

[0076] Step 14, Predictor: Combine graph convolutional networks and LSTM to learn a high-order representation of information cascade. First, capture structural information through graph convolutional networks, then capture time series dependencies through LSTM, and finally make predictions through multilayer perceptrons to obtain the scale of privacy information propagation over time.

[0077]

[0078] In the formula, Indicates privacy information I S Based on the propagation scale prediction results at time t, H GCN h LSTM and H BNN These represent the outputs of GCN, LSTM, and BNN, respectively. concat represents the concatenation operation, and MLP is used for the final prediction.

[0079] Step 2: In online social networks, based on the sensitivity of the privacy attributes contained in the privacy information and the dissemination scale obtained in Step 1, assess the influence of the privacy subject's privacy information recipients on the privacy subject.

[0080]

[0081] In the formula, δ is the natural decay rate of influence, representing the natural weakening of influence over time. That is, when a user reduces their speaking time or speaking frequency in OSNs, their influence will gradually decrease over time. For privacy information I S The scale of dissemination on online social networks, i.e., information visibility. It is privacy information I S Privacy attributes included From another perspective, ω represents the degree of interest users have in the privacy information on online social networks, or the quality of that information. α and β are weighting parameters that modulate the impact of information visibility and quality on influence in online social networks.

[0082] Step 3: Assess privacy risk measurement from both objectively faced risks and subjectively anticipated risks; among which, privacy risk R(I) S ,t) is:

[0083] R(I S ,t)=r(I S ,t)+N(I S ,t) (6)

[0084] In the formula, r(I) S ,t) represents different times t, I S The objective risks faced; N(I)S ,t) represents different times t, I S Risks based on subjective expectations;

[0085] The objective risks can be assessed from two aspects: the sensitivity and visibility of privacy information.

[0086]

[0087] In the formula, It is privacy information I S Privacy attributes included Sensitivity It is privacy information I S The future scale of dissemination in the current environment;

[0088] This invention considers the uncertainties inherent in privacy risk assessment, specifically the subjective factor of intention to conceal information. Therefore, fuzzy mathematics can be used to describe the uncertainty of privacy risk assessment. This invention takes the subjective intention to conceal information and calculates subjective privacy risk based on fuzzy theory. Based on the intention to conceal information... The corresponding fuzzy set and membership function can be set. Finally, the output logical language value is defuzzified. Since the region centroid method is applicable to Mamdani-type fuzzy models, this invention selects the region centroid method to transform the subjective intention to conceal into an accurate subjective privacy risk assessment quantity. Specifically, if the value range of the fuzzy set is [0,1], its calculation formula is as follows:

[0089]

[0090] In the formula, μ is the membership function. Indicating a desire to conceal, r(I) S ,t) represents objective risk.

[0091] Step 4: Determine privacy benefits based on the influence obtained in Step 2, determine privacy losses based on the privacy risks obtained in Step 3, and determine privacy utility based on privacy benefits and privacy losses.

[0092] Privacy benefits of information sharing: Influence is used to describe the impact of social benefits on individual decision-making. This is used to simulate the social benefits individuals gain when they decide to share information with their audience.

[0093] Privacy loss from sharing information: Privacy loss refers to an individual's assessment of privacy breaches during the decision-making process of sharing personal information. Therefore, the privacy risk value can be directly used to assess privacy loss; the privacy cost for individual S is...

[0094] Privacy benefits of not sharing information: Since online social networks are primarily platforms for communication between users, let's assess the social benefits, i.e., the impact, when individuals choose not to share personal information at the current moment.

[0095] Privacy loss from not sharing information: Similar to the privacy loss from sharing information, the privacy loss from not sharing information is determined using forward privacy risk. However, in this case, the privacy risk value is zero, and the privacy loss is zero.

[0096] The above analysis shows that the utility of sharing information is the sum of social benefits and privacy losses, that is... in It is a negative value, representing an individual's loss. Furthermore, the utility of not sharing information is solely determined by current social interests. And the loss of privacy in the previous moment The sum determines, that is Therefore, an individual S's decision about whether to share information with an audience can be determined by measuring utility value, as shown in the utility function below:

[0097]

[0098] In the formula, For the sake of privacy loss, at different times, when the privacy information of the privacy subject S is disseminated in OSNs, σ S =1, otherwise σ S =0.

[0099] Step 5: The problem of maximizing privacy utility in online social networks is summarized as a constrained Markov decision process in order to analyze the decision benefits at different time points and formulate appropriate strategies to achieve a balance.

