A method and apparatus for classifying privacy nodes on the edge of a communication network
By generating a public mask set and mask view, constraining the number of views exposed for each possible communication edge, and performing interval pruning and multi-view mean smoothing on the original classification score, a selective classification vector is constructed by combining the node classification interval and stability score. Laplace noise is added for global privacy determination, which solves the problems of unstable risk classification results and privacy leakage of intranet devices in the prior art. It achieves the protection of communication topology privacy while ensuring the stability and availability of device risk classification results.
Patent Information
- Application Number
- CN202611579618.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-13
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies struggle to protect the communication topology privacy of enterprise or industrial intranets while ensuring the stability and availability of device risk classification results, and there is a risk of privacy leakage due to changes in a single communication edge.
By generating a public mask set and masked views, the number of views exposed for each possible communication edge is constrained. The original classification scores are then subjected to interval pruning and multi-view mean smoothing. Combined with node classification interval and stability score, a selective classification vector is constructed, and Laplace noise is added for global privacy determination.
It achieves the stability and availability of device risk classification results while protecting the privacy of intranet communication topology, reduces the risk of privacy leakage, and adapts to the security operation and maintenance needs of enterprise and industrial intranets.
Smart Images

Figure CN123074848A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of communication network security, graph data processing, artificial intelligence and differential privacy technology, and in particular to a communication network edge privacy node classification method and apparatus for risk detection of intranet device nodes in enterprise / industrial intranets. Background Technology
[0002] In enterprise and industrial intranet network security operation and maintenance scenarios, a large number of network entities such as servers, office terminals, gateways, and IoT devices are deployed within the network. Risk identification and anomaly detection of various intranet devices are required to support security handling operations such as alarms, isolation, and security auditing. In communication network security management, network entities such as hosts, servers, terminals, routing devices, IoT devices, containers, virtual machines, or virtual network instances can typically be represented as nodes, and communication relationships such as connections, flows, or sessions can be represented as communication edges, thereby constructing a communication network graph reflecting the network's operational status. Graph Neural Networks (GNNs), by aggregating node features and communication relationships, can identify the operational status, security status, or risk level of nodes and have been widely used in intranet device risk identification scenarios such as network anomaly detection and malicious host identification. In actual deployment, the model is usually trained on historical communication graphs and the model parameters are frozen before inference is performed on the current sensitive inference communication graph, outputting the classification score or category label of each node. Since the current communication edge reflects the real communication sessions and service access relationships of intranet devices, it is sensitive topology data that needs to be strictly protected within the intranet. The published classification scores or category labels depend on these communication relationships, which may indirectly leak sensitive communication edge information and pose a potential threat to the network security of enterprise intranets and industrial networks.
[0003] Existing graph data privacy protection technologies include Differential Privacy (DP) training, graph structure perturbation, secure multi-party computation, and neighbor sampling, each targeting training data, input graph structure, computation process, or adjacency information used in a single inference iteration. DP training protects training data by introducing noise during the training phase, but struggles to constrain the frozen model's dependence on current communication graph edges during inference. Graph structure perturbation hides the true topology by randomly adding, deleting, or replacing graph structures, but this can corrupt graph structure information and affect the accuracy of intranet device risk classification. Secure multi-party computation focuses on protecting the multi-party collaborative computation process but does not protect against indirect dependencies of inference output on communication edges. Neighbor sampling reduces edge information exposure by limiting adjacency information used in a single inference iteration, but it cannot provide a unified privacy guarantee for the complete classification vector. However, none of these methods address the following technical problem: In enterprise or industrial intranet operations, when freezing a graph neural network to classify intranet device nodes using a sensitive inference communication graph, the classification vector may change with the addition or deletion of individual communication edges. Attackers can exploit this change to infer the true communication session relationships between intranet hosts or network devices. In existing processing methods, adding output noise node by node will affect the risk classification results of the device and bring about the privacy budget combination problem of multiple nodes; directly publishing the complete classification vector lacks stability judgment, making it difficult to combine the internal network security business requirements and uniformly decide whether to publish the risk identification results based on the stability of the classification vector to unilateral changes.
[0004] In summary, existing graph data privacy protection technologies struggle to provide edge differential privacy protection for complete classification vectors while protecting all inference communication edges within the intranet. Furthermore, they lack a unified stability assessment and release mechanism adapted to intranet security services, making it difficult to balance communication edge privacy protection with the availability of intranet device risk classification results. Summary of the Invention
[0005] This invention provides a method and apparatus for classifying privacy nodes on the edge of a communication network. It addresses the technical problem that existing technologies in enterprise / industrial intranet device risk identification services struggle to balance privacy protection on the edge of intranet communication with the usability of device risk classification results, cannot suppress the impact of changes in a single intranet communication session on the classification output, and risk leaking the true communication relationships of intranet devices in node risk classification results.
[0006] The method provided in this embodiment of the invention is adapted to the device risk identification service of enterprise and industrial intranets, and is used to complete the risk classification of intranet device nodes while protecting the privacy of intranet communication topology.
[0007] This invention provides a method for classifying privacy nodes on the edge of a communication network, executed by an electronic device, including:
[0008] The set of nodes in the sensitive inference communication graph constitutes the complete set of possible communication edges. A view number arrangement sequence is generated for each possible communication edge in the complete set of possible communication edges. The set of view numbers to which a single possible communication edge belongs is determined according to the preset edge exposure limit. Based on the possible communication edges corresponding to each view number, multiple public mask sets and public mask families are constructed.
[0009] Based on the intersection of the public mask set and the real communication edge set of the sensitive inference communication graph, a corresponding mask view is generated. All mask views and node features are input into the graph neural network with frozen parameters to obtain the original classification score of each node. The original classification scores are uniformly cropped to a set interval. The arithmetic mean of the cropped classification scores of the same node and the same category under multiple mask views is calculated to obtain the average classification score.
[0010] Traverse each node in the fixed release node set, determine the provisional category corresponding to the maximum average classification score of each node according to the preset category parallel processing rules, obtain the node classification interval based on the difference between the maximum average classification score and the second largest average classification score, set the to-be-released result of each node as the corresponding category label or rejection symbol according to the relationship between the node classification interval and the classification interval selection threshold, combine the to-be-released results of all nodes in the fixed release node set to obtain the to-be-released selective classification vector; discretize the calculation at intervals of twice the upper bound of the score sensitivity based on the difference between the node classification interval and the classification interval selection threshold to obtain the non-negative integer node stability score;
[0011] The selective classification vector to be published is used as the global proposed publishing result. The node stability score with the smallest value in the fixed publishing node set is selected as the query stability statistic. Laplace privacy noise of a preset scale is added to the query stability statistic to obtain the noise-added statistic. The global test is determined to be passed or failed based on the relationship between the noise-added statistic and the preset global test threshold.
