A method for detecting network content security risks, an electronic device, and a storage medium.

By using a non-learning multimodal simplex representation method, the simplex representation of multimodal network content is automatically extracted, which solves the problems of insufficient fusion of multi-source heterogeneous information and difficulty in capturing high-order correlation patterns in existing technologies, and improves the accuracy and stability of network content security risk detection.

CN123074853APending Publication Date: 2026-11-10CENT SOUTH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611584365.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing methods for detecting network content security risks are unable to fully integrate multi-source heterogeneous information, cannot effectively capture high-order correlation patterns, and rely on manually defined high-order relationships to make it difficult to automatically mine the complex correlation structure of network content.

Method used

Based on the non-learning multimodal simplex representation method, the simplex representation of multimodal network content is automatically extracted through random mapping and Hodge structure enhancement operators. 0-simplex, 1-simplex and 2-simplex features are constructed, and graph-level risk representation is generated by combining the recursive local sensitive hash generation mechanism.

Benefits of technology

It achieves unified risk characterization of multimodal network content, effectively identifies risky content in complex network environments, improves detection accuracy and stability, has strong adaptability, and is suitable for large-scale network content security risk detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074853A_ABST
    Figure CN123074853A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of data processing technology, specifically providing a method, electronic device, and storage medium for detecting network content security risks. The method includes: acquiring a multimodal network content risk graph; preprocessing the multimodal network content risk graph to obtain a set of k-simplexes; obtaining multi-order features of the simplexes; propagating the multi-order features of the simplexes to obtain a graph-level risk representation; constructing a classification detection model to obtain network content security detection results. This method uses a multimodal association graph composed of user information, content information, propagation relationships, and multimodal attribute information in a network content security scenario as input. By constructing a non-learning multimodal simplex representation, it maps complex network content security data into graph-level risk features of a unified dimension, and achieves the network content security risk detection task based on the graph-level representation combined with a traditional classification model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of data processing technology and relates to a network content security risk detection method, electronic device and storage medium based on non-learning multimodal simplex representation. Background Technology

[0002] In applications such as network content security supervision, risk content identification, violation information detection, multimodal content review, and online public opinion analysis, data can typically be represented as a multimodal network structure containing heterogeneous information from multiple sources, including text, images, and videos, as well as complex relationships between users, content, and dissemination behaviors. Entity nodes in network content not only possess multimodal semantic attributes such as text, images, and videos, but also exhibit various interactive relationships, including publishing, commenting, forwarding, liking, quoting, and related dissemination, as well as high-order group association patterns formed by multiple users, content, and topics. Therefore, in network content security risk detection tasks, it is necessary to start from the input network content association structure and node multimodal attributes to construct a unified risk representation that can simultaneously describe content semantic information, user interaction relationships, and high-order dissemination structure patterns. Based on this representation, network content risk category identification and abnormal content detection can be achieved.

[0003] However, existing methods for detecting network content security risks still have certain limitations. These include: First, traditional content security detection methods typically rely on single-modal information for feature learning, making it difficult to fully integrate multi-source heterogeneous information within network content. Second, existing graph structure modeling methods generally use ordinary graph structures composed of nodes and edges for information propagation, focusing primarily on pairwise relationships between users, content, and propagating entities. This fails to effectively capture high-order association patterns formed by multiple users, content, and topics in cyberspace. For example, during network propagation, multiple users spreading the same risky content, multiple pieces of content forming related propagation chains around the same event, and multiple users jointly participating in abnormal behavior patterns all demonstrate complex group interaction structures. Relying solely on ordinary node and edge relationships for modeling leads to a lack of crucial high-order propagation information in risk detection representations.

[0004] Secondly, while some existing methods attempt to model group relationships in network content using hypergraphs, hypergraphs typically rely on manually defined or prior-constructed high-order relationships, such as pre-specified user groups, interest sets, or propagation communities. This makes it difficult to automatically uncover the high-order risk propagation patterns implicit in the original network content's relational structure. In contrast, the simplex method can automatically extract nodes, edges, and closed interaction structures based on the original network relational graph structure and describe the relationships between different orders in a unified topological space. This makes it more suitable for modeling high-order risk patterns in complex network content security scenarios.

[0005] However, introducing simplexes into network content security risk detection tasks still presents the following challenges: On the one hand, how to construct simplex representations of different orders containing semantic information from multimodal content information such as text, images, and videos without training parameters; on the other hand, how to effectively integrate content semantic information, user interaction structures, and higher-order propagation relationships, and further generate graph-level risk features suitable for risk classification and detection tasks. Summary of the Invention

[0006] This invention proposes a network content security risk detection method, electronic device, and storage medium based on non-learning multimodal simplex representation, for realizing the task of multimodal risk content identification in complex network environments. The method first uses the network content security multimodal association graph structure and node multimodal attributes as unified input. Different modal features are transformed into a unified semantic space through random mapping, and user nodes, content nodes, interaction relationships, and high-order closure structures are automatically extracted to construct simplex representations of different orders. Then, the Hodge structure enhancement operator and multimodal semantic consistency modulation mechanism are used to achieve joint propagation of structural and semantic information between simplexes of different orders. Finally, a graph-level risk representation is obtained through a recursive local sensitive hash (LSH) generation mechanism, and combined with a classification model to complete the network content security risk category detection, thus forming a complete processing flow from multimodal network content input, simplex construction, non-learning representation generation to risk classification output.

