A federated graph synthesis method and system based on a differential privacy model

Abstract

This invention discloses a federated graph synthesis method and system based on a differential privacy model, belonging to the field of federated learning technology. It is applied to a federated graph synthesis system comprising one server and multiple clients. The method includes a community detection stage, an information extraction stage, and a graph reconstruction stage. This method achieves global graph synthesis that satisfies edge differential privacy without centralized original sensitive edge data or cross-client interaction of original data, thus ensuring data privacy from an architectural perspective. It designs a noise-resistant edge statistics mechanism between supernodes, converting the weighted supernode graph into an unweighted undirected graph to reduce the impact of high noise on community detection in federated scenarios. It also designs a local community optimization mechanism to improve the accuracy of community structure detection. Furthermore, it proposes a fine-grained edge construction strategy, adaptively stopping client contribution value collection based on the connection strength of supernode pairs. Finally, it designs a heterogeneity-aware information extraction mechanism, allocating differentiated privacy budgets based on edge participation patterns.

Description

Technical Field

[0001] This invention relates to the field of federated learning technology, and in particular to a federated graph synthesis method and system based on a differential privacy model. Background Technology

[0002] Graph data, with its precise characterization of node relationships, has been widely applied in key areas such as social network analysis, financial fraud detection, and personalized recommendations. However, this type of data often contains a large amount of sensitive information and is frequently stored across multiple institutions or platforms in real-world scenarios (i.e., a federated scenario), making it impossible to directly share or analyze it centrally.

[0003] Differential privacy graph synthesis technology can generate synthetic graphs that retain the structural characteristics of the original graph and satisfy privacy constraints. However, in federated environments, there are still few studies on graph synthesis that rely purely on differential privacy. When existing centralized differential privacy graph synthesis methods are directly transferred to federated scenarios, they face a series of problems such as noise accumulation, inefficient budget utilization, and insufficient privacy security, which lead to a serious degradation of the utility of the synthesized graph. Specifically, these include: (1) Noise interference causes distortion of the synthesized graph: In the federated parallel aggregation process, the independent noise injected by each client will accumulate linearly with the number of clients, causing serious distortion of the final statistics. At the same time, the method still uses a weighted supernode graph that is highly sensitive to noise for community detection. The reliability of community detection is low, which affects the quality of the synthesized graph. (2) High ineffective consumption of privacy budget and low utilization rate: In the edge contribution value collection stage, the contribution values ​​of all clients are collected indiscriminately for all supernode pairs (regardless of their connection strength), resulting in a large amount of privacy budget being ineffectively consumed. In the subsequent information extraction stage, a unified privacy budget is used to protect all edges. This results in insufficient protection for high-participation edges that have consumed the budget in the early stage, while causing budget waste for low-participation edges that have not consumed the budget in the early stage. The overall budget allocation is uneven and inefficient. (3) The privacy protection mechanism is weak and there is a risk of leakage: This method simply adds noise to the client contribution value and fails to effectively filter the noisy statistics that may contain a lot of original information after aggregation. Attackers can use these statistics and the distribution of client nodes to deduce the local subgraph structure in reverse with a high probability, which poses a risk of sensitive information leakage. In addition, some existing federated graph processing methods that do not rely on differential privacy, such as those combining secure multi-party computation (MPC) or privacy set intersection (PSI), have the core defect that they cannot provide the provable strong mathematical privacy guarantee that differential privacy has. In addition, these methods usually require the introduction of complex cryptographic protocols, which may lead to significant computation and communication overhead.

[0004] Therefore, there is an urgent need for a differential privacy graph synthesis scheme adapted to federated scenarios, which can achieve high-fidelity and high-security federated differential privacy graph synthesis without centralizing the original sensitive data. Summary of the Invention

[0005] To address the aforementioned issues, this invention discloses a federated graph synthesis method and system based on a differential privacy model. (1) This method proposes a differential privacy graph synthesis framework specifically adapted to federated scenarios, achieving the desired results without centralized original sensitive edge data or cross-client interaction of original data. - Edge differential privacy global graph synthesis ensures data privacy from an architectural perspective and avoids privacy leakage risks; (2) Noise robust community detection: Design a noise-resistant supernode edge statistics mechanism to convert the weighted supernode graph into an unweighted undirected graph, reduce the impact of high noise on community detection in federated scenarios, design a local community optimization mechanism to improve the detection accuracy of community structure and ensure the core structural characteristics of the synthesized graph; (3) Adaptive and differentiated privacy budget optimization: Propose a "fine-grained edge construction strategy" to adaptively stop collecting client contribution values ​​based on the connection strength of supernode pairs to reduce ineffective consumption; and design an "information extraction mechanism that perceives heterogeneity" to allocate differentiated privacy budgets based on the participation mode of edges.