[0012] If the global test passes, the selective classification vector to be published is output to the network security management device for node status management or security handling; if the global test fails, a full rejection output vector with a length equal to the total number of nodes in the fixed set of published nodes is output.
[0013] The device provided in this embodiment of the invention is adapted to the device risk identification service of enterprise and industrial intranets, and is used to complete the risk classification of intranet device nodes while protecting the privacy of intranet communication topology.
[0014] This invention provides a communication network edge privacy node classification device, executed by an electronic device, comprising:
[0015] The construction unit is used to construct a complete set of possible communication edges based on the node set of the sensitive inference communication graph, generate a view number arrangement sequence for each possible communication edge in the complete set of possible communication edges, determine the view number set to which a single possible communication edge belongs based on a preset edge exposure upper limit, and construct multiple public mask sets and public mask families based on the possible communication edges corresponding to each view number.
[0016] The first unit is used to generate a corresponding mask view based on the intersection of the public mask set and the real communication edge set of the sensitive inference communication graph. All mask views and node features are input into the parameter-frozen graph neural network to obtain the original classification score of each node. The original classification scores are uniformly cropped to a set interval. The arithmetic mean of the cropped classification scores of the same node and the same category under multiple mask views is calculated to obtain the average classification score.
[0017] The second obtaining unit is used to traverse each node in the fixed publishing node set, determine the provisional category corresponding to the maximum average classification score of each node according to the preset category parallel processing rules, obtain the node classification interval based on the difference between the maximum average classification score and the second largest average classification score, set the to-be-published result of each node as the corresponding category label or rejection symbol according to the relationship between the node classification interval and the classification interval selection threshold, combine the to-be-published results of all nodes in the fixed publishing node set to obtain the to-be-published selective classification vector; and perform discretization calculation with an interval of twice the upper bound of the score sensitivity based on the difference between the node classification interval and the classification interval selection threshold to obtain the non-negative integer node stability score.
[0018] The judgment unit is used to take the selective classification vector to be published as the global proposed publishing result, select the node stability score with the smallest value in the fixed publishing node set as the query stability statistic, add Laplace privacy noise of a preset scale to the query stability statistic to obtain a noisy statistic, and determine whether the global test passes or fails based on the relationship between the noisy statistic and the preset global test threshold.
[0019] The output unit is used to output the selective classification vector to be published to the network security management device if the global test passes, for node status management or security handling; if the global test fails, it outputs a full rejection output vector with a length equal to the total number of nodes in the fixed set of published nodes.
[0020] This invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform any of the above-described methods for classifying privacy nodes on the edge of a communication network.
[0021] This invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs any of the above-described methods for classifying privacy nodes on the edge of a communication network.
[0022] This invention provides a method and apparatus for classifying privacy nodes on the edge of a communication network. The method constrains the number of views exposed for each possible communication edge by publishing a mask set, limits the range of mask views to which a single possible communication edge belongs by combining a preset edge exposure upper limit, and performs interval pruning and multi-view mean smoothing on the original classification scores output by each mask view. This strictly constrains the maximum impact of the addition or subtraction of a single sensitive inference communication edge on the average classification score within the upper bound of score sensitivity, enabling quantifiable and provable control over internal network communication topology disturbances and ensuring the stable operation of risk classification services for internal network devices. The method is based on the node's maximum and second-largest average classification scores. The difference between the values is used to construct the node classification interval. A threshold and an upper bound for score sensitivity are selected based on the classification interval, and the node stability score, representing the edge editing radius, is discretized and calculated. The minimum value of the stability scores of all nodes within a fixed set of publishing nodes is selected as the query stability statistic, accurately representing the topological anti-disturbance stability of the overall selective classification vector to be published. This provides a quantitative and calculable reliable basis for the global privacy publishing decision to adapt to internal network security services. A unified publishing decision is achieved based on the privacy publishing mechanism. The overall decision is made to output either the selective classification vector to be published or a vector of complete rejection, based on the comparison between the noisy query stability statistic and the global test threshold. This method adds Laplace differential privacy noise only to the scalar query stability statistics, eliminating the need to add noise to the output of each node individually or to split and stack the privacy budget of multiple nodes. This significantly reduces the deployment difficulty of intranet security services while ensuring the integrity of global privacy protection. The method is compatible with pre-trained graph neural networks with frozen parameters for inference operations. During the training phase, communication topology data does not participate in the construction of sensitive inference mask views, thus avoiding the risk of sensitive topology information leakage at the source. The final standardized vector output can be directly used by network security management devices, supporting them in performing intranet security tasks such as node status management, risk alarms, topology auditing, and risk isolation. It is suitable for actual security operation and maintenance scenarios in enterprise intranets and industrial networks, demonstrating strong engineering practicality. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0024] Figure 1 This is a schematic flowchart of a method for classifying privacy nodes on the edge of a communication network, provided by an embodiment of the present invention.
[0025] Figure 2 This is a schematic diagram of a communication network edge privacy node classification device provided in an embodiment of the present invention. Detailed Implementation
[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] The method described in this invention is adaptable to network security operation and maintenance scenarios such as enterprise intranets and industrial networks, and is used to classify and identify risks in network devices such as servers, office terminals, gateways, and IoT devices within the intranet. Network devices are abstracted as nodes, and the actual communication sessions between devices are abstracted as sensitive inference communication edges. While protecting the privacy of the intranet communication topology, the method outputs node risk classification results, which are then delivered to network security management equipment to support security handling functions such as alarms, isolation, and security auditing.
[0028] Figure 1 This is a flowchart illustrating a method for classifying privacy nodes on the edge of a communication network according to an embodiment of the present invention. The following is in conjunction with... Figure 1 This invention provides a detailed description of a method for classifying privacy nodes on the edge of a communication network, based on embodiments of the present invention. For example... Figure 1 As shown, the method includes the following steps:
[0029] Step 101: Based on the node set of the sensitive inference communication graph, construct a complete set of possible communication edges. Generate a view number arrangement sequence for each possible communication edge in the complete set of possible communication edges. Determine the view number set to which a single possible communication edge belongs based on the preset edge exposure limit. Based on the possible communication edges corresponding to each view number, construct multiple public mask sets and public mask families.