[0007] This invention provides a method for detecting network content security risks, comprising the following steps: Step 1: Obtain a multimodal network content risk map ;in, It is a set of entity nodes in the network content association graph, including user nodes, content nodes, topic nodes, and other related entities; It is a set of relationships between nodes, including posting relationships, commenting relationships, liking relationships, forwarding relationships, citation relationships, and propagation relationships; Risk map of multimodal network content Preprocessing is performed to obtain the k-simplex set. ; Based on k-simplex sets We obtain the multi-order features of the simplex; the multi-order features of the simplex include 0-simplex features. 1-Simplified Features and 2-simplex characteristics ; Step 2: Set the number of propagation rounds to [number]. The number of layers in a single round of propagation is After propagating the multi-order features of the simplex, a graph-level risk representation is obtained. ; Step 3: Utilize graph-level risk representation Construct a classification and detection model, and apply it to the multimodal network content risk map. The corresponding risk characteristics are used to identify risk categories and obtain network content security detection results. .

[0008] Furthermore, we obtain the set of k-simplexes. The specific process is as follows: Construct 0-simplex and 1-simplex, and traverse the set of nodes in the network content security association graph. Add all entity nodes as 0-simplexes to the set. Each node can correspond to a user, content, topic, or other network entity. At the same time, traverse the set of relations Add all relationships between entities as 1-simplexes to the set. ; Based on the locally closed interaction structure in the original network content security association graph, node combinations that satisfy the pairwise association conditions between three entities are retrieved and added to the set as 2-simplexes. .

[0009] Furthermore, the specific process for obtaining the multi-order features of the simplex is as follows: Design a multimodal network content risk graph any node in The corresponding multimodal properties are ;in, The number of modes; For the first The node feature matrix corresponding to each mode; For any modality in the network content security multimodal graph, , The number of modalities in the multimodal graph for network content security; Set of k-simplexes and any node The corresponding multimodal attributes As input, we obtain the 0-simplex features respectively. 1-Simplified Features and 2-simplex characteristics ;in, For nodes The corresponding 0-simplex features are used to describe the multimodal semantic information of entities such as users, content, and topics; For the edge The corresponding 1-simplex feature is used to describe local risk association patterns in user interaction, content association, and propagation relationships; Triangular simplex The corresponding 2-simplex feature is used to represent the group propagation pattern formed by multiple users, content, and topics; Based on 0-simplex features 1-Simplified Features and 2-simplex characteristics We obtain multi-order simplex representations that range from node-level semantics to higher-order structure-level semantics. ;in To preserve the multimodal semantic information of network content entities at the node level, This refers to the local interactive relationships between users, content, and dissemination entities. This refers to the information on the high-order risk propagation topology formed by multiple entities.

[0010] Furthermore, the 0-simplex feature was obtained. The specific process is as follows: Traverse the 0-simplex and add it to the set Each 0-simplex; For nodes The corresponding 0-simplex, its text features Image features By splicing, we obtain nodes. Characteristic representation of 0-simplex .

[0011] Furthermore, the 1-simplex feature was obtained. The specific process is as follows: Traversal 1 - Adding the simplex to the set Each of the 1-simulacra; For the edge The corresponding 1-simplex is obtained by performing feature interaction based on the 0-simplex features at its two endpoints. ; side Characteristic representation of 1-simplex The expression is as follows: ; in, This is the Hadamard product operation. For max pooling aggregation operation, For nodes Characteristic representation of 0-simplex , .

[0012] Furthermore, construct 2-simplex features The process is as follows: Traversal 2 - Adding the simplex to the set Each of the 2-simulacra; Obtaining the triangular simplex 1-simplex features corresponding to the three internal edges , and ; Will , and By performing aggregation, the feature representation of the 2-simplex is obtained. ; 2- Simplex characteristics The expression is as follows: .

[0013] Furthermore, in step two, the 0-simplex feature... 1-Simplified Features and 2-simplex characteristics The propagation process at each layer is as follows: Step 2.1: For the simplex set Define the boundary matrix and ;in, This represents the boundary relationship between nodes and edges. Indicate the boundary relationship between the sides and the triangular simplex; Step 2.2: Based on the boundary matrix and Construct Hodge Laplace matrices of different orders ; Step 2.3: Apply the Hodge Laplace matrix After normalization, the normalized Hodge Laplace matrix is ​​obtained. ; Step 2.4: Construct the local topology enhancement matrix and multi-scale Hodge structure enhancement operators ; Step 2.5, for the first The characteristic matrix of the first-order simplex Construct two sets of random semantic mapping matrices with different mapping forms. and ;in, The two are based on a shared random projection space to form two sets of asymmetric semantic representations; the random mapping parameters are predetermined, remain fixed during the representation generation process and do not participate in training; Step 2.6: Based on the first random semantic mapping matrix and the second random semantic mapping matrix Two random semantic representations were obtained respectively. and For any two of the first Simplex Define asymmetric semantic consistency ,in, ; Step 2.7: Perform symmetry processing on the original semantic consistency matrix. ,get ; Step 2.8: Enhance the matrix using the Hodge structure. With the symmetricized multimodal semantic consistency matrix Construct the final propagation operator ; Step 2.9, for the first Recursive binary encoding based on random projection hashing is performed on the simplex of order 1; initialization No. During the round of iteration, structure-semantic propagation is performed first. Then, using a random projection matrix Perform LSH mapping Finally, a binary feature representation is generated using a sign function. ,in, ;go through After round of iterations , obtained the The final characteristic representation of an order simplex; Step 2.10, for the first Simplex ; in, ; when hour, ; Step 2.11: Segment the simplex representations of different orders. To obtain a graph-level risk representation .