[0006] To understand this invention, the following well-known technical concepts must first be clarified:

[0007] (1) Differential privacy graph synthesis: The aim is to generate a synthetic graph that satisfies differential privacy, making its key statistical characteristics highly similar to the original graph. The synthetic graph must be able to support downstream graph analysis tasks such as graph classification, community detection, and link prediction, without requiring additional privacy budget.

[0008] (2) CL model (Zhonglu model): It is a classic graph reconstruction model. It generates intra-community edges based on the given target degree of nodes within the community, with the probability of generating intra-community edges by the ratio of the product of node degrees to the total number of edges; inter-community edges are generated by probability sampling based on the number of inter-community edges.

[0009] (3) Federated learning architecture: This is a distributed collaborative architecture whose core is "server + multiple clients". Under this architecture, the original data is always stored on the local client, and the task is completed by exchanging intermediate parameters (such as model parameters and noisy statistics), thus eliminating the direct leakage of the original data from the architecture.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] The first aspect of this invention proposes a federated graph synthesis method based on a differential privacy model, applied to a federated graph synthesis system comprising one server and multiple clients, wherein each client holds a local subgraph of the global graph. The method includes the following steps:

[0012] Community Detection Phase: The server randomly divides the global nodes into multiple coarse-grained supernodes and distributes the supernode division to each client. Each client calculates the contribution value of the number of edges within and between its local supernodes, injects first Laplace noise matching the privacy budget, and uploads it to the server. The server constructs fine-grained edges, sequentially aggregates the noisy contribution values ​​uploaded by each client, determines whether edges exist between pairs of supernodes based on an adaptive threshold pruning strategy, constructs an unweighted, undirected supernode graph, and performs community detection on this graph to obtain the initial community division. After receiving the community division results, each client optimizes the initial community division locally using its local data to obtain the final community division and uploads it to the server.

[0013] Information extraction phase: The server divides the unvisited edge set and the visited edge set based on the accumulated stopping records from the fine-grained edge construction phase, and sends the sampling probability corresponding to this state to each client; each client performs differentiated sampling of its original edges based on the sent sampling probability to obtain the edge set to be counted; each client calculates the local intra-community degree and inter-community edge number contribution value corresponding to the edge set to be counted, injects a second Laplace noise, and uploads it to the server; the server aggregates to obtain the global noisy intra-community degree sequence and the noisy inter-community edge number.

[0014] Graph Reconstruction Phase: The server reconstructs the graph based on the global noisy intra-community degree sequence and the number of edges between noisy communities. Specifically, for each community, based on the noisy intra-community degree sequence, the Chung-Lu (CL) model is used to generate intra-community edges. Specifically, the probability of an edge between nodes is determined by the ratio of the product of node degrees to the total number of edges within the community. For different communities, based on the number of edges between noisy communities, node pairs are randomly sampled from the corresponding two communities to generate inter-community edges. Finally, the intra-community edges and inter-community edges are integrated to output a global differential privacy graph.

[0015] Optionally, in the step of the client calculating the contribution values ​​of the number of internal edges of the local supernode and the number of edges between supernodes, the formula for calculating the contribution value of the number of internal edges of the supernode is as follows: The starting point between supernodes is The middle and end points are at The formula for calculating the contribution value of the number of edges in the equation is as follows: ,in, For the client's identification number, The contribution value to the number of edges for the client. supernode China belongs to the client The set of nodes, It is a node The adjacency vector, Indicates the existence of from arrive edge It is an indicator function.

[0016] Optionally, the first Laplace noise scale is After injection, the noisy contribution value is obtained. ,in, The privacy budget allocated for fine-grained edge construction during the community testing phase.

[0017] Optionally, during the fine-grained edge construction step on the server side, the server accumulates and aggregates the results in real time. Set threshold ,like Immediately stop collecting contribution values ​​for this supernode pair from subsequent clients, determine if an edge exists between the supernode pair, construct an undirected edge between the supernodes, and record the set of client nodes for which the contribution value of this supernode pair has not been accumulated; if ≤ Then continue collecting until all clients have been traversed, where the threshold is... The amplitude is set according to the Laplace noise theory.