[0030] Step 102: Generate a corresponding mask view based on the intersection of the public mask set and the real communication edge set of the sensitive inference communication graph. Input all mask views and node features into the parameter-frozen graph neural network to obtain the original classification score of each node. Crop the original classification scores to a set range. Calculate the arithmetic mean of the cropped classification scores of the same node and the same category under multiple mask views to obtain the average classification score.
[0031] Step 103: Traverse each node in the fixed publishing node set, determine the provisional category corresponding to the maximum average classification score of each node according to the preset category parallel processing rules, obtain the node classification interval based on the difference between the maximum average classification score and the second largest average classification score, set the to-be-published result of each node as the corresponding category label or rejection symbol according to the relationship between the node classification interval and the classification interval selection threshold, combine the to-be-published results of all nodes in the fixed publishing node set to obtain the to-be-published selective classification vector; discretize the calculation based on the difference between the node classification interval and the classification interval selection threshold, with twice the upper bound of the score sensitivity as the interval, to obtain the non-negative integer node stability score;
[0032] Step 104: Take the selective classification vector to be published as the global proposed publishing result, select the node stability score with the smallest value in the fixed publishing node set as the query stability statistic, add Laplace privacy noise of a preset scale to the query stability statistic to obtain the noise-added statistic, and determine whether the global test passes or fails based on the relationship between the noise-added statistic and the preset global test threshold.
[0033] Step 105: If the global test passes, the selective classification vector to be published is output to the network security management device for node status management or security handling; if the global test fails, a full rejection output vector with a length equal to the total number of nodes in the fixed publishing node set is output.
[0034] The subject of this communication network edge privacy node classification method is an electronic device with data processing and network security computing capabilities, specifically a network electronic device with computing power such as a network security management device, a server, or an industrial control processing terminal.
[0035] Step 101 is the first step of the method provided in this embodiment of the invention, used to construct a family of public masks decoupled from the set of real communication edges of the sensitive inference communication graph, and further generate mask views to prepare for subsequent graph neural network inference based on multiple sets of mask views.
[0036] The input data and parameters used in this embodiment of the invention are all obtained in advance or offline before the method is executed, and are not calculated or generated within the method. This includes: a sensitive inference communication graph. Node feature matrix Graph Neural Networks with Frozen Parameters Fixed release node set Total number of masked views Preset edge exposure limit The constraints satisfy... The parameters mentioned above are explained in detail below:
[0037] Specifically, the sensitive inference communication graph is collected from the network security system, and includes the following:
[0038] (1)
[0039] in, Represents a sensitive reasoning communication graph. The set of nodes representing the sensitive reasoning communication graph. This represents the set of real communication edges in the sensitive reasoning communication graph. In practical applications, network entities are abstracted as nodes to obtain the set of nodes in the sensitive reasoning communication graph, and network communication interaction records are abstracted as communication edges to obtain the set of real communication edges in the sensitive reasoning communication graph. The set of real communication edges in the sensitive reasoning communication graph belongs to the protected privacy input.
[0040] Node feature matrix The node feature matrix is pre-calculated based on statistical information such as network traffic, ports, protocols, and device attributes. Each row corresponds to a node's multidimensional statistical features, which are used for graph neural network inference. This is publicly available information and does not rely on sensitive reasoning. It represents the set of real communication edges in the communication graph.
[0041] Graph Neural Networks with Frozen Parameters This was trained offline: training was performed on an independent training communication graph and training node label dataset, and all network model parameters were frozen after training. In this embodiment of the invention, the network model parameters are not updated during the entire inference process; the training communication graph and the sensitive inference communication graph share the same set of nodes in the sensitive inference communication graph. The edges of the training communication graph do not participate in the inference operations of the subsequent masked view.
[0042] Fixed release node set This is a non-empty set, predetermined by publicly available information before privacy classification, eliminating the need to read the actual communication edge set of the sensitive inference communication graph; the set of publishing nodes is fixed. It is the set of nodes in the sensitive reasoning communication graph. A subset of represents the nodes from which classification results need to be output in subsequent embodiments of the present invention. It can also be set to... That is, all nodes are selected as the nodes to be output.
[0043] Total number of masked views Preset edge exposure limit These are pre-defined public hyperparameters, which can be set manually or obtained through parameter tuning using a verification set; total number of mask views. This represents the set of public masks to be generated and the total number of mask views; preset edge exposure limit. Used to constrain the maximum number of possible communication edges that can be included. A set of public masking codes.
[0044] It should be noted that: node feature matrix Frozen graph neural network model parameters, fixed set of publishing nodes And the public mask family, which remains unchanged between edge-adjacent inputs where only a single communication edge differs; the so-called edge-adjacent input is a pair of sensitive inference communication graphs defined by edge differential privacy, which are composed of two sensitive inference communication graphs; the set of nodes of the two sensitive inference communication graphs. They are completely identical, with the only difference between their actual communication edge sets being one communication edge. That is, by adding or deleting a communication edge on one of the sensitive inference communication graphs, another sensitive inference communication graph can be obtained. This is used to simulate the input disturbance scenario of adding or disconnecting a single communication session in the intranet, where the only difference is the addition or removal of a single communication edge.
[0045] In step 101, a complete set of possible communication edges is constructed based on the node set of the sensitive inference communication graph, and its expression is as follows:
[0046] (2)
[0047] in, Indicates the complete set of possible communication edges. Indicates that by node With nodes The unordered pairs of nodes formed The set of nodes representing the sensitive reasoning communication graph.
[0048] In this embodiment of the invention, the complete set of possible communication edges consists of all pairs of distinct unordered node pairs in the node set of the sensitive inference communication graph; each unordered node pair is called a possible communication edge, which only represents the potential possibility of communication between two nodes and does not represent that the edge corresponding to the node pair actually exists in the sensitive inference communication graph; the process of constructing the complete set of possible communication edges only uses the node set of the sensitive inference communication graph and does not read the actual communication edge set of the sensitive inference communication graph.
[0049] In step 101, a pre-set, fixed public random seed is also required. The public random seed is a public quantity that remains unchanged between adjacent inputs. The same input paired with the same public random seed can produce a definite and reproducible output result.
[0050] For each potentially communicative edge in the complete set of potentially communicative edges, perform the following processing:
[0051] (1) Generate the normalized endpoint identifier of the possible communication edge: sort the two endpoints of the possible communication edge according to the pre-agreed fixed sorting rules and splice them to obtain a unique normalized endpoint identifier; for the same possible communication edge, regardless of the writing order of the endpoints, the same normalized endpoint identifier is obtained. The normalized endpoint identifier belongs to public information and does not read the real communication edge set of the sensitive reasoning communication graph.