[0014] Furthermore, the specific process of step three is as follows: Representing risks at the graph level As input to the classification model; Let the network content security training dataset be... ,in, Indicates the first Risk characterization corresponding to each network content sample; Indicates the corresponding risk label; For binary risk detection tasks, a logistic regression classifier can be used. ,in, Represents the Sigmoid function; Output the corresponding risk prediction results .

[0015] As a further aspect of the present invention, the present invention also provides an electronic device, including a memory, one or more processors, and one or more programs stored in the memory, said one or more programs including instructions for executing the network content security risk detection method as described above.

[0016] As a further aspect of the present invention, the present invention also provides a storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing the network content security risk detection method as described above.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) To address the problem that network content security scenarios simultaneously contain multi-source heterogeneous information such as text, images, and videos, and that there are complex relationships and high-order group interaction structures among users, content, propagation behavior, and interaction relationships, this invention proposes a network content security risk detection method based on non-learning multimodal simplex representation. This method takes a multimodal association graph composed of user information, content information, propagation relationships, and multimodal attribute information in the network content security scenario as input. By constructing a non-learning multimodal simplex representation, complex network content security data is mapped into graph-level risk features of a unified dimension. The network content security risk detection task is then achieved based on this graph-level representation combined with a traditional classification model.

[0018] (2) This invention addresses the complex high-order relationships that exist in the process of network content dissemination, user interaction, and content association. Based on the original network content association graph structure, it automatically constructs a simplex representation without the need for manual definition of high-order relationships or pre-construction of hyperedges. User nodes, content nodes, topic nodes, and related entities are mapped to 0-simplexes, user interaction relationships, content association relationships, and dissemination relationships are mapped to 1-simplexes, and group association patterns formed by multiple users around the same content, theme, or dissemination event are mapped to 2-simplexes. Furthermore, it utilizes multimodal information such as text, images, and videos to construct node-level semantic features. Through random mapping, different modal features are transformed into a unified semantic space, and edge-level simplex features are generated by the interaction of features between nodes. Combined with the high-order group association structure, high-order simplex features are constructed, so that the local semantic patterns, dissemination association patterns, and group risk dissemination structures in the network content are explicitly expressed, while preserving multi-source heterogeneous content information.

[0019] (3) This invention proposes a non-learning multimodal simplex representation construction mechanism. By utilizing the internal topological association relationship of the simplex and the Hodge structure enhancement operator, information propagation between simplexes of the same order is realized. The topological propagation process is modulated by the multimodal semantic consistency relationship, so that the semantic information of content such as text, image, and video is jointly fused with the user interaction structure, content propagation structure and high-order group association information. Furthermore, through recursive random mapping and hash generation mechanism, simplex structures of different orders and semantic information are compressed to form a graph-level risk representation, so that the multimodal semantic features and complex topological risk patterns in network content security data can participate in the risk detection process.

[0020] (4) In response to the risk detection needs of large-scale, multi-source heterogeneous content data and complex interconnected networks in network content security scenarios, this invention constructs a classification and detection mechanism based on graph-level risk representation. By constraining the generated graph-level representation, the semantic relationship, propagation structure relationship and high-order interaction mode between network content are maintained, so that the obtained risk features can effectively distinguish different types of network security risks. Combined with classification models such as logistic regression and support vector machine, the risk category identification and risk status detection in large-scale network content data are realized, thereby improving the risk detection efficiency in complex network environments.

[0021] (5) The method of the present invention can effectively improve the multimodal data representation capability in the process of network content security risk detection, so that the generated graph-level risk representation can simultaneously include semantic information of multimodal content such as text, images, and videos, user interaction relationship information and high-order propagation topology information, thereby improving the accuracy, stability and adaptability of risk content identification, risk category classification and abnormal propagation pattern detection tasks in complex network environments.

[0022] (6) This invention proposes a non-learning multimodal simplex representation method for network content security scenarios. This method can effectively solve the problems in the existing technology, such as the difficulty in fully utilizing high-order risk propagation relationships, insufficient fusion of multimodal content semantics and network structure information, and limited generalization ability of risk representation. While retaining the semantic information of multimodal content such as text, images, and videos, it fully explores the multi-order topological relationships formed in the process of user interaction, content association, and group propagation, and generates a unified risk representation suitable for network content security risk detection tasks, thereby improving the accuracy, stability, and adaptability of risk identification in complex network environments.

[0023] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0024] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating a network content security risk detection method based on non-learning multimodal simplex representation in Embodiment 1 of the present invention. Detailed Implementation

[0025] To make the above-mentioned objects, features, and advantages of the present invention clearer and easier to understand, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that the accompanying drawings of the present invention are all in a simplified form and use non-precise proportions, and are only used to facilitate and clearly illustrate the implementation of the present invention.

[0026] Example 1: See Figure 1 As shown in the figure, the network content security risk detection method based on non-learning multimodal simplex representation provided in this embodiment includes the following steps: Step 1: Obtain a multimodal network content risk map ;in, It is a set of entity nodes in the network content association graph, including user nodes, content nodes, topic nodes, and other related entities; It is a set of relationships between nodes, including posting relationships, commenting relationships, liking relationships, forwarding relationships, citation relationships, and propagation relationships; Risk map of multimodal network content Preprocessing is performed to obtain the k-simplex set. ; Based on k-simplex sets The multi-order features of the simplex are obtained.