[0018] Optionally, the local optimization specifically involves: the client calculating the number of connections between each node and each community as a quality score, and redistributing node community affiliation through an exponential mechanism; using the number of connections as the scoring function for the exponential mechanism, with a sensitivity of 1; wherein, the formula for calculating the number of connections is:

[0019] ;

[0020] in, For nodes, For the target community, It is a node The adjacency vector, This indicates that there are edges in the graph. ;

[0021] The probability calculation formula used in the exponential mechanism is:

[0022] ;

[0023] in, Represents a node With candidate communities The number of connections in a node. , Adjust the privacy budget for node community affiliation in the index mechanism.

[0024] Optionally, the differential sampling specifically involves: including all unvisited edge sets AE into the edge set to be counted AE', and for edges in the visited edge set NE, using probability... Samples are added to AE' with a sampling probability of . ,in, The total privacy budget allocated for the information extraction phase; after differential sampling, edges in AE' uniformly use a higher privacy budget. Laplace noise is added for protection, to balance the statistical accuracy of different edge sets with the privacy budget consumption.

[0025] Optionally, the step of each client calculating the contribution values ​​of local intra-community degree and inter-community edge count for the edge set to be counted, injecting a second Laplace noise, and then uploading it to the server specifically includes:

[0026] Each client performs local statistics for each node it holds. The client calculates the community to which a given edge set AE' belongs, based on the edge set to be counted after differential sampling. Within the community , For each community involved in AE' Calculate the contribution value of the number of edges between communities. Next, the client will provide services within the community. Injection scale is The noise, after perturbation, generates noisy statistics. Transmitted to the server; contribution value of edge count between communities Injection scale is The noise and disturbance are used to obtain the noise contribution value. Then, the noisy contribution value is uploaded to the server.

[0027] Optionally, the privacy budget allocated during the community testing phase Privacy budget allocated during the information extraction phase satisfy: , The total global differential privacy budget can be adjusted according to task requirements, for example, by setting... , ;and Further decomposed into fine-grained edge construction budget Optimize budgets with local communities ,satisfy During the information extraction phase, a personalized differential privacy-based design is adopted. For connections from unvisited edge sets AE, the available budget is... For a connection in the visited edge set NE, its available budget is .

[0028] A second aspect of this invention proposes a federated graph synthesis system based on a differential privacy model, used to perform the method described in the first aspect, the system comprising a server and multiple clients; wherein the server comprises:

[0029] Supernode partitioning module: used to randomly partition the global nodes into multiple coarse-grained supernodes;

[0030] Fine-grained edge construction module: used to sequentially aggregate noisy contribution values ​​uploaded by each client, determine whether there are edges between pairs of supernodes based on an adaptive threshold pruning strategy, and construct an unweighted and undirected supernode graph;

[0031] Community detection module: used to perform initial community partitioning on unweighted supernode graphs;

[0032] Heterogeneity information extraction module: It is used to divide the edge set (unvisited edge set and visited edge set) according to the cumulative stopping record of the supernode pair in the fine-grained edge construction stage, allocate differentiated privacy budgets for different connection sets, calculate the corresponding sampling probability, and send the sampling instructions and probabilities to each client;

[0033] Global statistics aggregation module: used to aggregate noisy intra-community degree and inter-community edge count uploaded by the client;

[0034] Graph Reconstruction Module: Used to generate a global differential privacy graph using a CL model and a random sampling strategy;

[0035] Each client includes:

[0036] Differential sampling module: Used to receive the sampling probability sent by the server and perform differential sampling on the corresponding edges held locally to determine the local edge set to be counted;

[0037] Local contribution calculation module: used to calculate the contribution values ​​of the number of internal edges of local supernodes, the number of edges between supernodes, the degree within the community, and the number of edges between communities;

[0038] Noise injection module: Used to inject Laplace noise into the contribution value based on the privacy budget issued by the server;

[0039] Local optimization module: Used to optimize the community affiliation of nodes based on local data through an exponential mechanism.

[0040] A third aspect of the present invention provides a terminal configured with a memory and a processor, the memory being adapted to store multiple instructions, and the processor being adapted to call the instructions in the memory to perform steps implementing the method as described in the first aspect.