[0052] (2) Input the normalized endpoint identifiers and the public random seed into the public deterministic permutation function to generate The view numbering sequence.
[0053] (3) Select the first view number in the sequence. Each view number constitutes the set of view numbers to which this possible communication edge belongs. Naturally satisfies that each possible communication edge can be assigned to at most one of the following groups: Constraints on a set of public masks.
[0054] After traversing all possible communication edges and completing the allocation, construct the first one one by one. A set of public masks , ;
[0055] Public mask set For the complete set of possible communication edges A subset; if a possible communication edge The set of view numbers to which it belongs Includes view number Then the possible communication edge is classified into the first... A set of public masks .
[0056] All By combining a set of public masks, we obtain a family of public masks. .
[0057] In this embodiment of the invention, the entire generation process of the mask family is disclosed, relying only on the complete set of possible communication edges, the normalized endpoint identifiers of each possible communication edge, and a publicly disclosed random seed, without reading or relying on the actual set of communication edges in the sensitive inference communication graph. When two sets of inputs are edge-adjacent inputs, the set of nodes in the sensitive inference communication graph... When the public random seed remains unchanged, the resulting public mask families are completely identical.
[0058] The construction logic of the aforementioned public mask family has the following practical significance in intranet business scenarios: it does not rely on real intranet communication session data to generate the mask structure, thus avoiding the leakage of communication relationships between intranet devices during the mask construction process. When a single communication session is added or disconnected on the intranet, the mask system remains unchanged, ensuring the stability of the subsequent risk classification and reasoning process.
[0059] In step 102, firstly, for each set of public masks obtained in step 101, the intersection of the set of public masks and the set of real communication edges of the sensitive inference communication graph is calculated. The communication edges that belong to both the set of public masks and the set of real communication edges of the sensitive inference communication graph are retained to form the corresponding mask view. The mask view is shown below:
[0060] (3)
[0061] in, The set of nodes representing the sensitive reasoning communication graph. This represents the set of real communication edges in a sensitive reasoning communication graph. Indicates the first A set of public masking codes Indicates the first A masked view, Indicates the mask view number. , Indicates the total number of masked views.
[0062] It should be noted that the public mask set is a predetermined public structure and does not depend on the set of real communication edges in the sensitive inference communication graph. The set of real communication edges in the sensitive inference communication graph is the entire set of actual communication edges that exist in the current sensitive inference communication graph to be inferred. This set is generated by actual network communication behavior (real connections, sessions, traffic interaction records) and is the sensitive topology data that needs privacy protection and prevention of leakage in this invention. This step takes the intersection of the public mask set and the set of real communication edges, and uses the publicly fixed mask structure to filter the private real communication edges. This allows for real business inference while limiting the influence of a single sensitive communication edge on the inference result.
[0063] Then, all masked views and node feature matrices are input into a pre-trained, parameter-frozen graph neural network to perform inference operations, obtaining the K-dimensional raw classification score for each node, where... Throughout the inference process, the parameters of the graph neural network model remain fixed, without parameter updates or retraining. The offline-trained model can be used directly for online inference.
[0064] After obtaining the K-dimensional original classification scores, a uniform interval clipping process is performed on each dimension of the original classification scores to clip the original classification scores of each dimension to a fixed interval. , The cropping classification score is obtained. Interval cropping can suppress extreme abnormal scores under a single view, avoid the score disturbance caused by extreme values amplifying a single sensitive communication edge, and provide a prerequisite for obtaining a quantifiable upper bound on the score sensitivity.
[0065] After completing the interval cropping process of the original classification scores, the cropped classification scores corresponding to each masked view can be obtained. The classification scores after cropping a single masked view satisfy the following calculation formula:
[0066] (4)
[0067] in, Indicates the first The classification score matrix after cropping a masked view This represents the interval pruning operator, used to constrain the model's original inference score to a fixed interval. Inside, This indicates the upper and lower bounds for cropping the category score. , The parameter is Parameter-frozen graph neural networks Represents network model parameters, Indicates the first A masked view, This represents the node feature matrix.
[0068] To eliminate random perturbation errors in single-view reasoning and improve the overall stability of classification scores, the following measures are taken: The arithmetic mean of the cropping classification scores of the same node and the same category under each masked view is used to obtain the average classification score. The average classification score satisfies the following calculation formula:
[0069] (5)
[0070] in, This represents the average classification score matrix for multiple views. Indicates the total number of mask views. This represents the summation operator. This indicates the view traversal sequence number. By balancing the mean of multiple views, extreme score deviations in a single view can be effectively suppressed, providing a stable calculation basis for the subsequent fluctuation range of the quantified score sensitivity.
[0071] Based on the constraints imposed by the preset edge exposure limit in step 101, and combined with the score pruning and multi-view averaging processing mechanism, the impact range of a single sensitive communication edge disturbance on the classification results can be strictly quantified.
[0072] In this embodiment of the invention, for any real number The clipping operation is defined as:
[0073] (6)
[0074] in, This represents the standard score clipping operator. This indicates the upper and lower bounds for cropping the category score. This represents the real number representing the category score to be clipped. This represents the operator for finding the maximum value.
[0075] When a single sensitive inference communication edge is added or deleted, the maximum change in the score value corresponding to any node and any category in the average classification score matrix does not exceed the upper bound of the score sensitivity. The upper bound of the score sensitivity is determined by the following formula:
[0076] (7)
[0077] in, This indicates the preset edge exposure limit. This indicates the upper bound of the score sensitivity. Indicates the total number of mask views. This indicates the upper and lower bounds for cropping the category score.
[0078] It should be noted that the upper bound of the score sensitivity is determined only by the three publicly available hyperparameters mentioned above, and is independent of the model structure of the parameter-frozen graph neural network, the network model parameters, and the topology of the sensitive inference communication graph. This enables the controllability and quantification of sensitive topological perturbations in the parameter-frozen graph neural network inference scenario, providing a rigorous and provable mathematical foundation for subsequent node classification interval calculation, node stability score quantification, and global differential privacy verification.
[0079] In step 103, each node in the fixed publishing node set is traversed, and subsequent operations are performed only on nodes belonging to that set; the fixed publishing node set is determined in advance by public information and remains unchanged between edge-adjacent inputs.
