A method and system for assessing privacy risks and utility of synthetic graph data

By employing a two-stage workflow and a quantitative evaluation method for privacy gains and utility losses, this study addresses the challenge of unifying the assessment of privacy risks and utility in synthetic graph data. It enables comparable assessments and specific references within the same scenario, providing a more accurate evaluation framework for the privacy protection of synthetic graphs.

CN123116339APending Publication Date: 2026-11-17HARBIN INST OF TECH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611537585.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-30
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies lack a unified assessment method for evaluating the privacy risks and utility of synthetic graph data, making it difficult to compare privacy and utility performance in the same scenario, and making it difficult to observe the actual level of protection. There is a lack of empirical evidence of the reduction in privacy risks with reference to the original graph and the difference assessment under the same query workload.

Method used

A two-stage workflow is adopted, organizing the differential private graph synthesis method into a representation stage and a private learning and synthesis stage. Privacy risks and utility are quantified by privacy gain and utility loss indicators, and evaluated in conjunction with a consistent threat model and attack settings. Probabilistic constructors, community-based and block-based generators, and neural graph generators are used, combined with privacy implementations such as Laplace mechanism and differential privacy stochastic gradient descent, to quantify privacy gains and utility losses.

Benefits of technology

It enables a unified assessment of the privacy risks and utility of synthetic graphs in the same scenario, directly quantifies the reduction in privacy risks and utility differences of target elements, reduces duplication of work, facilitates the comparison and selection of methods, provides specific privacy concepts and budget selection references, and improves the comparability of cross-method assessments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116339A_ABST
    Figure CN123116339A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for evaluating privacy risks and utility of synthetic graph data, relating to the field of graph data privacy protection technology. The method organizes the Differential Private Graph Synthesis (DPGS) method into a two-stage workflow: a representation stage and a private learning and synthesis stage. It uses two indicators—privacy gains and utility losses—to jointly evaluate the release of the synthetic graph. Privacy gains are determined by comparing the difference in attack advantages between releasing the synthetic graph and releasing the original graph under member inference attacks and attribute inference attacks. Utility losses are determined by calculating the task output difference between the synthetic graph and the original graph under the same graph query workload, using a relative error form or a task-specific difference function form. This invention provides empirical evidence for the selection of methods, privacy budgets, and privacy concepts in graph data publishing and supports the expansion of new DPGS methods, attacks, and utility indicators.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of graph data privacy protection technology, and in particular to a method and system for assessing the privacy risks and utility of synthetic graph data. Background Technology

[0002] The privacy risk and utility assessment of synthetic graph data is used to quantify two aspects of the effect after the synthetic graph is published: first, privacy risk, which examines whether attackers can infer the membership or sensitive attributes of target elements in the original graph based on the published results and auxiliary knowledge under a set threat model, in order to determine how much the synthetic graph reduces the inference risk compared to the published original graph; second, utility, which examines how close the synthetic graph is to the output of the original graph under the same set of graph queries. Scalars are mostly used for relative error, while structured objects such as distributions are measured by difference.

[0003] Currently, in the evaluation of differential private graph synthesis methods, due to the differences in graph representation, privacy implementation, and generators among various methods, existing evaluations mostly report utility under given privacy concepts and budgets. They lack empirical evidence of privacy risk reduction with reference to the original graph publication and the difference in task outputs under the same query workload. The conclusions at both ends are often taken from different threat models or datasets, and the gap between nominal budget and actual protection is not easy to observe. This makes it difficult to compare privacy and utility performance across methods in the same scenario, and the empirical evidence that can be relied upon is limited.

[0004] Therefore, a method and system for assessing privacy risks and utility of synthetic graph data are proposed to address the above issues. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for assessing privacy risks and utility of synthetic graph data, in order to address the problems mentioned in the background above.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a method for assessing privacy risks and utility of synthetic graph data, comprising the following steps:

[0007] S1. The Difference Private Graph Synthesis (DPGS) method is organized into a two-stage workflow, including a representation stage and a private learning and synthesis stage. The representation stage converts the original graph... Mapping to a compact representation of the structural signal targeted by the capture method, the private learning and synthesis stages take this compact representation as input, enforce privacy under the selected privacy concept and privacy budget, and generate a synthetic graph. ;

[0008] S2. The privacy and utility behaviors of synthetic graph release are characterized by two indicators: privacy gains and utility losses. Privacy gains quantify the impact of publishing synthetic graphs under the same attack settings. Compared to releasing the original image Reduced empirical privacy risks, the utility loss quantifies the synthetic graph under the same graph query workload. Compared with the original image Differences in task output between them;

[0009] S3. Under consistent threat models, attack settings, datasets, and downstream tasks, conduct a joint privacy and utility evaluation of the release of the synthetic graph GS based on privacy gains and utility losses.