[0027] Specifically, existing network content security risk detection methods typically rely on text content features, image features, and user interaction relationships for risk assessment. These methods struggle to fully utilize hidden high-order group association patterns during network propagation, resulting in risk representations that fail to comprehensively describe the semantic and structural information within complex network environments. To address this, this invention introduces a simplex structure. By automatically extracting high-order closure relationships from network content association graphs, it transforms traditional network association graphs into simplex representations containing topological relationships of different orders.

[0028] Preferred, multimodal network content risk map It also contains rich semantic and structural information. The content attributes of user-posted text, images, and videos describe the semantic features of entity nodes, while user comments, reposts, likes, posts, and content association behaviors reflect the structural relationships between nodes. To retain this information, this invention first extracts simplex structures of different orders from the multimodal association graph of the input network content. Then, based on the multimodal content features, it constructs simplex representations of corresponding orders, providing a foundation for subsequent non-learning simplex propagation and risk characterization generation.

[0029] Preferably, the k-simplex set is obtained. The specific process is as follows: Multimodal network content risk map As input, construct a set of k-simulacra containing all k-simulacra. This involves automatically extracting simplex structures of different dimensions from the original network content security association graph; specifically, constructing 0-simplexes and 1-simplexes, and traversing the node set in the network content security association graph. Add all entity nodes as 0-simplexes to the set. Among them, the set of nodes Each node in the network can correspond to a user, content, topic, or other network entity. In this embodiment, each node contains at least text modality and image modality, and can still be further extended to multimodal data such as video, audio, and user behavior. At the same time, traverse the set of relations Add all relationships between entities as 1-simplexes to the set. ; Based on the locally closed interaction structure in the original network content security association graph, node combinations that satisfy the pairwise association conditions between three entities are retrieved and added to the set as 2-simplexes. Specifically, in this embodiment, during the process of network content dissemination, if multiple users interact around the same risk event, multiple contents form an associated dissemination relationship around the same theme, or multiple entities form a closed dissemination link, then the corresponding three-node closed structure can be mapped to a 2-simplex.

[0030] Preferably, in the multimodal association graph of network content security, nodes typically contain multiple modal attributes such as text, images, and videos. This modal information can reflect the semantic features of the content and plays an important role in accurately identifying risky content. Therefore, after completing the simplex extraction, this invention constructs simplex features of various orders based on a multimodal mapping mechanism to fully preserve the multi-source semantic information in the network content and enhance the expressive power of complex propagation structures. The specific method for obtaining the multi-order features of the simplex is as follows: Set of k-simplexes and any node The corresponding multimodal attributes As input, we obtain the 0-simplex features respectively. 1-Simplified Features and 2-simplex characteristics ;in, For nodes The corresponding 0-simplex features are used to describe the multimodal semantic information of entities such as users, content, and topics; For the edge The corresponding 1-simplex feature is used to describe local risk association patterns in user interaction, content association, and propagation relationships; Triangular simplex The corresponding 2-simplex feature is used to represent the group propagation pattern formed by multiple users, content, and topics.

[0031] For further optimization, the features of each order of simplex are constructed according to the following steps: Step 1.1: Construct 0-simplex features The process is as follows: Traverse the 0-simplex and add it to the set Each 0-simplex; For nodes The corresponding 0-simplex, its text features Image features By splicing the components, we obtain the feature representation of the 0-simplex. Through this step, the original multimodal semantic information of the nodes is uniformly mapped to the 0-simplex representation.

[0032] node Characteristic representation of 0-simplex The expression is as follows: ; Step 1.2: Construct 1-simplex features The process is as follows: Traversal 1 - Adding the simplex to the set Each of the 1-simulacra; For the edge The corresponding 1-simplex is obtained by performing feature interaction based on the 0-simplex features at its two endpoints. ; side Characteristic representation of 1-simplex The expression is as follows: ; in, This is the Hadamard product operation. For max pooling aggregation operation, For nodes Characteristic representation of 0-simplex , This step constructs an edge-level structural representation by fusing multimodal semantic information between connected nodes. This enables the 1-simplex to not only describe the connection relationships between nodes, but also to capture the semantic similarity, differences, and interaction relationships between nodes, providing local structural information for subsequent graph-level representation generation.

[0033] Step 1.3: Construct 2-simplex features The process is as follows: Traversal 2 - Adding the simplex to the set Each of the 2-simulacra; Obtaining the triangular simplex 1-simplex features corresponding to the three internal edges , and ; Will , and By performing aggregation, the feature representation of the 2-simplex is obtained. ; 2-Characteristic representation of the simplex The expression is as follows: .

[0034] Through this step, the edge-level interaction information inside the triangular closed structure is further compressed into a higher-order simplex, enabling the 2-simplex to explicitly represent the higher-order topological patterns formed by multiple nodes and their corresponding multimodal semantic relationships.

[0035] Following the steps described above, after constructing 0-simplex, 1-simplex, and 2-simplex features respectively, we can obtain multi-order simplex representations ranging from node-level semantics to higher-order structural semantics. ;in To preserve the multimodal semantic information of network content entities at the node level, This refers to the local interactive relationships between users, content, and dissemination entities. This refers to the information on the high-order risk propagation topology formed by multiple entities.