[0080] Specifically, for the current node Read the average classification score of all categories for this node output in step 102, and determine the provisional category corresponding to the node with the highest average classification score according to the preset category tie processing rules. The details are as follows:
[0081] (8)
[0082] in, Represents a node Provisional category with the highest average classification score Represents a collection of category indexes. This represents the total number of categories in the model classification. , Indicates the category traversal index. Represents a node category The corresponding average classification score, This represents the maximum value index operator.
[0083] It should be noted that the default category tie handling rule is a pre-defined public rule used to handle the tie case where the average classification scores of multiple categories of the same node are tied and the maximum value is taken. This rule remains unchanged between adjacent edge inputs.
[0084] Next, calculate the nodes. Node classification interval The node classification margin is the difference between the node's maximum average classification score and the maximum average classification score of all other categories excluding the provisional category. It characterizes the confidence level of the node's classification result; a larger classification margin indicates higher discriminative power. The formula for calculating the node classification margin is as follows:
[0085] (9)
[0086] in, Represents a node The node classification interval, Represents a node category The corresponding average classification score, This indicates the category traversal index.
[0087] Select a threshold between the node classification interval and the preset classification interval. The comparison is performed to generate a single node's result to be published. The preset classification interval selection threshold is greater than zero. If a node's classification interval is greater than or equal to the classification interval selection threshold, then... Then set the provisional category This serves as the category label to be published for that node; if the node's category interval is less than the category interval selection threshold, i.e. If so, the result to be published will be set to a rejection symbol.
[0088] In this embodiment of the invention, the result to be published by the current node is determined by the following formula:
[0089] (11)
[0090] in, The symbol represents rejection. Represents a node The results to be released Represents a node Corresponding category tags, Indicates the node classification interval. This indicates the threshold for selecting the classification interval. .
[0091] Furthermore, the fixed set of publishing nodes The results to be published from all nodes within the node are combined to obtain the selective classification vector to be published. The details are as follows:
[0092] (12)
[0093] in, This represents the selective classification vector to be published. node The results to be released Represents a fixed set of publishing nodes. Indicates a fixed set of publishing nodes All nodes in the pending results A vector formed sequentially by node number.
[0094] It should be noted that the selective classification vector to be released can contain both category labels and rejection symbols. This vector is only an intermediate result to be released and has not yet been output to the public. Whether this vector can be officially output to the public will be determined based on the global privacy check in the subsequent steps.
[0095] Furthermore, based on the difference between the node classification interval and the selected threshold, a sensitivity upper bound of twice the score is applied. The discretization step size is used for discretization calculations to obtain non-negative integer node stability scores. In this embodiment, the node stability score represents the edge editing radius that ensures the node's published result remains unchanged. That is, if edge editing occurs in the communication graph and the edge editing distance does not exceed the node stability score, the node's published result will not change.
[0096] The node stability score is determined by the following formula:
[0097] (13)
[0098] in, Represents a node The node stability score, This indicates the upper bound of the score sensitivity. This indicates the floor operator. This represents the floor operator. Represents a fixed set of publishing nodes. This represents the selective classification vector to be published. Indicates the node classification interval. This indicates the threshold for selecting the classification interval.
[0099] In the intranet device risk identification service, the node stability score represents whether the risk assessment result of the device can remain unchanged when a certain number of communication sessions change in the intranet communication topology. The higher the stability score, the stronger the ability of the risk label output by the device to resist topology disturbances and the lower the risk of privacy leakage.
[0100] In step 104, this step does not modify the publishing results of a single node. Instead, it uses the proposed publishing results of the entire fixed publishing node set as the verification object. By calculating the set query stability statistic, adding Laplace privacy noise, and comparing thresholds, it completes the decision of whether to publish or reject globally in one go. This constrains the privacy leakage risk caused by topology disturbances of sensitive inference communication edges from an overall perspective. The global test judgment result is then output to step 105 for the final output operation.
[0101] Specifically, the selective classification vector to be published generated in step 103 is used as the global proposed publication result. This proposed publication result contains the content to be published of all nodes in the fixed set of publication nodes. The content can be category labels or rejection symbols. No local modifications are made to the proposed publication result during the verification process.
[0102] Extracting a fixed set of publishing nodes The minimum node stability score among all nodes in the set is selected as the query stability statistic. Using the minimum value as the statistic represents using the weakest node within the set as the global stability benchmark. This ensures that all nodes in the set satisfy the privacy stability constraints and avoids privacy leaks caused by the instability of individual nodes. The query stability statistic is determined by the following formula:
[0103] (14)
[0104] in, This indicates a query stability statistic. Represents a node The node stability score, Indicates the node sequence number. Represents a fixed set of publishing nodes. This represents the minimum value operator.
[0105] Furthermore, by pre-setting a fixed public hyperparameter noise scale, Laplace privacy noise with a scale of λ is added to the original noiseless query stability statistic to generate a noisy statistic, as shown below:
[0106] (15)
[0107] in, This represents the noise statistics. This indicates a query stability statistic. λ represents the Laplace privacy noise sample value, i.e., the specific value obtained by sampling from the Laplace privacy noise distribution, and λ represents the scale hyperparameter of the Laplace noise, which is pre-configured by the service.
[0108] It should be noted that adding Laplace privacy noise is a differential privacy perturbation technique. By masking the sensitive topological features carried by the original statistics with noise, it reduces the possibility of attackers inferring sensitive inference communication edges and achieves differential privacy protection at the set level.
[0109] In this embodiment of the invention, the noise statistics are numerically compared with a preset global test threshold to complete the one-time privacy release mechanism determination. The privacy release mechanism only makes a unified decision on the complete proposed release results for a fixed set of release nodes, and does not individually modify or reject the proposed release results of a particular node. The output result of the privacy release mechanism is determined by the following formula:
[0110] (16)
[0111] in, Represents a set of nodes The corresponding privacy release mechanism output results, This represents the selective classification vector to be published, i.e., the global proposed publication result. This represents the Laplace privacy noise sample value. This indicates the preset global test threshold. This represents the differential privacy parameter. Represents the natural logarithm operator. This represents the total number of nodes in a fixed set of publishing nodes. This indicates a total rejection output vector.
[0112] In this embodiment, when the noise statistics are greater than the preset global test threshold, the global test is deemed to have passed, the overall privacy risk of the proposed results is controllable, and the release can be executed; when the noise statistics are less than or equal to the preset global test threshold, the global test is deemed to have failed, the proposed results have a potential privacy leak risk, the release of valid classification results is prohibited, and a full rejection output vector is output.