[0010] Preferably, the private learning and synthesis phase is defined by two coupled components: a generator and a privacy implementation. The generator specifies how to generate a graph from a compact representation, and the privacy implementation defines the privacy concept and a differential privacy DP mechanism for privatizing the generation process. The generator includes a probabilistic constructor, a community- and block-based generator, and a neural graph generator. The privacy implementation relies on the Laplace mechanism, the exponential mechanism, and differential privacy stochastic gradient descent DP-SGD.

[0011] Preferably, the family of representations for compact representation in the representation stage includes degree-based statistics, adjacency-based counting, matrices, community-level summaries, and continuous node representations, i.e., node embeddings; wherein degree-based statistics are used by DP-dK and DGG, adjacency-based representations are used by TmF and PrivSKG, community structures are used by PrivGraph, and node embeddings are used by PrivDPR.

[0012] Preferably, the privacy implementation enforces privacy through a corresponding privacy-enforcing (DP) mechanism. When multiple methods employ the same DP mechanism, the implementation of the DP mechanism and related privacy calculations are aligned, and the relevant privacy parameter settings are further aligned to ensure that DP remains effective and comparable across graphs at different scales. For DP-SGD-based methods, According to the number of protected units Relevant rules are set to enable The value varies with the number of protected units. The value decreases as it increases, under node-level DP. That is, the number of nodes in the original graph, under edge-level DP. That is, the number of edges in the original graph; let... This is the original image. and These are its node set and edge set, respectively. and , respectively representing the original image The number of nodes and the number of edges.

[0013] Preferably, after privacy protection is enforced, each method constructs a graph generator based on its selected compact representation and privatized output, and generates a synthetic graph from the graph generator. The method-specific strategies of the graph generator include probabilistic graph reconstruction, top edge selection, community-based graph generation, BTER-based graph generation, and embedding-based edge sampling. The generator design of each method is preserved, and relevant implementations are reused to support controlled comparisons when the generator construction and generation process is shared by multiple methods and future extensions.

[0014] Preferably, the privacy benefits are defined in the following way:

[0015] In the publishing environment, trusted data curators hold the original images. Apply the DPGS mechanism to M and publish the synthetic graph. After observing the release, the attacker attempts to infer sensitive information about the protected elements in the original graph using auxiliary background knowledge;

[0016] Each privacy threat is modeled as attacker A, who aims to infer target elements. The secret, the target element Representing the target node and edges, the secret corresponds to the target's membership and related sensitive attributes; defining the attack advantage. Measure when the target exists in the published graph The degree to which an attacker's reasoning ability is enhanced when not present is calculated based on the difference in success rates between inclusion and exclusion attacks; the greater the attack advantage, the higher the privacy risk of the target element.

[0017] The privacy benefit is achieved by releasing the composite image. Instead of the original image To reduce the part of the attack advantage, namely:

[0018] .

[0019] Preferably, when the secret is the member identity of the target, the attacker is instantiated as a member inference attack, including node target member relationship inference and edge target member relationship inference; in the associative game between the attacker and the challenger, the attack advantage, assuming the secret bits are a priori equal, is determined by the difference between the probability of the attacker's output indicating the existence of the target in the target inclusion case and the target exclusion case, which is equivalent to twice the probability of the attacker guessing that the secret bit is equal to one; since the attacker can directly check the member identity with a probability of 1 when the original graph is released, that is, the attack advantage under the release of the original graph is 1, the privacy gain is:

[0020] ;

[0021] in The attack advantage released by the composite graph; This indicates that synthetic release offers no protection against member inference. This indicates that attackers have no advantage over those who guess randomly.