[0041] The beneficial effects of this invention are that it provides a federated differential privacy graph synthesis method and system, aiming to achieve an efficient, secure, and high-fidelity federated differential privacy graph synthesis scheme in federated scenarios, specifically including:

[0042] (1) Solve the problem of linear growth of differential privacy noise in federated scenarios and improve the utility of synthetic graph structure. This invention transforms the supernode weighted graph into a noise-resistant unweighted undirected graph through client-level aggregation strategy and adaptive threshold pruning strategy, effectively offsetting the negative impact of noise as the number of clients increases and avoiding the degradation of synthetic graph structure utility.

[0043] (2) Solve the problems of inefficient use and uneven allocation of privacy budget, and reduce ineffective consumption and overprotection. This invention designs a fine-grained edge construction mechanism, which adaptively stops collecting client contribution values ​​based on the connection strength of supernodes, avoiding budget waste caused by indiscriminate full collection; at the same time, based on personalized differential privacy design, an information extraction mechanism that perceives heterogeneity is designed, which allocates differentiated budgets according to the differences in edge participation modes, making full use of the remaining budget of unvisited edges and improving the ability to preserve the structure of the synthetic graph.

[0044] (3) Addressing the privacy leakage risk of federated graph synthesis. This invention filters unreliable statistics through threshold pruning, transmitting only noisy aggregation results to avoid leakage of original data or sensitive information. It also blocks the possibility of reverse deducing local subgraphs on the client side through aggregated statistics from the architecture, thus ensuring data privacy and security in federated scenarios. Attached Figure Description

[0045] Figure 1 This is a flowchart of the federated graph synthesis method based on the differential privacy model of the present invention;

[0046] Figure 2 This is a flowchart of the fine-grained edge construction method of the present invention. Detailed Implementation

[0047] 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 a part of the embodiments of the present invention, not all of them. 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. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. 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.

[0048] Example 1

[0049] This embodiment addresses the differential privacy graph synthesis problem in a federated environment, proposing a federated graph synthesis method based on a differential privacy model. This method is applied to a federated graph synthesis system comprising one server and multiple clients (≥2), where each client... A local subgraph of the global graph, containing a set of disjoint nodes. The goal is to synthesize a set of adjacent vectors that satisfy the condition without aggregating the original edge data. - Edge-difference privacy global graph G′. The method includes the following steps, such as Figure 1 As shown:

[0050] (1) Community testing phase

[0051] a. The server randomly divides the global nodes into m coarse-grained supernodes. Each supernode contains approximately n / m nodes. This partitioning is independent of edge data, does not leak privacy, and distributes the supernode partitioning to each client.

[0052] For each pair of supernodes for which statistics need to be collected The server follows a fixed order (e.g.) The server accumulates the noisy contribution values ​​reported by each client. Specifically, the server knows which nodes each supernode contains and which clients these nodes are distributed across. When it needs to collect... and When calculating the number of edges between supernodes, the server will sequentially send requests to those nodes in the supernode sequence. The client that owns the node initiates the request.

[0053] The client calculates the contribution values ​​of the number of internal edges within local supernodes and the number of edges between supernodes, injects first Laplace noise matching the privacy budget, and then uploads it to the server: Client Upon receiving information regarding supernode pairs After the query, based on its local storage, belonging to Node set The local contribution value of the adjacency vectors of the supernode is calculated. Specifically, the formula for calculating the contribution value of the number of internal edges of the supernode is as follows: The contribution value of the number of edges between supernodes (i.e., the starting point is at...) The middle and end points are at The formula for calculating the number of sides in a given area is: ,in, For the client's identification number, The contribution value to the number of edges for the client. supernode China belongs to the client The set of nodes, It is a node The adjacency vector, Indicates the existence of from arrive edge This is an indicator function. Subsequently, the client uses the privacy budget allocated to this step. The contribution value is injected with the first Laplace noise (noise scale is ). ), to obtain the noisy contribution value ,in, The privacy budget is allocated for fine-grained edge construction during the community detection phase and uploaded to the server. The process for collecting the number of edges within supernodes is similar. The server is responsible for coordinating the entire sequential collection and aggregation process.

[0054] b. The server performs fine-grained edge construction, sequentially aggregating the noisy contribution values ​​uploaded by each client. Based on an adaptive threshold pruning strategy, it determines whether edges exist between pairs of supernodes, constructs an unweighted, undirected supernode graph, and executes the Louvain algorithm on this graph for community detection to obtain the initial community partitioning results. Each community contains several supernodes.