[0036] The aforementioned multi-order simplex serves as the input for the subsequent non-learning simplex representation process. Through Hodge topology propagation, structure enhancement based on multimodal semantic consistency, and information interaction between simplexes of different orders, it further integrates semantic and topological information at different levels. After multi-order simplex representation generation and graph-level readout, it obtains risk feature representations of a unified dimension. Finally, it is combined with a classification model to realize the task of identifying network content security risk categories and detecting risks.

[0037] Step 2: Multi-order simplex propagation and graph-level risk representation generation layer; Existing network content security risk detection methods typically rely on information such as text content, image content, and user interaction relationships for risk assessment, making it difficult to fully utilize the complex relationships and high-order propagation patterns existing in cyberspace. Furthermore, traditional graph representation methods usually rely on ordinary graph structures composed of nodes and edges for information propagation, focusing primarily on local relationships between users, content, and propagating entities, neglecting the high-order group propagation structures formed by multiple users, content, and topics. This results in risk representations that struggle to simultaneously retain multimodal content semantic information and complex network propagation topology information.

[0038] To address this, this invention proposes a risk feature generation method based on non-learning multimodal simplex representation. Without requiring end-to-end parameter training, it generates risk representations of multimodal association graphs related to network content security through topology propagation, asymmetric semantic consistency modeling, and a recursive local sensitive hashing (LSH) generation mechanism. Specifically, after completing multimodal semantic initialization and simplex construction, this invention first constructs Hodge Laplacian matrices of different orders using the simplex boundary matrix, establishing topological relationships between nodes, edges, and higher-order closed structures. Subsequently, drawing on the idea of ​​asymmetric random mapping, it constructs two different forms of random semantic mapping for simplexes of different orders based on a shared random projection space. The feature interaction after bi-branch mapping characterizes the multimodal semantic consistency relationship between simplexes of the same order, and this semantic consistency information is used to modulate the Hodge topology propagation process. Finally, through recursive LSH projection and binarization operations, the structure-semantic information of simplexes of different orders is compressed into a specific-dimensional representation, and a unified-dimensional risk feature representation is obtained through graph-level readout.

[0039] Preferably, for multi-order simplex propagation processes, the definition is as follows: Input: Number of propagation rounds And the 0-simplex features obtained in the previous stage 1-Simplified Features and 2-simplex characteristics ; Output: After Final graph-level risk representation after layer propagation .

[0040] As a further embodiment, in step two, the 0-simplex feature is... 1-Simplified Features and 2-simplex characteristics The propagation process at each layer is as follows: (a) Hodge multi-scale topology enhancement; Step 2.1: For the simplex set Define the boundary matrix and ; in, This represents the boundary relationship between nodes and edges. Indicate the boundary relationship between the sides and the triangular simplex; Step 2.2: Based on the boundary matrix and Construct Hodge Laplace matrices of different orders Hodge Laplace matrices of different orders Includes the Hodge Laplace matrix used to represent the topological relationships between nodes. The Hodge Laplace matrix is ​​used to represent shared nodes and shared triangle relationships between edges. And the Hodge Laplace matrix used to represent the shared edge relationships between triangular simplexes. ; in, ; ; ; Step 2.3: To enhance the structural stability of networks of different sizes, the Hodge Laplace matrix is ​​normalized to obtain the normalized Hodge Laplace matrix. ;in, ; Step 2.4: Further construct the local topology enhancement matrix ;in, Represents the Hadamard product; Furthermore, in order to simultaneously preserve the simplex's intrinsic information, local topological associations, and multi-hop structural dependencies, this invention further constructs a multi-scale Hodge structural enhancement operator. ;in, , , This represents the self-preservation term of the simplex; This represents a first-order local topology enhancement term; This represents the second-order structure diffusion term.

[0041] (II) Multimodal semantic consistency modeling of asymmetric random mapping; To enable the propagation process to simultaneously consider network topology and multimodal semantic information, this invention further constructs a multimodal semantic consistency relationship between simplexes of the same order through a two-branch asymmetric random semantic mapping.

[0042] Step 2.5, for the first eigenvalue matrix of the first-order simplex Construct two sets of first random semantic mapping matrices with different mapping forms. and the second random semantic mapping matrix ;in, The two are based on a shared random projection space to form two sets of asymmetric semantic representations; For the first The order is the input feature dimension of the pure shape. The dimension of the public semantic representation after random semantic mapping (i.e., the dimension of the shared random projection space); the random mapping parameters are predetermined, remain fixed during the representation generation process and do not participate in training.

[0043] Step 2.6: Based on the first random semantic mapping matrix and the second random semantic mapping matrix Two sets of random semantic representations were obtained respectively. and For any two of the first Simplex Define asymmetric semantic consistency ,because ,therefore .

[0044] Step 2.7: Perform symmetry processing on the original semantic consistency matrix. ,get .

[0045] (III) Semantic consistency-aware Hodge semantic modulation propagation; Step 2.8: Enhance the matrix using the Hodge structure described above. and the symmetricized multimodal semantic consistency matrix Constructing the final propagation operator ;in, Represents a matrix of all 1s; This propagation mechanism enhances information exchange between simplexes with strong topological correlations and further increases the propagation weight between simplexes with high multimodal semantic consistency.