[0022] Preferably, when the secret is a sensitive attribute of the target, the attacker is instantiated as an attribute inference attacker: the attacker observes a portion of the target record. And attempts to infer the values ​​of the retained sensitive attributes based on the released graph and auxiliary knowledge P. The challenger assigned sensitive attribute values ​​to a subset of target records based on the overall reference graph. This constitutes a complete target record. ;

[0023] The attack advantage is the difference between the probability that the estimated value equals the true sensitive attribute value when the target is included and the probability that the estimated value equals the true sensitive attribute value when the target is excluded; the privacy gain is the difference between the attack advantage under the release of the original graph and the attack advantage under the release of the synthetic graph, i.e.:

[0024] .

[0025] Preferably, the utility loss is defined in a task-based manner, given a graph query task T, let Representation diagram The task output above, for the original graph and composite image The loss of utility is denoted as A smaller utility loss value indicates better utility preservation;

[0026] For tasks whose output is scalar and aggregate statistics, the utility loss is instantiated using relative error:

[0027] ;

[0028] For tasks where the output is a distribution, vector, or other structured objects, use a task-specific difference function. definition:

[0029] ;

[0030] This is used when DIS is a similarity metric where a larger value is always better. Convert it into a loss so that the utility loss is always as small as possible.

[0031] A privacy risk and utility assessment system for synthetic graph data includes a unified assessment framework and modular method integration components;

[0032] The unified evaluation framework introduces two metrics, privacy gains and utility losses, to characterize the privacy and utility behavior of releasing a synthetic graph. The privacy gains measure the degree to which releasing a synthetic graph reduces empirical privacy risks compared to releasing the original graph, and the utility losses measure the difference between the synthetic graph and the original graph under the same graph query workload.

[0033] The modular method integration component integrates representative DPGS methods within a shared organizational and evaluation framework. By dividing each DPGS method into a representation phase and a private learning and synthesis phase, it localizes method differences into three parts: graph representation, privacy implementation, and graph generator construction and generation. It also aligns comparable components and evaluation conditions where applicable for controlled comparisons. Specifically, when multiple methods rely on the same representation family, the corresponding implementation components and key preprocessing choices are aligned, including degree-based methods using normalized degree histograms under the same construction rules, adjacency-based methods using consistent sparse coding, and community methods using a fixed-set Leuven algorithm.

[0034] Privacy gains and utility losses were measured across different methods under the same threat model, attack instantiation, target selection rules, and utility settings.

[0035] The system supports the integration of additional DPGS methods by mapping the pipeline of new methods to the same two-stage organization, and extends the system by incorporating additional attack and utility metrics through privacy gains and utility losses defined at the framework level.

[0036] The present invention has the following beneficial effects:

[0037] 1. In this invention, by unifying various methods into a two-stage process and aligning comparable components such as histogram construction, sparse coding, and community detection, the problem of different methods handling graph representation and privacy implementation, and the difficulty in attributing cross-method differences to the method design itself, is improved. This makes it easier to distinguish the relative performance under different representation families or different mechanisms. The comparison results largely reflect the differences brought about by representation selection, mechanism, or generator design, rather than inconsistencies between the implementation level and the preprocessing level. This helps to reduce the repetitive work of rebuilding the pipeline for each method and makes it easier for new methods to be integrated in the same way.

[0038] 2. In this invention, by defining privacy benefits with the original graph publication as a reference, and calculating them as the difference in attack advantage under both membership inference and attribute inference, the invention improves the situation where the strength of protection can only be inferred based on nominal privacy concepts and budgets, while the actual level of protection is not easily observable. As a result, the risk reduction of the inference of the membership and sensitive attributes of target elements after the synthetic graph is published can be directly quantified, which helps to compare the privacy performance of various methods under similar actual protection levels, without having to rely on nominal budgets as the sole basis, and provides a more specific reference for the selection of privacy concepts and budgets at the time of publication.

[0039] 3. In this invention, by calculating the utility loss using relative error or difference functions under the same query workload and jointly evaluating it under consistent threat models, attack settings, datasets, and downstream tasks, the invention improves the situation where privacy and utility conclusions are scattered across different settings and difficult to compare in the same scenario. As a result, the privacy and utility performance of each method under the same release settings can be observed simultaneously, providing more sufficient empirical information when selecting methods, privacy budgets, and privacy concepts for specific release tasks. New methods can be accessed using the same two-stage organization, and new attack and utility metrics can be incorporated at the framework level. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating a method for assessing privacy risks and utility of synthetic graph data according to the present invention.