[0113] In step 105, when the global test passes, it indicates that the entire set of selective classification vectors to be released meets the edge difference privacy and security conditions, and the risk of privacy leakage is controllable. The complete selective classification vectors to be released are then output to the network security management device as is. After receiving the vector, the network security management device can perform node status management or security handling services such as node status display, risk alarm, risk device isolation, and communication security auditing based on the vector content.
[0114] When the global test fails, it indicates that the current proposed result is insufficient to resist topology edge disturbances, and direct output will pose a privacy risk of leaking sensitive inference communication edge topology information. In this case, the selective classification vector to be published is not used, and a full rejection output vector is generated and output. The length of the full rejection output vector is equal to the total number of nodes in the fixed publishing node set, and each element of the vector is a rejection symbol. After receiving the full rejection output vector, the network security management device discards all classification information obtained from this inference and does not perform automatic security actions such as alarms or isolation based on the inference result.
[0115] Therefore, the method provided by the embodiments of the present invention does not simply output mathematical vectors. The output results directly serve the internal network security operation and maintenance: when the global test is passed, the vectors can be used to carry out risk alarms and handling of internal network devices; when the global test fails, all device risk identification results in this round are abandoned to avoid inferring internal network device communication sessions indirectly due to risk labels.
[0116] Example
[0117] This embodiment applies to a scenario of enterprise intranet terminal risk detection. The enterprise intranet deploys a total of 12 devices, including business servers, office terminals, gateways, and IoT devices. Communication session records are generated between these devices. The network security system collects the connection and session records between these devices to construct a sensitive inference communication graph. The set of nodes in the sensitive reasoning communication graph. Includes all 12 intranet devices, and a set of real communication edges in the sensitive inference communication graph. Records of the actual communication relationships between devices constitute sensitive topology data that requires strict protection.
[0118] The security operations and maintenance requirements are: to classify and identify the risks of 8 core business devices on the internal network, determining whether the devices are in a normal or abnormal risk state, while ensuring that the classification results cannot be used to infer the actual communication connections between the devices. Traditional methods using graph neural network inference result in unbounded fluctuations in classification results as communication connections change. These changes in output can be used to infer device communication relationships, leading to potential topology privacy breaches. Therefore, this method is adopted to achieve node risk classification while protecting privacy.
[0119] Pre-configure public hyperparameters: Total number of mask views Preset edge exposure limit Category score cropping upper and lower bounds Classification interval selection threshold Differential privacy parameters , Calculate the upper bound of the score sensitivity: Calculate the preset global test threshold: .
[0120] From the node set of the sensitive reasoning communication graph Eight core business devices were selected to form a fixed release node set. Privacy classification reasoning is performed only for devices within this set.
[0121] Execute step 101, relying on the node set A complete set of possible communication edges is constructed, and a public mask family is generated based on a public random seed, further yielding 10 sets of mask views. The public mask family generation process does not read the actual internal network communication records, satisfying quantity constraints, decoupling constraints, and invariance constraints, ensuring that the public mask family will not change when any communication session is added or removed from the internal network.
[0122] Proceed to step 102, freeze the input parameters of the graph neural network, including the mask view and the node feature matrix. It consists of service statistics such as traffic, port, and protocol data for each device. After inference, score trimming, and averaging across multiple views, an average classification score matrix is obtained. This is subject to an upper bound on score sensitivity. The constraint ensures that when any communication session is added or disconnected on the intranet, the average classification score fluctuation of any device and any risk category will not exceed 0.6, thus achieving quantitative constraint on topology disturbance.
[0123] Proceed to step 103, targeting the fixed set of publishing nodes. Eight core devices are used for computation, categorized by type. These represent the normal state and the abnormal risk state of the equipment, respectively. Based on the average classification score, the provisional category and node classification interval for each piece of equipment are calculated sequentially; then, combined with... ,by Using the discretization step size, the node stability score of each device is calculated. The node stability scores obtained for the 8 devices are as follows: Threshold selection based on classification interval. Perform selective output determination and assemble to obtain the selective classification vector to be published. In the vector, some devices output normal / abnormal risk labels, while devices with insufficient confidence output rejection symbols.
[0124] Step 104: Take the minimum stability score of the 8 device nodes to obtain the query stability statistic. Add Laplace privacy noise to this statistic to obtain a noisy statistic. Set up two possible actual business scenarios:
[0125] Operating Condition 1: The calculated noise statistic is 2.40, which meets the requirements. The global test passed.
[0126] Operating Condition 2: The calculated result of the noise statistics is 1.10, which satisfies the condition. The global test failed.
[0127] Step 105, execute the business output:
[0128] When in working condition one, after the global test passes, the selective classification vector to be released is output to the network security management device. The operation and maintenance platform reads the vector content and performs alarms, link audits, and risk isolation on devices that output abnormal risk labels; for devices that output rejection symbols, it indicates that the confidence of this inference is insufficient, and no risk action is triggered for that device, only log recording.
[0129] When in condition two, if the global test fails, output a full rejection vector of length 8. Upon receiving the full rejection output vector, the network security management device directly invalidates the classification inference results of all 8 core devices in this round, without generating any alarms, isolation, or other automatic handling actions, thus avoiding the leakage of internal network communication topology information due to the direct output of classification results.
[0130] Implementation Conclusions in this Scenario: In the scenario of risk detection for core devices within an enterprise intranet, this method can classify devices as normal or abnormal. By relying on the decoupled design of the public mask family, it avoids introducing sensitive communication session information during the mask construction stage. By relying on the upper bound of score sensitivity, it controls the fluctuation of classification scores caused by changes in a single communication session to within 0.6. It retains the risk identification capability of high-confidence devices through node-level selective output. Finally, it achieves overall release or overall rejection through set-level global differential privacy verification. This method not only meets the business needs of enterprise intranet security operations for risk identification of core business devices, but also prevents attackers from inferring the real communication session relationships between intranet devices through classification outputs, solving the topology privacy leakage problem caused by direct inference from traditional graph neural networks. Furthermore, the entire process uses offline-trained graph neural networks with frozen parameters, eliminating the need for model retraining. It can be directly integrated with existing enterprise intranet security operations platforms, demonstrating strong practicality for engineering implementation.