[0046] (iv) Recursive binary representation generation based on random projective hashing; Step 2.9: After semantic enhancement propagation, for the first... Recursive binary encoding based on random projection hashing is performed on the simplex of order 1; initialization No. During the round of iteration, structure-semantic propagation is performed first. ; Using random projection matrix Perform LSH mapping ; Binary feature representation is generated using a sign function. ,in, ; go through After round of iterations , obtained the The final characteristic representation of an order simplex.

[0047] (v) Generation of graph-level risk representations; Since the number of nodes, edges, and higher-order structures in the graphs associated with different network content varies, this invention uses an average readout method to generate graph-level risk representations.

[0048] Step 2.10, for the first Simplex ;when hour, ; in, .

[0049] Step 2.11: Segment the simplex representations of different orders. To obtain a graph-level risk representation .

[0050] This risk representation can include multimodal semantic information such as node-level text and images, and can be further extended to modalities such as video and audio; it also includes local structural information of user interaction and content dissemination; and a high-order risk propagation mode formed by multiple entities.

[0051] Specifically, the generated risk graph representation is input into classification models such as logistic regression and support vector machines to achieve network content security risk category detection.

[0052] Step 3: Network content security risk detection; After completing multi-order simplex propagation and generating graph-level risk representations, the system has obtained the risk feature representations corresponding to each network content security multimodal association graph. To achieve automated risk detection of large-scale multimodal content in complex network environments, it is necessary to further utilize the aforementioned graph-level risk representation to construct a classification detection model, and complete risk category identification based on the risk features corresponding to the input network content graph.

[0053] According to the task definition of this invention, the objective of network content security risk detection is to detect network content security risks in a given network content security multimodal correlation graph. In this context, a non-learning multimodal simplex representation method is used to generate a graph-level risk representation that simultaneously reflects content semantic information, user interaction relationships, and higher-order propagation structures. Based on this representation, the existence of security risks in the network content and its risk category are determined. Therefore, in this invention, the risk detection module is based on the risk representation obtained by the aforementioned multi-order simplex propagation and graph-level risk representation generation module. By training an offline risk classification model and conducting an online risk detection process, a complete network content security risk identification process can be achieved.

[0054] The process of detecting network content security risks is defined as follows: Input: Multimodal Relationship Diagram for Network Content Security And the graph-level risk representation obtained through the non-learning multi-order simplex representation generation module. ; Output: Network content security detection results , ,in, This represents the Sigmoid function. This indicates the predicted risk category. The network content security risk detection process is handled according to the following steps: This step enables a complete detection process from inputting multimodal information about network content to outputting risk categories.

[0055] Compared with the prior art, the method of this application has the following beneficial effects: ① This application addresses the challenges of complex multimodal information, diverse propagation relationships, and the difficulty in effectively modeling high-order group associations in network content security scenarios. It constructs a simplex representation automatically extracted from the original network association structure, mapping users, content, and high-order propagation relationships to a unified multi-order topological space, thus achieving automatic modeling of complex network content association patterns. By designing a non-learning multimodal simplex representation mechanism, utilizing random mapping, Hodge structure enhancement, and multi-order simplex information fusion methods, it effectively combines multimodal semantic information from text, images, and videos with network propagation topological information, enabling the generated risk representation to simultaneously retain content semantic features and high-order risk propagation patterns. Risk detection is achieved by combining graph-level risk representation with a classification model, realizing automated risk identification for large-scale network content data. This method eliminates the need for manually defining high-order relationships, does not rely on large-scale end-to-end training processes, and possesses good automation capabilities, structural representation capabilities, and cross-scenario adaptability, meeting the needs of multimodal content security risk detection in complex network environments.

[0056] ② Addressing the challenge that existing network content security risk detection methods struggle to simultaneously characterize multimodal content semantic information and high-order propagation topology information in scenarios where data includes text, images, and videos, and where complex relationships exist between users, content, topics, and propagation behaviors, this invention proposes a network content security risk detection method based on non-learning multimodal simplex representation. This method uses a multimodal association graph in the network content security field as a unified input, constructing a complete processing flow from multimodal data input, simplex structure construction, multi-order topology enhancement, non-learning risk representation generation, to risk classification and detection output. While preserving the multimodal semantic information of user, content, and topic nodes, it automatically extracts high-order interaction relationships from the network association structure and constructs a unified graph-level risk representation that integrates content semantic information, user interaction information, and propagation structure information. Ultimately, this achieves risk content identification, risk category classification, and abnormal propagation pattern detection in network content security scenarios.

[0057] ③ Existing network content security risk detection methods typically characterize risks based on text features, image features, or ordinary graph structures. They mainly utilize local correlation information between content nodes, user nodes, and interaction relationships, making it difficult to fully leverage higher-order propagation patterns formed by multiple users, content, and topics. This results in risk representations that cannot fully describe group relationships in complex network environments. Furthermore, some existing higher-order graph representation methods rely on manually defined hyperedges or pre-specified group relationships, such as manually constructing user groups, interest sets, or propagation communities, making it difficult to automatically discover potential higher-order risk association structures from raw network content interaction data. This invention constructs a simplex representation of the input network content security multimodal association graph, automatically extracting entity nodes, content relationships, and closed interaction structures to form 0-simplexes, 1-simplexes, and 2-simplexes, respectively. Further, by combining multimodal content attribute information such as text, images, and videos, a multi-order simplex representation is constructed, encompassing entity-level semantics, content-level semantics, interaction relationship semantics, and higher-order propagation structure semantics. Compared to existing methods, this invention eliminates the need for manual definition of higher-order risk relationships. It can automatically discover and utilize higher-order topological structures such as user collaborative behavior, content association propagation, and group diffusion patterns based on the original network content association structure, thereby enhancing the ability of network content security risk representation to express complex propagation relationships.