[0041] Figure 2 This is a schematic diagram of the overall architecture of a privacy risk and utility assessment system for synthetic graph data according to the present invention;

[0042] Figure 3 This invention provides a comparative graph of node member inference attack (nMIA) under similar empirical privacy protections, representing a privacy risk and utility assessment system for synthetic graph data.

[0043] Figure 4 This diagram compares a privacy risk and utility assessment system for synthetic graph data of the present invention against node attribute inference attacks (nAIA) under similar empirical privacy protection. Detailed Implementation

[0044] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0045] Example 1, please refer to Figures 1-4 As shown: A method for assessing privacy risks and utility of synthetic graph data, comprising the following steps:

[0046] S1. The Difference Private Graph Synthesis (DPGS) method is organized into a two-stage workflow, including a representation stage and a private learning and synthesis stage. The representation stage converts the original graph... Mapping to a compact representation of the structural signal targeted by the capture method, the private learning and synthesis stages take this compact representation as input, enforce privacy under the selected privacy concept and privacy budget, and generate a synthetic graph. ;

[0047] S2. The privacy and utility behaviors of synthetic graph release are characterized by two indicators: privacy gains and utility losses. Privacy gains quantify the impact of publishing synthetic graphs under the same attack settings. Compared to releasing the original image Reduced empirical privacy risks, utility loss quantified in synthetic graphs under the same graph query workload. Compared with the original image Differences in task output between them;

[0048] S3. Under consistent threat models, attack settings, datasets, and downstream tasks, conduct a joint privacy and utility evaluation of the release of the synthetic graph GS based on privacy gains and utility losses.

[0049] The private learning and synthesis phase is defined by two coupled components: a generator and a privacy implementation. The generator specifies how to generate graphs from compact representations, including probabilistic constructors, community-based or block-based generators, and neural graph generators. The privacy implementation defines the privacy concept and the differential private graph synthesis DP mechanism used for the privatization generation process, relying on the Laplacian mechanism, the exponential mechanism, and differential privacy stochastic gradient descent DP-SGD. The differential privacy stochastic gradient descent DP-SGD clips the gradient of each sample during training and averages it with Gaussian noise. Its privacy cost is calculated using privacy accounting methods such as moment accounting or RényiDP, thereby constructing a private generator that satisfies formal differential privacy guarantees.

[0050] During the representation phase, the compact representation family includes degree-based statistics, adjacency-based counting or matrices, community-level summaries, and continuous node representations, i.e., node embeddings. Degree-based statistics are used by DP-dK and DGG, adjacency-based representations are used by TmF and PrivSKG, community structures are used by PrivGraph, and node embeddings are used by PrivDPR. When multiple methods depend on the same representation family, the corresponding implementation components and key preprocessing choices are aligned, including degree-based methods using normalized degree histograms under the same construction rules, adjacency-based methods using consistent sparse coding, and community methods using the Leuven algorithm with a fixed setting.

[0051] Privacy is enforced through appropriate privacy-enforcing mechanisms. When multiple methods employ the same privacy-enforcing mechanism, the implementation of the DP mechanism and related privacy calculations are aligned, and the relevant privacy parameter settings are further aligned to ensure that DP remains effective and comparable across graphs at different scales. For DP-SGD-based methods, [the following steps are taken]. Configure it according to the rules related to the number of protected units N, so that The value varies with the number of protected units. The value decreases as it increases, under node-level DP. That is, the original image The number of nodes in the middle, under edge-level DP That is, the original image The number of sides in the middle; let This is the original image. and These are its node set and edge set, respectively. and , respectively representing the original image The number of nodes and the number of edges; specifically, set... This setting adapts to the protected units and map scale.

[0052] After privacy protection is enforced, each method constructs a graph generator based on its chosen compact representation and privatized output, and generates a synthetic graph from the graph generator. The method-specific strategies for the graph generator include probabilistic graph reconstruction, top edge selection, community-based graph generation, BTER-based graph generation, and edge sampling based on embedding. The generator design for each method is preserved, and the relevant implementations are reused to support controlled comparisons when the generator construction and generation process is shared by multiple methods and future extensions.