[0100] Specifically, it includes:

[0101] Step 51: In the Markov decision-making process, the one-step reward function is jointly determined by the current state of the system and the selected action. This invention selects the utility function U(S) as the reward function in the Markov decision-making process. This function simultaneously considers the social benefits and privacy losses when sharing information and not sharing information, and varies with the privacy subject S's choice σ at different decision moments. S And change;

[0102] Step 52: Relative Value Iteration (RVI) is used to solve for the optimal policy. First, the Lagrange method is used to transform the CMDP problem into an unconstrained MDP problem. By introducing the Lagrange multiplier β (β>0), the utility function U(S) and the risk function R(I) are compared. SThe optimal policy is solved by constructing a Lagrange reward function L(S,t) from the above Lagrange reward function, which simultaneously considers the balance between privacy utility and risk, and then solves the optimal policy through the relative value iteration method.

[0103] Step 53: When transforming the CMDP problem into an unconstrained MDP problem, the reward function L(S,t) is used as the objective function for optimization; at this point, the constraint (i.e., the maximum risk R that the privacy subject can bear) is... max ) by R(I S It is determined that there is a relationship between the Lagrange multiplier β and the maximum risk; the Q-learning algorithm is used to determine the maximum risk R that the privacy subject can bear. max And the corresponding Lagrange multiplier β * and through β * By controlling the balance between privacy risks and utility, the reward function is optimized under the unconstrained MDP problem, and the optimal strategy is finally solved.

[0104] Step 54: To obtain the optimal strategy, first test the Lagrange multiplier β. * By perturbating, we obtain β - =β * -Δβ and β + =β * +Δβ; then the perturbed β - and β + Substituting the reward function L(S,t), the corresponding two pure policies δ are solved using the relative value iteration method. - and δ + And combine the utility function U(S) and the risk function R(I) S Calculate the average risk R for each of the following: ,t) - and R + Since the optimal policy is a mixture of two pure policies, this invention defines a parameter q such that... Based on the specific value of q obtained from this relationship, the optimal strategy can ultimately be obtained at each decision point through a comprehensive balance between utility and risk:

[0105]

[0106] like Figure 1 As shown, the present invention provides a privacy protection system for information sharing in a social network environment, comprising:

[0107] The privacy information dissemination scale prediction module is used to implement step 1 of the above privacy protection method;

[0108] The influence measurement and update module is used to implement step 2 of the above privacy protection method;

[0109] The privacy risk measurement module is used to implement step 3 of the above privacy protection method;

[0110] The privacy utility measurement module is used to implement step 4 of the above privacy protection method;

[0111] The privacy policy decision module is used to implement step 5 of the above privacy protection method.

[0112] The advantages of this invention are:

[0113] 1. Current privacy protection technologies for information sharing in online social networks mainly rely on static information and user personal information to formulate privacy protection strategies, ignoring the dynamically changing user roles, user relationships, and information dissemination in online social networks. This invention effectively formulates privacy protection strategies that conform to the dynamic dissemination of privacy information by introducing the characteristics of privacy information dissemination in online social networks.

[0114] 2. This invention assesses the influence of privacy subjects and update methods and privacy sharing risks by evaluating the scale of privacy information dissemination, and effectively evaluates the privacy utility that meets users' personalized needs.

[0115] 3. This invention maximizes the privacy utility obtained by constraining the Markov decision process and balancing the gains and losses of privacy information sharing, thereby formulating a privacy protection strategy that meets the personalized needs of users and achieving the goal of balancing privacy protection and privacy sharing.

[0116] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for protecting privacy in information sharing within a social network environment, characterized in that, include: Step 1: Establish a cascading scale prediction model for privacy information at different points in time to obtain the scale of privacy information propagation over time; Specifically, it includes: Step 11: Based on graph convolutional networks, capture and model the contextualized structural patterns in the information cascade graph and the implicit relationships of users in social networks to obtain the node feature matrix H. GCN ; Step 12: Use the self-attention mechanism in Transformer to model the temporal dependencies in information propagation and obtain the output result H. Trans ; Step 13: A Bayesian neural network was used to model the changes and uncertainties in information propagation and information growth, and the output result H was obtained. BNN ; Step 14: First, capture structural information using a graph convolutional network; then, capture temporal dependencies using an LSTM; finally, predict the propagation scale of privacy information over time using a multilayer perceptron. In step 14, the privacy information I... S Propagation scale prediction results at time t The calculation formula is: In the formula, Indicates privacy information I S Based on the propagation scale prediction results at time t, H GCN h LSTM and H BNN These represent the outputs of GCN, LSTM, and BNN, respectively. concat represents the concatenation operation, and MLP is used for the final prediction. Step 2: Based on the sensitivity of the privacy attributes contained in the privacy information and the dissemination scale obtained in Step 1, assess the influence of the privacy subject's privacy information recipients on the privacy subject. Step 3: Assess privacy risks from both objective and subjective perspectives. Step 4: Determine privacy benefits based on the influence obtained in Step 2, determine privacy losses based on the privacy risks obtained in Step 3, and determine privacy utility based on privacy benefits and privacy losses. Step 5: The problem of maximizing privacy utility in online social networks is summarized as a constrained Markov decision process to analyze the decision payoffs at different time points and formulate appropriate strategies to achieve balance. In the Markov decision process, the reward function at one step is determined by the current state of the system and the selected action. The privacy utility obtained in Step 4 is used as the reward function in the Markov decision process.