[0131] In summary, this invention provides a method for classifying edge privacy nodes in a communication network. This method constructs a public mask family that relies solely on publicly available information such as the node set and public random seed of the sensitive inference communication graph, without reading the actual communication edge set of the sensitive inference communication graph. The public mask family remains unchanged under edge adjacency input conditions. A preset edge exposure upper limit constrains the number of possible communication edges that can be included in the public mask set, thus avoiding topological information leakage during the mask generation stage. A provable upper bound for score sensitivity is obtained by inferring from the mask view, pruning classification scores, and calculating the arithmetic mean of the pruned classification scores under multiple views. This limits the fluctuation of the average classification score caused by the addition or removal of a single sensitive inference communication edge to a fixed range. This upper bound for score sensitivity is determined only by publicly available hyperparameters such as the total number of mask views, the preset edge exposure upper limit, and the upper and lower bounds of the classification score pruning. It eliminates the need to retrain the already deployed graph neural network with frozen parameters, making it highly practical for engineering applications. The method utilizes node classification interval quantification to achieve node-level selective output, satisfying the requirements of... Nodes with a classification interval threshold are output with provisional categories as labels for the categories to be published. If the threshold is not met, a rejection symbol is output, balancing individual node privacy protection with the usability of inference results. The ability of each node's published results to resist topology edge editing disturbances is quantified by node stability scores. The minimum value of the node stability scores within a fixed set of publishing nodes is selected as the query stability statistic. Laplace privacy noise is introduced to obtain a noisy statistic to complete set-level global differential privacy verification. A decision of overall publishing or overall rejection is adopted to achieve full-domain privacy protection, avoiding privacy leaks caused by local weak nodes. Based on the global test judgment results, selective classification vectors or full rejection output vectors are output differently, with the output format adaptable to the security handling business logic of network security management devices, facilitating integration with existing security operation and maintenance systems. All hyperparameters, constraints, and calculation formulas are clearly and completely defined, and the execution process is reproducible. It can be applied to enterprise intranets and can also be extended to use scenarios such as industrial networks with high requirements for topology privacy protection.
[0132] Based on the same inventive concept, this invention provides a communication network edge privacy node classification device. Since the principle of this device in solving the technical problem is similar to that of a communication network edge privacy node classification method, the implementation of this device can refer to the implementation of the method, and the repeated parts will not be described again.
[0133] This device is used to implement the aforementioned intranet device risk identification method and can be deployed in a server or network security management device. For example... Figure 2 As shown, the device includes a construction unit 201, a first obtaining unit 202, a second obtaining unit 203, a judgment unit 204, and an output unit 205.
[0134] Construction unit 201 is used to construct a complete set of possible communication edges based on the node set of the sensitive inference communication graph, generate a view number arrangement sequence for each possible communication edge in the complete set of possible communication edges, determine the view number set to which a single possible communication edge belongs based on a preset edge exposure upper limit, and construct multiple public mask sets and public mask families based on the possible communication edges corresponding to each view number.
[0135] The first obtaining unit 202 is used to generate a corresponding mask view based on the intersection of the public mask set and the real communication edge set of the sensitive inference communication graph. All mask views and node features are input into the parameter-frozen graph neural network to obtain the original classification score of each node. The original classification scores are uniformly cropped to a set interval. The arithmetic mean of the cropped classification scores of the same node and the same category under multiple mask views is calculated to obtain the average classification score.
[0136] The second obtaining unit 203 is used to traverse each node in the fixed publishing node set, determine the provisional category corresponding to the maximum average classification score of each node according to the preset category parallel processing rules, obtain the node classification interval based on the difference between the maximum average classification score and the second largest average classification score, set the to-be-published result of each node as the corresponding category label or rejection symbol according to the relationship between the node classification interval and the classification interval selection threshold, combine the to-be-published results of all nodes in the fixed publishing node set to obtain the to-be-published selective classification vector; and perform discretization calculation with an interval of twice the upper bound of the score sensitivity based on the difference between the node classification interval and the classification interval selection threshold to obtain the non-negative integer node stability score.
[0137] The judgment unit 204 is used to take the selective classification vector to be published as the global proposed publishing result, select the smallest node stability score in the fixed publishing node set as the query stability statistic, add Laplace privacy noise of a preset scale to the query stability statistic to obtain the noise-added statistic, and determine whether the global test passes or fails based on the relationship between the noise-added statistic and the preset global test threshold.
[0138] The output unit 205 is used to output the selective classification vector to be published to the network security management device if the global test passes, for node status management or security handling; if the global test fails, it outputs a full rejection output vector with a length equal to the total number of nodes in the fixed publishing node set.
[0139] It should be understood that the units included in the above-described communication network edge privacy node classification device are merely a logical division based on the functions implemented by the device. In practical applications, the above units can be superimposed or split. Furthermore, the functions implemented by the communication network edge privacy node classification device provided in this embodiment correspond one-to-one with the communication network edge privacy node classification method provided in the above-described embodiment. The more detailed processing flow implemented by this device has been described in detail in the above-described method embodiment one, and will not be described in detail here.
[0140] Another embodiment of the present invention provides a computer device, the computer device including: a processor and a scene database; the scene database is used to store computer program code, the computer program code including computer instructions; when the processor executes the computer instructions, the electronic device performs each step of the communication network edge privacy node classification method shown in the above method embodiment.
[0141] Another embodiment of the present invention provides a computer-readable storage medium storing computer instructions that, when executed on a computer device, cause the computer device to perform the various steps of the communication network edge privacy node classification method shown in the above method embodiment.
[0142] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for classifying privacy nodes on the edge of a communication network, executed by an electronic device, characterized in that, include: The set of nodes in the sensitive inference communication graph constitutes the complete set of possible communication edges. A view number arrangement sequence is generated for each possible communication edge in the complete set of possible communication edges. The set of view numbers to which a single possible communication edge belongs is determined according to the preset edge exposure limit. Based on the possible communication edges corresponding to each view number, multiple public mask sets and public mask families are constructed. Based on the intersection of the public mask set and the real communication edge set of the sensitive inference communication graph, a corresponding mask view is generated. All mask views and node features are input into the graph neural network with frozen parameters to obtain the original classification score of each node. The original classification scores are uniformly cropped to a set interval. The arithmetic mean of the cropped classification scores of the same node and the same category under multiple mask views is calculated to obtain the average classification score. Traverse each node in the fixed release node set, determine the provisional category corresponding to the maximum average classification score of each node according to the preset category parallel processing rules, obtain the node classification interval based on the difference between the maximum average classification score and the second largest average classification score, set the to-be-released result of each node as the corresponding category label or rejection symbol according to the relationship between the node classification interval and the classification interval selection threshold, combine the to-be-released results of all nodes in the fixed release node set to obtain the to-be-released selective classification vector; discretize the calculation at intervals of twice the upper bound of the score sensitivity based on the difference between the node classification interval and the classification interval selection threshold to obtain the non-negative integer node stability score; The selective classification vector to be published is used as the global proposed publishing result. The node stability score with the smallest value in the fixed publishing node set is selected as the query stability statistic. Laplace privacy noise of a preset scale is added to the query stability statistic to obtain the noise-added statistic. The global test is determined to be passed or failed based on the relationship between the noise-added statistic and the preset global test threshold. If the global test passes, the selective classification vector to be released will be output to the network security management device for node status management or security handling. If the global test fails, the output will be a rejection vector equal to the total number of nodes in the fixed set of publishing nodes.