[0058] ④ To address the problem that traditional risk detection methods struggle to uniformly model content semantic information, user interaction information, and high-order propagation structure information due to the complex relationships between simplexes of different orders, this invention designs a non-learning multi-order simplex representation mechanism. In this mechanism, firstly, the topological relationships between simplexes of different orders are utilized to enhance the same-order structure of the 0-simplex, 1-simplex, and 2-simplex. Secondly, simplex boundary relationships are used to achieve information interaction between entity nodes, content relationships, and high-order propagation structures, enabling low-order multimodal semantic information to diffuse to high-order propagation structures, while high-order risk topological information can feed back to enhance low-order entity representations. Through a multi-layered propagation process, this invention can jointly fuse multimodal semantic information such as text and images in network content, user interaction relationships, and group propagation patterns, and generate a unified-dimensional network content security risk representation through multi-order simplex graph-level readout. Compared to traditional risk detection methods based solely on text content or local relationships, this invention can obtain a more complete multimodal network risk structure representation, improving the ability to recognize risk patterns in complex network environments.

[0059] ⑤ After obtaining the multi-order simplex graph-level risk representation, this invention further designs a classification and identification method for network content security risk detection. This method first utilizes a non-learning multimodal simplex representation module to encode the user relationship graph, content association graph, and multimodal propagation graph in network content security data, obtaining a graph-level risk representation in a unified feature space. Subsequently, the generated risk representation is input into a classification model, and risk identification is completed based on the feature differences between different risk categories. When new network content data is input, the corresponding risk features are obtained through the same non-learning representation process, and the risk detection results are output using the classification model. This method overcomes the limitations of traditional network content security detection methods that rely on single-modal information, keyword matching, or local content analysis. It achieves joint analysis of multimodal content information, user relationships, and high-order propagation patterns through a unified graph-level risk semantic space, making it suitable for applications such as network rumor detection, malicious content identification, illegal information detection, abnormal propagation behavior analysis, and simplex and multi-type network security risk classification.

[0060] ⑥ The network content security risk detection method based on non-learning multimodal representation proposed in this invention can effectively integrate the semantic information of multimodal content such as text, images, and videos with user nodes, content nodes, interaction relationships, and high-order propagation topology information, thereby obtaining a more comprehensive risk representation in network content security risk detection tasks. Compared with existing risk detection methods based on ordinary content features or simple relationship modeling, this invention can fully utilize user group interaction, content association propagation, and high-order risk diffusion patterns to improve the ability to distinguish between different risk categories and the detection accuracy in complex network environments. Furthermore, because this invention adopts a non-learning multimodal simplex representation method, it does not rely on a large-scale end-to-end training process, and has good automation, generalization, and scenario adaptability capabilities, meeting the practical application needs of large-scale network content security risk detection tasks.

[0061] Example 2: As a further embodiment of the present invention, the present invention also provides an electronic device, comprising: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0062] In practical use, users can interact with servers, which are also devices, via a network through electronic devices acting as terminals, to receive or send messages. Terminal devices are generally various devices equipped with a display and used through a human-machine interface, including but not limited to smartphones, tablets, laptops, and desktop computers. Various specific application software can be installed on these terminals as needed, including but not limited to web browsers, instant messaging software, social media platforms, and shopping apps.

[0063] Furthermore, the server is a network service provider that offers various services, such as a backend server that provides corresponding calculation services for the multimodal network content risk map received from the terminal device. This enables the processing of the network content security risk detection method based on non-learning multimodal simplex representation, calculates the network content security detection results, and finally returns them to the terminal device.

[0064] Example 3: As a further embodiment of the present invention, the present invention also provides a storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for executing the network content security risk detection method based on non-learning multimodal simplex representation as described above.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting network content security risks, characterized in that, Includes the following steps: Step 1: Obtain a multimodal network content risk map ;in, It is a set of entity nodes in the network content association graph, including user nodes, content nodes, and topic nodes; It is a set of relationships between nodes, including posting relationships, commenting relationships, liking relationships, forwarding relationships, citation relationships, and propagation relationships; Risk map of multimodal network content Preprocessing is performed to obtain the k-simplex set. ; Based on k-simplex sets We obtain the multi-order features of the simplex; the multi-order features of the simplex include 0-simplex features. 1-Simplified Features and 2-simplex characteristics ; Step 2: Set the number of propagation rounds to [number]. The number of layers in a single round of propagation is After propagating the multi-order features of the simplex, a graph-level risk representation is obtained. ; Step 3: Utilize graph-level risk representation Construct a classification and detection model, and apply it to the multimodal network content risk map. The corresponding risk characteristics are used to identify risk categories and obtain network content security detection results. .

2. The network content security risk detection method according to claim 1, characterized in that, Obtain the set of k-simplexes The specific process is as follows: Construct 0-simplex and 1-simplex, and traverse the set of nodes in the network content security association graph. Add all entity nodes as 0-simplexes to the set. Each node corresponds to a user, content, and topic. At the same time, traverse the set of relations Add all relationships between entities as 1-simplexes to the set. ; Based on the locally closed interaction structure in the original network content security association graph, node combinations that satisfy the pairwise association conditions between three entities are retrieved and added to the set as 2-simplexes. .