[0053] Privacy benefits are defined as follows:

[0054] In the publishing environment, trusted data curators hold the original images. Apply the DPGS mechanism to M and publish the synthetic graph. After observing the release, the attacker attempts to infer sensitive information about the protected elements in the original graph using auxiliary background knowledge;

[0055] Each privacy threat is modeled as attacker A, whose aim is to infer target elements. The secret, the target element Representing the target node and edges, the secret corresponds to the target's membership and related sensitive attributes; defining the attack advantage. Measure when the target exists in the published graph The degree to which an attacker's reasoning ability is enhanced when not present is calculated based on the difference in success rates between inclusion and exclusion attacks; the greater the attack advantage, the higher the privacy risk of the target element.

[0056] Privacy benefits are obtained by releasing synthetic images. Instead of the original image To reduce the part of the attack advantage, namely:

[0057] .

[0058] When the secret is the target's membership identity, the attacker is instantiated as a membership inference attacker, including node target membership inference and edge target membership inference; in the associative game between the attacker and the challenger, the challenger samples whether the target is retained in the secret position of the original graph. If the secret bit is 1, then the target element will be... Retain in the original image; if the value is 0, then... Replace with non-target elements extracted from the corresponding non-target pool induced by the overall reference graph to obtain a graph with missing targets; challengers are ranked according to the number of target nodes. and the number of target edges From private generator Sampling yields composite image Sample the public bit b, and release the original graph when the public bit is 0. When the value is 1, the composite image is released. As a published image Attackers observe published graphs Public bits, target element And auxiliary knowledge P, outputting a guess about the membership of the target element. The attack advantage, assuming the secret bits are a priori equal, is determined by the difference between the probability that the attacker's output represents the existence of the target in the case of target inclusion and target exclusion. This is equivalent to doubling the attacker's guess equal to the secret bit and subtracting one, i.e.:

[0059] ;

[0060] Since an attacker can directly check member identities with a probability of 1 when the original graph is released, the attack advantage under the condition of the original graph being released is 1. Therefore, the privacy gain is: ;in The attack advantage released by the composite graph; This indicates that synthetic release offers no protection against member inference. This indicates that the attacker has no advantage over random guessing; when the synthetic graph is released, the attacker is instantiated as a black-box supervised classifier, which constructs shadow graphs containing and not containing the target under auxiliary knowledge P, applies the same synthesis process to obtain the shadow synthetic graph, extracts graph-level features and target-related features from it, and trains a binary classifier to predict the membership of the target element based on the released synthetic graph.

[0061] When the secret is a sensitive attribute related to the target, the attacker is instantiated as an attribute inference attacker: the attacker observes part of the target record. And attempts to infer the values ​​of the retained sensitive attributes based on the released graph and auxiliary knowledge P. The challenger assigned sensitive attribute values ​​to a subset of target records based on the overall reference graph. This constitutes a complete target record. The sampling characterizes whether the complete target is contained in the secret bit of the original graph, and constructs the original graph accordingly. And according to the number of target nodes and the number of target edges The composite graph is obtained by sampling from the private generator. The challenger samples the public bit and releases the original graph when the public bit is 0. When the value is 1, the composite image is released. As a published image Attackers observe published graphs Public bits, partial target records And auxiliary knowledge P, outputting an estimate of the retained sensitive attributes. The attack advantage is the difference between the probability that the estimated value equals the true sensitive attribute value when the target is included and the probability that the estimated value equals the true sensitive attribute value when the target is excluded.

[0062] ;

[0063] The privacy gain is the difference between the attack advantage under the release of the original graph and the attack advantage under the release of the composite graph, i.e.:

[0064] ;

[0065] Treating attribute inference as a supervised learning problem: Attackers utilize the released graph and partial target records And auxiliary knowledge P, train predictor variables from observed target information to retain sensitive attributes, and instantiate predictor variables into classifiers or regression variables according to the type of sensitive attribute considered in the attack.

[0066] The larger the PG value, the more significantly attribute leakage can be reduced by replacing the original release with the synthetic release.