2. The method as described in claim 1, characterized in that, The construction of multiple sets of public masks and families of public masks specifically includes: Each possible communication edge can be assigned to at most r sets of public masks; The generation process of the public mask family does not read the set of real communication edges of the sensitive inference communication graph; Two sensitive inference communication graphs with the same set of nodes and a single differing communication edge are selected as edge adjacency inputs. The public mask families generated by these edge adjacency inputs are completely identical. This indicates the preset upper limit for edge exposure.
3. The method as described in claim 1, characterized in that, The complete set of possible communication edges is determined by the following formula: ; The masked view is determined by the following formula: ; in, Indicates the complete set of possible communication edges. Indicates that by node With nodes The unordered pairs of nodes formed Let each represent any two nodes in the sensitive reasoning communication graph. The set of nodes representing the sensitive reasoning communication graph. This represents the set of real communication edges in a sensitive reasoning communication graph. Indicates the first A set of public masking codes Indicates the first A masked view, Indicates the mask view number. , Indicates the total number of masked views. This represents the intersection operator of sets.
4. The method as described in claim 1, characterized in that, The original classification score is determined using the following formula: ; The average classification score is determined by the following formula: ; The upper bound of the score sensitivity is determined by the following formula: ; in, Indicates the first The classification score matrix after cropping a masked view This represents the interval clipping operator. This indicates the upper and lower bounds for cropping the category score. , The parameter is Parameter-frozen graph neural networks Represents network model parameters, Indicates the first A masked view, Represents the node feature matrix, This represents the average classification score matrix for multiple views. Indicates the total number of mask views. Indicates the view traversal sequence number. This indicates the preset edge exposure limit. This indicates the upper bound of the score sensitivity. This represents the summation operator.
5. The method as described in claim 1, characterized in that, The provisional category corresponding to the maximum average classification score of each node is determined by the following formula: ; The node classification interval is determined by the following formula: ; in, Represents a node Provisional category with the highest average classification score Represents a collection of category indexes. This represents the total number of categories in the model classification. , Indicates the category traversal index. Represents a node category The corresponding average classification score, This represents the maximum value index operator. This represents the maximum value operator. Represents a node The node classification interval, Indicates the node sequence number.
6. The method as described in claim 1, characterized in that, The results to be published at each node are determined using the following formula: ; The selective classification vector to be published is determined by the following formula: ; The node stability score is determined by the following formula: ; in, The symbol represents rejection. Represents a node The results to be released Represents a node Corresponding category tags, Indicates the node classification interval. This indicates the threshold for selecting the classification interval. , Represents a node The node stability score, This indicates the upper bound of the score sensitivity. This indicates the floor operator. This represents the floor operator. Represents a fixed set of publishing nodes. This represents the selective classification vector to be published. Indicates the node sequence number. Represents a fixed set of publishing nodes All nodes in the list of pending results A vector formed sequentially by node number.
7. The method as described in claim 1, characterized in that, The proposed global release result is determined using the following formula: ; The global test is determined by the following formula: ; in, This represents the query stability statistic. Represents a node The node stability score, Indicates the node sequence number. Represents a fixed set of publishing nodes. This represents the minimum value operator. Represents a set of nodes The corresponding privacy release mechanism output results, This represents the selective classification vector to be published, i.e., the edge-adjacent input to the global proposed publication result. This represents the Laplace privacy noise sample value. This indicates the preset global test threshold. This represents the differential privacy parameter. Represents the natural logarithm operator. This represents the total number of nodes in a fixed set of publishing nodes. This indicates a total rejection output vector.
8. A privacy node classification device for a communication network, executed by an electronic device, characterized in that, include: The construction unit is used to construct a complete set of possible communication edges based on the node set of the sensitive inference communication graph, generate a view number arrangement sequence for each possible communication edge in the complete set of possible communication edges, determine the view number set to which a single possible communication edge belongs based on a preset edge exposure upper limit, and construct multiple public mask sets and public mask families based on the possible communication edges corresponding to each view number. The first unit is used to generate a corresponding mask view based on the intersection of the public mask set and the real communication edge set of the sensitive inference communication graph. All mask views and node features are input into the parameter-frozen graph neural network to obtain the original classification score of each node. The original classification scores are uniformly cropped to a set interval. The arithmetic mean of the cropped classification scores of the same node and the same category under multiple mask views is calculated to obtain the average classification score. The second obtaining unit is used to traverse each node in the fixed publishing node set, determine the provisional category corresponding to the maximum average classification score of each node according to the preset category parallel processing rules, obtain the node classification interval based on the difference between the maximum average classification score and the second largest average classification score, set the to-be-published result of each node as the corresponding category label or rejection symbol according to the relationship between the node classification interval and the classification interval selection threshold, combine the to-be-published results of all nodes in the fixed publishing node set to obtain the to-be-published selective classification vector; and perform discretization calculation with an interval of twice the upper bound of the score sensitivity based on the difference between the node classification interval and the classification interval selection threshold to obtain the non-negative integer node stability score. The judgment unit is used to take the selective classification vector to be published as the global proposed publishing result, select the node stability score with the smallest value in the fixed publishing node set as the query stability statistic, add Laplace privacy noise of a preset scale to the query stability statistic to obtain a noisy statistic, and determine whether the global test passes or fails based on the relationship between the noisy statistic and the preset global test threshold. The output unit is used to output the selective classification vector to be released to the network security management device if the global test passes, for node status management or security handling. If the global test fails, the output will be a rejection vector equal to the total number of nodes in the fixed set of publishing nodes.
9. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, causes the processor to perform a method for classifying privacy nodes on the edge of a communication network as described in any one of claims 1 to 7.
10. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform a method for classifying edge privacy nodes in a communication network as described in any one of claims 1 to 7.