2. The privacy protection method for information sharing in a social network environment as described in claim 1, characterized in that, In step 11, H GCN The calculation formula is: In the formula, H GCN This represents the node feature matrix obtained through GCN. This represents the adjacency matrix after adding self-loops. yes degree matrix; H (l) W represents the feature representation of the nodes in the l-th layer. (l) It is a layer-specific trainable weight matrix, and σ is an activation function that introduces non-linearity; In step 12, H Trans The calculation formula is: In the formula, Q = W Q H GCN K = W K H GCN V = W V H GCN These are the query, key, and value matrices derived from the input sequence, W. Q W K W V These are trainable weight matrices for the query, key, and value, respectively; d k It is the dimension of the key; In step 13, H BNN The calculation formula is: In the formula, H BNN H represents the output of the Bayesian neural network. Trans This represents the output of the Transformer. It is the posterior distribution of the weights given data D, p(y|H) Trans W) is a given input H Trans The likelihood of the output with weight W.

3. The privacy protection method for information sharing in a social network environment as described in claim 2, characterized in that, In step 2, influence The calculation formula is: In the formula, δ is the natural decay rate of influence. For privacy information I S The scale of propagation at time t-1 in online social networks, It is privacy information I S Privacy attributes included The sensitivity is α and β, which are weighting parameters.

4. The privacy protection method for information sharing in a social network environment as described in claim 3, characterized in that, In step 3, the privacy risk R(I) S ,t) is: R(I S ,t)=r(I S ,t)+N(I S ,t) (6) In the formula, r(I) S ,t) represents different times t, I S The objective risks faced; N(I) S ,t) represents different times t, I S Risks based on subjective expectations; In the formula, It is privacy information I S Privacy attributes included Sensitivity It is privacy information I S The future scale of dissemination in the current environment; In the formula, μ is the membership function. Indicating a desire to conceal, r(I) S ,t) represents objective risk.

5. The privacy protection method for information sharing in a social network environment as described in claim 4, characterized in that, In step 4, the formula for calculating the privacy utility U(S) is: In the formula, For the privacy loss at time t and time t-1, σ represents the privacy loss when the privacy information of the privacy subject S is propagated in OSNs at different times. S =1, otherwise σ S =0.

6. The privacy protection method for information sharing in a social network environment as described in claim 5, characterized in that, Step 5 specifically includes: Step 51: In the Markov decision process, the one-step reward function is jointly determined by the current state of the system and the selected action. The utility function U(S) is selected as the reward function in the Markov decision process. This function simultaneously considers the social benefits and privacy losses when sharing information and not sharing information, and varies with the privacy subject S's choice σ at different decision moments. S And change; Step 52: Relative value iteration is used when solving for the optimal policy. First, the Lagrange method is used to transform the CMDP problem into an unconstrained MDP problem. By introducing the Lagrange multiplier β, the utility function U(S) and the risk function R(I) are compared. S The optimal policy is solved by constructing a Lagrange reward function L(S,t) from the above Lagrange reward function, which simultaneously considers the balance between privacy utility and risk, and then solves the optimal policy through the relative value iteration method. Step 53: When transforming the CMDP problem into an unconstrained MDP problem, the reward function L(S,t) is used as the objective function for optimization; at this time, the constraints are R(I... S It is determined that there is a relationship between the Lagrange multiplier β and the maximum risk; the Q-learning algorithm is used to determine the maximum risk R that the privacy subject can bear. max And the corresponding Lagrange multiplier β * and through β * By controlling the balance between privacy risks and utility, the reward function is optimized under the unconstrained MDP problem, and the optimal strategy is finally solved. Step 54: To obtain the optimal strategy, first test the Lagrange multiplier β. * By perturbating, we obtain β - =β * -Δβ and β + =β * +Δβ; then the perturbed β - and β + Substituting the reward function L(S,t), the corresponding two pure policies δ are solved using the relative value iteration method. - and δ + And combine the utility function U(S) and the risk function R(I) S Calculate the average risk for each of the following: ,t) and Since the optimal strategy is a hybrid of two pure strategies, it can be obtained at each decision point through a comprehensive balance of utility and risk: In the formula, 7. A privacy protection system for information sharing in a social network environment, used to implement the privacy protection method for information sharing in a social network environment as described in any one of claims 1 to 6, characterized in that, include: The privacy information dissemination scale prediction module is used to implement step 1 of the privacy protection method; The influence measurement and update module is used to implement step 2 of the privacy protection method; The privacy risk measurement module is used to implement step 3 of the privacy protection method. The privacy utility measurement module is used to implement step 4 of the privacy protection method; The privacy policy decision module is used to implement step 5 of the privacy protection method.

Images