[0055] To address the issue of weighted graph distortion caused by the linear increase in noise with the number of clients, FGDPsyn designed a fine-grained edge construction mechanism and a local community optimization mechanism. The total budget is divided into... ( (for community testing) and ( (used for information extraction), among which Further split into (Fine-grained edge construction) and (Local community optimization) to meet the needs of During the information extraction phase, a personalized differential privacy-based design is adopted. For connections from unvisited edge sets AE, the available budget is... For a connection in the visited edge set NE, its available budget is .

[0056] The fine-grained edge construction mechanism is a core optimization in the community detection phase, aiming to address the noise accumulation problem in federated scenarios and improve the robustness of community detection. This mechanism is executed in the early stages of the community detection phase and consumes a significant amount of budget. The specific process is as follows: Figure 2 As shown.

[0057] After global supernode partitioning and initialization, the client calculates the local contribution values ​​of the number of internal edges and the number of edges between supernodes based on the local subgraph, and injects the first Laplacian noise (noise scale is...). Then upload to the server. The server accumulates and aggregates the results in real time. Set threshold (i.e., 0.5 times the theoretical amplitude of Laplace noise), if Immediately stop collecting contribution values ​​for this supernode pair from subsequent clients, determine if an edge exists between the supernode pair, construct an undirected edge between the supernodes, and record the set of client nodes for which the contribution value of this supernode pair has not been accumulated; if ≤ If the data collection fails, it continues until all clients have been traversed. This mechanism reduces invalid data collection and noise accumulation through an early stopping strategy, while constructing an unweighted supernode graph to replace the traditional weighted graph, significantly improving the noise resistance of community detection.

[0058] c. To improve the accuracy of community segmentation, each client, after receiving the community segmentation results, utilizes the remaining privacy budget. traverse each node in the graph The initial community division is optimized locally to obtain the final community division result. And upload it to the server; specifically, for each node The client uses the locally stored adjacency vector of the node. Calculate its division with the community Each community The number of connections is used as a quality score, which is then used through an exponential mechanism to... Reassign community affiliation; use the number of connections as the scoring function of the exponential mechanism, with a sensitivity of 1, where the formula for calculating the number of connections is:

[0059] ;

[0060] in, For nodes, For the target community, It is a node The adjacency vector, This indicates that there are edges in the graph. ;

[0061] The probability calculation formula used in the exponential mechanism is:

[0062] ;

[0063] in, Represents a node With candidate communities The number of connections in a node. , Adjust the privacy budget for node community affiliation in the index mechanism.

[0064] The optimized community affiliation results are uploaded to the server, which then integrates all client results to obtain the final community division. This ensures that the community structure more closely resembles the true characteristics of the original map.

[0065] (2) Information extraction stage: This stage aims to extract information based on the final community division results. The algorithm collects the intra-community node degree and inter-community edge count for graph reconstruction. Due to the heterogeneity of edge participation patterns caused by the fine-grained edge construction mechanism, this embodiment employs a heterogeneity-aware information extraction mechanism. This mechanism is the core optimization in the information extraction phase, as the fine-grained edge construction mechanism leads to heterogeneous edge set participation patterns. Based on the accumulation stop record in the fine-grained edge construction phase, the server classifies edges associated with subsequent client nodes that have not triggered accumulation into the unvisited edge set AE, and the remaining accumulated edges into the visited edge set NE. Some edges (unvisited edge set AE) are not consumed. Budget, some edges (visited edge set NE) have been consumed. Budget. To fully utilize the privacy budget, the server divides the edge set into unvisited edge sets (AEs, with remaining budget). ) and visited edge set (NE, remaining budget is ), and allocate different privacy budgets to different edge sets; based on the heterogeneity partitioning results of the edge sets, the client performs differential sampling on the unvisited edge set and the visited edge set, and merges them to obtain the edge set to be counted; the differential sampling is as follows: the unvisited edge set AE is fully included in the edge set to be counted AE'; for the edges in the visited edge set NE, with probability Samples are added to AE' with a sampling probability of . ,in, The privacy budget allocated for fine-grained edge computing during the community testing phase. The total privacy budget allocated for the information extraction phase; after differential sampling, edges in AE' uniformly use a higher privacy budget. Laplace noise is added for protection, balancing the statistical accuracy of different edge sets with the privacy budget. After edge set classification and differential sampling are completed, the server will output the final community partitioning results. and will And the issuance of statistical instructions.