3. The network content security risk detection method according to claim 1 or 2, characterized in that, The specific process for obtaining the multi-order features of the simplex is as follows: Construct a multimodal network content risk graph any node in The corresponding multimodal properties are ;in, The number of modes; For the first The node feature matrix corresponding to each mode; For any modality in the network content security multimodal graph, , The number of modalities in the multimodal graph for network content security; Set of k-simplexes and any node The corresponding multimodal attributes As input, we obtain the 0-simplex features respectively. 1-Simplified Features and 2-simplex characteristics ;in, For nodes The corresponding 0-simplex features are used to describe the multimodal semantic information of entities such as users, content, and topics; For the edge The corresponding 1-simplex feature is used to describe local risk association patterns in user interaction, content association, and propagation relationships; Triangular simplex The corresponding 2-simplex feature is used to represent the group propagation pattern formed by multiple users, content, and topics; Based on 0-simplex features 1-Simplified Features and 2-simplex characteristics We obtain multi-order simplex representations that range from node-level semantics to higher-order structure-level semantics. ;in To preserve the multimodal semantic information of network content entities at the node level, This refers to the local interactive relationships between users, content, and dissemination entities. This refers to the information on the high-order risk propagation topology formed by multiple entities.

4. The network content security risk detection method according to claim 3, characterized in that, Obtain 0-simplex features The specific process is as follows: Traverse the 0-simplex and add it to the set Each 0-simplex; For nodes The corresponding 0-simplex, its text features Image features By splicing, we obtain nodes. Characteristic representation of 0-simplex .

5. The network content security risk detection method according to claim 3, characterized in that, 1-Simplified Features The specific process is as follows: Traversal 1 - Adding the simplex to the set Each of the 1-simulacra; For the edge The corresponding 1-simplex is obtained by performing feature interaction based on the 0-simplex features at its two endpoints. ; side Characteristic representation of 1-simplex The expression is as follows: ; in, This is the Hadamard product operation. For max pooling aggregation operation, For nodes The characteristic representation of the 0-simplex, .

6. The network content security risk detection method according to claim 3, characterized in that, The process of constructing 2-simplex features is as follows: Traversal 2 - Adding the simplex to the set Each of the 2-simulacra; Obtaining the triangular simplex 1-simplex features corresponding to the three internal edges , and ; Will , and By performing aggregation, the feature representation of the 2-simplex is obtained. ; 2-Characteristic representation of the simplex The expression is as follows: 。 7. The network content security risk detection method according to any one of claims 4-6, characterized in that, In step two, the 0-simplex feature... 1-Simplified Features and 2-simplex characteristics The propagation process at each layer is as follows: Step 2.1: For the simplex set Define the boundary matrix and ;in, This represents the boundary relationship between nodes and edges. Indicate the boundary relationship between the sides and the triangular simplex; Step 2.2: Based on the boundary matrix and Construct Hodge Laplace matrices of different orders ; Step 2.3: Apply the Hodge Laplace matrix After normalization, the normalized Hodge Laplace matrix is ​​obtained. ; Step 2.4: Construct the local topology enhancement matrix and multi-scale Hodge structure enhancement operators ; Step 2.5, for the first eigenvalue matrix of the first-order simplex Construct two sets of first random semantic mapping matrices with different mapping forms. and the second random semantic mapping matrix ;in, The two are based on a shared random projection space to form two sets of asymmetric semantic representations; For the first The order is the input feature dimension of the pure shape. The dimension of the public semantic representation after random semantic mapping; Step 2.6: Based on the first random semantic mapping matrix and the second random semantic mapping matrix Two random semantic representations were obtained respectively. and For any two of the first Simplex Define asymmetric semantic consistency ,in, ; Step 2.7: Perform symmetry processing on the original semantic consistency matrix. ,get ; Step 2.8: Enhance the matrix using the Hodge structure. With the symmetricized multimodal semantic consistency matrix Construct the final propagation operator ; Step 2.9, for the first Recursive binary encoding based on random projection hashing is performed on the simplex of order 1; initialization No. During the round of iteration, structure-semantic propagation is performed first. Then, using a random projection matrix Perform LSH mapping Finally, a binary feature representation is generated using a sign function. ,in, ;go through After round of iterations , obtained the The final characteristic representation of an order simplex; Step 2.10, for the first Simplex ;when hour, ; Step 2.11: Segment the simplex representations of different orders. To obtain a graph-level risk representation .

8. The network content security risk detection method according to claim 7, characterized in that, The specific process of step three is as follows: Representing risks at the graph level As input to the classification model; Let the network content security training dataset be... ,in, Indicates the first Risk characterization corresponding to each network content sample; Indicates the corresponding risk label; For binary risk detection tasks, a logistic regression classifier is used. ,in, Represents the Sigmoid function; Output the corresponding risk prediction results .

9. An electronic device, characterized in that, It includes a memory, one or more processors, and one or more programs stored in the memory, said one or more programs including instructions for performing the network content security risk detection method as described in any one of claims 1-8.

10. A storage medium, characterized in that, It includes one or more programs that are executed by one or more processors of an electronic device, the one or more programs including instructions for performing the network content security risk detection method as described in any one of claims 1-8.