[0067] Utility loss is defined in a task-based manner, given a graph query task T, let Representation diagram The task output above, for the original graph and composite image The loss of utility is denoted as A smaller utility loss value indicates better utility preservation; for tasks with outputs of scalars and aggregate statistics, the utility loss is instantiated using relative error:

[0068] ;

[0069] For tasks where the output is a distribution, vector, or other structured objects, use a task-specific difference function. definition:

[0070] ;

[0071] when Use when the similarity index is as high as possible. Convert this into a loss, so that the utility loss is always as small as possible; in the case of publishing a private generator, adjust the loss based on the number of target nodes from the private generator. and the number of target edges The utility loss of each of the sampled composite graphs is calculated, and the average of the multiple utility losses is taken to report the task-level average utility loss. Specifically, 15 graph query tasks are used, including the number of nodes, the number of edges, the number of triangles, the average degree, the degree variance, the degree distribution, the diameter, the average shortest path length, the distance distribution, the global clustering coefficient, the average clustering coefficient, community detection, modularity, the homogeneity coefficient, and the eigenvector centrality score. The corresponding dissimilarity measures include KL divergence, Hellinger distance, Kolmogorov-Smirnov statistic, NMI, and mean absolute error (MAE).

[0072] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for privacy risk and utility assessment oriented to synthetic graph data, characterized in that, Includes the following steps: S1, organizing the differentially private graph synthesis, DPGS, method as a two-stage workflow comprising a representation stage and a private learning and synthesis stage, wherein the representation stage maps the original graph to a compact representation of the structural signal that the mapping is targeted at, and the private learning and synthesis stage takes the compact representation as input, enforces privacy under the chosen privacy notion and privacy budget, and generates the synthetic graph ; S2, representing the privacy and utility behavior of the synthetic graph release through two indicators, privacy gain and utility loss, where the privacy gain quantifies the released synthetic graph under the same attack setting Reduced empirical privacy risk, the utility loss quantifies the task output difference between the synthetic graph and the original graph under the same graph query workload; S3. Under consistent threat model, attack setting, dataset, and downstream task, evaluate the synthetic graph according to privacy gain and utility loss. Release the privacy and utility joint evaluation.

2. The method for assessing privacy risks and utility of synthetic graph data according to claim 1, characterized in that: The private learning and synthesis phase is defined by two coupled components: a generator and a privacy implementation. The generator specifies how to generate a graph from a compact representation, while the privacy implementation defines the privacy concept and a differential privacy DP mechanism for privatizing the generation process. The generator includes a probabilistic constructor, a community- and block-based generator, and a neural graph generator. The privacy implementation relies on the Laplace mechanism, the exponential mechanism, and differential privacy stochastic gradient descent DP-SGD.

3. The method for assessing privacy risks and utility of synthetic graph data according to claim 1, characterized in that: The compact representation family in the representation phase includes degree-based statistics, adjacency-based counting, matrices, community-level summaries, and continuous node representations, i.e., node embeddings; among which degree-based statistics are used by DP-dK and DGG, adjacency-based representations are used by TmF and PrivSKG, community structures are used by PrivGraph, and node embeddings are used by PrivDPR.

4. The method for assessing privacy risks and utility of synthetic graph data according to claim 2, characterized in that: The privacy implementation enforces privacy through a corresponding privacy-enforcing (DP) mechanism. When multiple methods employ the same DP mechanism, the implementation of the DP mechanism and related privacy calculations are aligned, and the relevant privacy parameter settings are further aligned to ensure that DP remains effective and comparable across graphs at different scales. For DP-SGD-based methods, According to the number of protected units Relevant rules are set to enable The value varies with the number of protected units. The value decreases as it increases, under node-level DP. That is, the number of nodes in the original graph, under edge-level DP. That is, the number of edges in the original graph; let... This is the original image. and These are its node set and edge set, respectively. and , respectively representing the original image The number of nodes and the number of edges.

5. The method for assessing privacy risks and utility of synthetic graph data according to claim 2, characterized in that: After privacy protection is enforced, each method constructs a graph generator based on its selected compact representation and privatized output, and generates a synthetic graph from the graph generator. The method-specific strategies of the graph generator include probabilistic graph reconstruction, top edge selection, community-based graph generation, BTER-based graph generation, and embedding-based edge sampling. The generator design for each method is preserved, and when the generator construction and generation process is shared by multiple methods and future extensions, the relevant implementations are reused to support controlled comparisons.