[0066] Each client calculates the contribution values ​​of the local intra-community degree and inter-community edge count for the edge set to be counted, injects a second Laplace noise, and then uploads it to the server; specifically including:

[0067] Each client performs local statistics for each node it holds. The client calculates the community to which a user belongs based on the edge set AE' to be counted, which is constructed after differential sampling. Within the community , For each community involved in AE' Calculate the contribution value of the number of edges between communities. Next, the client will provide services within the community. Injection scale is The noise, after perturbation, generates noisy statistics. Transmitted to the server; contribution value of edge count between communities Injection scale is The noise and disturbance are used to obtain the noise contribution value. Then, the noisy contribution value is uploaded to the server.

[0068] The server performs global aggregation on the client to obtain the global noisy community degree sequence and the number of edges between noisy communities, which serve as the core input for graph reconstruction.

[0069] (3) Graph reconstruction stage: The server generates intra-community edges through the CL model based on the global noisy intra-community degree sequence and the number of inter-community edges through the noisy inter-community edge count, generates inter-community edges through the random sampling strategy, integrates intra-community edges and inter-community edges, and outputs a global differential privacy graph.

[0070] The FGDPsyn solution's overall privacy budget is divided into two parts: Assigned to the community testing phase, Assigned to the information extraction stage. Strictly adheres to differential privacy principles regarding sequence combination, parallel combination, and post-processing invariance.

[0071] During the community testing phase, privacy budget Consumed. Specifically, the fine-grained edge construction process consumes... In this process, for the supernode pairs that are ultimately determined to have "edges" (i.e., connections that do not belong to the unvisited edge set AE), their statistics are collected from the client and scald with a metric. The Laplace noise satisfies - Differential privacy. For connections in the unvisited edge set AE, they were not visited during the community detection phase, satisfying 0-differential privacy. Subsequently, in the local community optimization step, each node redistributes its community through an exponential mechanism, a step that consumes budget. and satisfy - Differential privacy. According to the sequence combination theorem of differential privacy, the overall community detection phase satisfies ( Differential privacy, i.e. - Differential privacy.

[0072] In the information extraction phase, a set of edges to be counted, AE', is constructed through differential sampling: initially, AE' = AE (all unvisited edges are included), and then edges in the visited edge set NE are counted with probability. Sampling is added to AE'. All edges in AE' use a privacy budget uniformly. Laplace noise is added. This design ensures that the unvisited edge set AE fully utilizes its unconsumed budget from the previous stage, while the visited edge set NE incurs additional privacy overhead through sampling control, achieving overall satisfaction. - Security requirements under differential privacy.

[0073] The graph reconstruction phase only performs post-processing on the statistics output from the first two phases that are already protected by differential privacy, without consuming additional privacy budget.

[0074] In summary, according to the sequence combination theorem, the entire FGDPsyn process satisfies... - Differential privacy.

[0075] Example 2

[0076] A federated graph synthesis system based on a differential privacy model is provided for performing the method described in Example 1. The system includes a server and multiple clients; wherein the server includes:

[0077] Supernode partitioning module: used to randomly partition the global nodes into multiple coarse-grained supernodes;

[0078] Fine-grained edge construction module: used to sequentially aggregate noisy contribution values ​​uploaded by each client, determine whether there are edges between pairs of supernodes based on an adaptive threshold pruning strategy, and construct an unweighted and undirected supernode graph;

[0079] Community detection module: used to perform initial community partitioning on unweighted supernode graphs;

[0080] Heterogeneity information extraction module: It is used to divide the edge set (unvisited edge set and visited edge set) according to the cumulative stopping record of the supernode pair in the fine-grained edge construction stage, allocate differentiated privacy budgets for different connection sets, calculate the corresponding sampling probability, and send the sampling instructions and probabilities to each client;

[0081] Global statistics aggregation module: used to aggregate noisy intra-community degree and inter-community edge count uploaded by the client;

[0082] Graph Reconstruction Module: Used to generate a global differential privacy graph using a CL model and a random sampling strategy;

[0083] Each client includes:

[0084] Differential sampling module: Used to receive the sampling probability sent by the server and perform differential sampling on the corresponding edges held locally to determine the local edge set to be counted;

[0085] Local contribution calculation module: used to calculate the contribution values ​​of the number of internal edges of local supernodes, the number of edges between supernodes, the degree within the community, and the number of edges between communities;

[0086] Noise injection module: Used to inject Laplace noise into the contribution value based on the privacy budget issued by the server;

[0087] Local optimization module: Used to optimize the community affiliation of nodes based on local data through an exponential mechanism.

[0088] Example 3

[0089] A terminal is provided, comprising a memory and a processor. The memory is adapted to store multiple instructions, and the processor is adapted to invoke the instructions in the memory to execute the steps of the method described in Embodiment 1. The terminal includes, but is not limited to, terminal devices such as servers, mobile phones, computers, and tablet computers.

[0090] Specifically, in this embodiment of the invention, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0091] It should also be understood that the memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0092] Application examples

[0093] Taking a social network scattered across multiple data silos as an example, a global graph needs to be synthesized for downstream model training. This invention's method (FGDPsyn) is compared with a centralized, federated direct-adaptation version (Fed-PrivGraph). In the Fed-PrivGraph method, due to the use of a weighted supernode graph and indiscriminate full parallel collection, Laplacian noise accumulates linearly with the number of clients, severely damaging the structured features of the synthesized graph.

[0094] This invention significantly reduces inefficient privacy budget consumption and effectively resists high noise in federated scenarios by transforming unweighted, undirected supernode graphs and implementing an early stopping mechanism for fine-grained edge construction. Experiments verify that, under the same global differential privacy budget (e.g., ... Under the given conditions, compared to the Fed-PrivGraph baseline, the synthetic graph generated by this invention reduces the degree KL-Divergence by approximately 50% and improves the normalized mutual information (NMI) index for community discovery by nearly 50%, significantly enhancing the high fidelity and data availability of the synthetic graph in downstream applications such as node influence assessment and community network analysis.

[0095] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A federated graph synthesis method based on a differential privacy model, characterized in that, A federated graph synthesis system comprising one server and multiple clients, each client holding a local subgraph of the global graph, the method includes the following steps: Community Detection Phase: The server randomly divides the global nodes into multiple coarse-grained supernodes and distributes the supernode division to each client. Each client calculates the contribution values ​​of the number of edges within and between its local supernodes, injects first Laplace noise matching the privacy budget, and uploads it to the server. The server constructs fine-grained edges, sequentially aggregates the noisy contribution values ​​uploaded by each client, determines whether edges exist between pairs of supernodes based on an adaptive threshold pruning strategy, constructs an unweighted, undirected supernode graph, and performs community detection on this graph to obtain the initial community division. After receiving the community division results, each client optimizes the initial community division locally using its local data to obtain the final community division and uploads it to the server. Information extraction phase: The server divides the unvisited edge set and the visited edge set based on the accumulated stopping records from the fine-grained edge construction phase, and sends the corresponding sampling probabilities to each client; each client performs differentiated sampling of its original edges according to the sampling probabilities locally to obtain the edge set to be counted, calculates the contribution values ​​of the local intra-community degree and inter-community edge count corresponding to the edge set to be counted, injects second Laplace noise, and uploads it to the server; the server aggregates to obtain the global noisy intra-community degree sequence and the noisy inter-community edge count. Graph reconstruction phase: The server generates intra-community edges based on the global noisy intra-community degree sequence and the number of noisy inter-community edges using the CL model, generates inter-community edges using a random sampling strategy, integrates intra-community edges and inter-community edges, and outputs a global differential privacy graph.

2. The federated graph synthesis method based on a differential privacy model as described in claim 1, characterized in that, In the steps of calculating the contribution values ​​of the number of internal edges of local supernodes and the number of edges between supernodes on the client side, the formula for calculating the contribution value of the number of internal edges of supernodes is as follows: The starting point between supernodes is The middle and end points are at The formula for calculating the contribution value of the number of edges in the equation is as follows: ,in, For the client's identification number, The contribution value to the number of edges for the client. supernode China belongs to the client The set of nodes, It is a node The adjacency vector, Indicates the existence of from arrive edge It is an indicator function.

3. The federated graph synthesis method based on a differential privacy model as described in claim 1, characterized in that, The first Laplace noise scale is After injection, the noisy contribution value is obtained. ,in, The privacy budget allocated for fine-grained edge construction during the community testing phase.