6. The method for assessing privacy risks and utility of synthetic graph data according to claim 1, characterized in that, The privacy benefits are defined in the following ways: In the publishing environment, trusted data curators hold the original images. Apply the DPGS mechanism to M and publish the synthetic graph. After observing the release, the attacker attempts to infer sensitive information about the protected elements in the original graph using auxiliary background knowledge; Each privacy threat is modeled as attacker A, who aims to infer target elements. The secret, the target element Representing the target node and edges, the secret corresponds to the target's membership and related sensitive attributes; defining the attack advantage. Measure when the target exists in the published graph The degree to which an attacker's reasoning ability is enhanced when not present is calculated based on the difference in success rates between inclusion and exclusion attacks; the greater the attack advantage, the higher the privacy risk of the target element. The privacy benefit is achieved by releasing the composite image. Instead of the original image To reduce the part of the attack advantage, namely: 。 7. The method for assessing privacy risks and utility of synthetic graph data according to claim 6, characterized in that: When the secret is the target's membership identity, the attacker is instantiated as a member inference attacker, including node target membership inference and edge target membership inference. In the associative game between the attacker and the challenger, the attack advantage, assuming the secret bits are a priori equal, is determined by the difference between the attacker's output probability of the target's existence in the target inclusion case and the target exclusion case. This is equivalent to doubling the attacker's guess equal to the secret bit probability minus one. Since the attacker can directly check the membership identity with probability 1 when the original graph is released, i.e., the attack advantage under the original graph release is 1, the privacy gain is: ; in The attack advantage released by the composite graph; This indicates that synthetic release offers no protection against member inference. This indicates that attackers have no advantage over those who guess randomly.

8. The method for assessing privacy risks and utility of synthetic graph data according to claim 6, characterized in that, When the secret is a sensitive attribute of the target, the attacker is instantiated as an attribute inference attacker: the attacker observes part of the target record. And attempts to infer the values ​​of the retained sensitive attributes based on the released graph and auxiliary knowledge P. The challenger assigned sensitive attribute values ​​to a subset of target records based on the overall reference graph. This constitutes a complete target record. ; The attack advantage is the difference between the probability that the estimated value equals the true sensitive attribute value when the target is included and the probability that the estimated value equals the true sensitive attribute value when the target is excluded; the privacy gain is the difference between the attack advantage under the release of the original graph and the attack advantage under the release of the synthetic graph, i.e.: 。 9. The method for assessing privacy risks and utility of synthetic graph data according to claim 1, characterized in that: The utility loss is defined in a task-based manner, given a graph query task T, let Representation diagram The task output above, for the original graph and composite image The loss of utility is denoted as A smaller utility loss value indicates better utility preservation; For tasks whose output is scalar and aggregate statistics, the utility loss is instantiated using relative error: ; For tasks where the output is a distribution, vector, or other structured objects, use a task-specific difference function. definition: ; This is used when DIS is a similarity metric where a larger value is always better. Convert it into a loss so that the utility loss is always as small as possible.

10. A privacy risk and utility assessment system for synthetic graph data, referring to the privacy risk and utility assessment method for synthetic graph data as described in any one of claims 1-9, characterized in that, This includes a unified evaluation framework and modular methodology integration components; The unified evaluation framework introduces two metrics, privacy gains and utility losses, to characterize the privacy and utility behavior of releasing a synthetic graph. The privacy gains measure the degree to which releasing a synthetic graph reduces empirical privacy risks compared to releasing the original graph, and the utility losses measure the difference between the synthetic graph and the original graph under the same graph query workload. The modular method integration component integrates representative DPGS methods within a shared organizational and evaluation framework. By dividing each DPGS method into a representation phase and a private learning and synthesis phase, it localizes method differences into three parts: graph representation, privacy implementation, and graph generator construction and generation. It also aligns comparable components and evaluation conditions where applicable for controlled comparisons. Specifically, when multiple methods rely on the same representation family, the corresponding implementation components and key preprocessing choices are aligned, including degree-based methods using normalized degree histograms under the same construction rules, adjacency-based methods using consistent sparse coding, and community methods using a fixed-set Leuven algorithm. Privacy gains and utility losses were measured across different methods under the same threat model, attack instantiation, target selection rules, and utility settings. The system supports the integration of additional DPGS methods by mapping the pipeline of new methods to the same two-stage organization, and extends the system by incorporating additional attack and utility metrics through privacy gains and utility losses defined at the framework level.