4. The federated graph synthesis method based on a differential privacy model as described in claim 1, characterized in that, In the server-side fine-grained edge construction step, the adaptive threshold pruning strategy specifically involves the server-side real-time accumulation of aggregation results. Set threshold ,like Immediately stop collecting contribution values ​​for this supernode pair from subsequent clients, determine if an edge exists between the supernode pair, construct an undirected edge between the supernodes, and record the set of client nodes for which the contribution value of this supernode pair has not been accumulated; if ≤ Then continue collecting until all clients have been traversed, where the threshold is... The amplitude is set according to the Laplace noise theory.

5. The federated graph synthesis method based on a differential privacy model as described in claim 1, characterized in that, The local optimization specifically involves: the client calculating the number of connections between each node and each community as a quality score, and then redistributing node community affiliation through an exponential mechanism; the number of connections is used as the scoring function for the exponential mechanism, where the formula for calculating the number of connections is: ； in, For nodes, For the target community, It is a node The adjacency vector, This indicates that there are edges in the graph. ; The probability calculation formula used in the exponential mechanism is: ； in, Represents a node With candidate communities The number of connections in a node. , Adjust the privacy budget for node community affiliation in the index mechanism.

6. The federated graph synthesis method based on a differential privacy model as described in claim 1, characterized in that, Differential sampling specifically involves: including all unvisited edge sets AE into the edge set AE' to be counted; and for edges in the visited edge set NE, using probability... Samples are added to AE' with a sampling probability of . ,in, Total privacy budget allocated for the information extraction phase.

7. The federated graph synthesis method based on a differential privacy model as described in claim 1, characterized in that, The steps for each client to calculate the contribution values ​​of the local intra-community degree and inter-community edge count for the edge set to be counted, inject a second Laplace noise, and then upload the results to the server include: Each client performs local statistics for each node it holds. The client calculates the community to which a user belongs based on the edge set AE' to be counted, which is constructed after differential sampling. Within the community , For each community involved in AE' Calculate the contribution value of the number of edges between communities. Next, the client will provide services within the community. Injection scale is The noise, after perturbation, generates noisy statistics. Transmitted to the server; contribution value of edge count between communities Injection scale is The noise and disturbance are used to obtain the noise contribution value. Then, the noisy contribution value is uploaded to the server.

8. The federated graph synthesis method based on a differential privacy model as described in claim 1, characterized in that, Privacy budget allocated during the community testing phase Privacy budget allocated during the information extraction phase satisfy: ,in, The total budget for global differential privacy; and Further decomposed into fine-grained edge construction budget Optimize budgets with local communities ,satisfy During the information extraction phase, a personalized differential privacy-based design is adopted. For connections from unvisited edge sets AE, the available budget is... For a connection in the visited edge set NE, its available budget is .

9. A federated graph synthesis system based on a differential privacy model, characterized in that, The system is used to perform the method according to any one of claims 1 to 8, the system comprising a server and a plurality of clients; wherein the server comprises: Supernode partitioning module: used to randomly partition the global nodes into multiple coarse-grained supernodes; Fine-grained edge construction module: used to sequentially aggregate noisy contribution values ​​uploaded by each client, determine whether there are edges between pairs of supernodes based on an adaptive threshold pruning strategy, and construct an unweighted and undirected supernode graph; Community detection module: used to perform initial community partitioning on unweighted supernode graphs; Heterogeneity information extraction module: used to divide the edge set according to the accumulated stopping records of the supernode pair in the fine-grained edge construction stage, allocate differentiated privacy budgets, calculate the corresponding sampling probabilities, and send the sampling instructions and probabilities to each client; Global statistics aggregation module: used to aggregate noisy intra-community degree and inter-community edge count uploaded by the client; Graph Reconstruction Module: Used to generate a global differential privacy graph using a CL model and a random sampling strategy; Each client includes: Differential sampling module: Used to receive the sampling probability sent by the server and perform differential sampling on the corresponding edges held locally to determine the local edge set to be counted; Local contribution calculation module: used to calculate the contribution values ​​of the number of internal edges of local supernodes, the number of edges between supernodes, the degree within the community, and the number of edges between communities; Noise injection module: Used to inject Laplace noise into the contribution value based on the privacy budget issued by the server; Local optimization module: Used to optimize the community affiliation of nodes based on local data through an exponential mechanism.

10. A terminal, characterized in that, The device is equipped with a memory and a processor, the memory being adapted to store multiple instructions, and the processor being adapted to invoke the instructions in the memory to perform the steps of implementing the method as described in any one of claims 1 to 8.

Images