Social network dynamic credibility assessment method and system based on blockchain fingerprint

Abstract

The application discloses a social network dynamic credibility evaluation method and system based on a blockchain fingerprint, wherein a content feature extraction module processes text features and analyzes multimedia content; a blockchain hash generation module adopts an SHA-3 algorithm and Beidou timing to construct double-channel verification; a data storage module realizes efficient storage based on an improved Merkle Patricia tree; a content propagation management module traces information variation by using a dynamic directed acyclic graph; a dynamic reputation scoring module realizes minute-level user portrait updating and reputation evaluation through six-dimensional features; an intelligent contract module establishes a three-level response mechanism to realize precise regulation and control; and a closed-loop optimization module continuously improves system performance through federated learning and a DAO mechanism. The application effectively solves the problems of difficult tracing, single evaluation and lagging regulation and control in traditional social network information governance, and realizes dynamic credibility management of the whole process from content production, propagation to feedback.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of data processing, in particular to a social network dynamic credibility evaluation method and system based on a blockchain fingerprint. BACKGROUND

[0002] The reliability problem of social network information dissemination has become a global challenge, and the spread of false information far exceeds that of real content. The World Health Organization has listed "information epidemic" as a public health threat. The current technical solution has significant limitations. Traditional blockchain evidence only supports linear recording and cannot track multiple version variants. Artificial intelligence detection is difficult to identify complex dissemination patterns due to single-point analysis and data silos. The existing evaluation system has problems such as single dimension and update lag. So far, there is no method for accurate evidence and tracking of dynamically disseminated social network information. Therefore, a social network dynamic credibility evaluation method and system based on a blockchain fingerprint are proposed to overcome the above defects and improve the credibility and reliability of information dissemination in social networks. SUMMARY

[0003] The application overcomes the defects of the prior art and provides a social network dynamic credibility evaluation method and system based on a blockchain fingerprint, which is important to improve the credibility and reliability of information dissemination in social networks.

[0004] To achieve the above purpose, the first aspect of the application provides a social network dynamic credibility evaluation method based on a blockchain fingerprint, which comprises the following steps:

[0005] Real-time capture of multi-modal content data from a target social network platform, feature extraction of the obtained multi-modal content data, and transmission of the encrypted multi-modal content features to the next processing unit after encryption processing of the extracted multi-modal content features;

[0006] According to the received encrypted multi-modal content features, a blockchain multi-branch fingerprint is generated, and the improved Merkle Patricia tree structure is used for fingerprint storage to obtain a blockchain fingerprint library;

[0007] A dynamic directed acyclic graph is constructed based on the branch fingerprint data stored in the blockchain fingerprint library, and the dynamic directed acyclic graph is used for abnormal propagation detection and tracing to obtain abnormal propagation detection information;

[0008] According to the decrypted multi-modal content features and the abnormal propagation detection information, a multi-dimensional credibility evaluation is performed on the target user to generate credibility evaluation information;

[0009] Based on the preset smart contract hierarchical response mechanism, it is judged whether regulation is needed based on the credibility evaluation information, and a jury mechanism is introduced to optimize controversial regulation cases.

[0010] In the scheme, the multi-modal content data captured in real time from the target social network platform is subjected to feature extraction, and the extracted multi-modal content features are encrypted and transmitted to the next processing unit. Specifically, the process includes:

[0011] The multi-modal content data is captured in real time from the target social network platform using a distributed crawler technology. The multi-modal content data includes text published by users, images uploaded, and video files.

[0012] For the captured text content data, non-semantic symbols are filtered through a regular expression, and the sentence is then broken down into a word token sequence using a word segmentation engine to generate a word token set. The word token set is used as input for the TF-IDF algorithm for initial feature processing.

[0013] The term frequency-inverse document frequency weight of each word in the document is calculated to generate a primary feature representation to represent the keyword distribution feature. The primary feature representation is input into a BERT model based on the Transformer architecture for deep semantic embedding to obtain the first modal content feature.

[0014] For the captured image and video data, the PHASH perceptual hashing algorithm is used to downsample the image or video key frame to a preset pixel specification and convert it to a grayscale image. The low-frequency coefficients are extracted through discrete cosine transform to generate a visual fingerprint. The StyleGAN3 discriminator model is called synchronously to analyze the local statistical features of the pixel matrix in layers and output the confidence probability value of the content being deep fake to obtain the second modal content feature.

[0015] The first modal content feature and the second modal content feature are combined to form a multi-modal content feature, which is encrypted through the SM4 national encryption algorithm and transmitted to the next processing unit through a dedicated secure channel.

[0016] In the scheme, the multi-modal content features are encrypted and stored in the Merkle Patricia tree structure to obtain a blockchain fingerprint library. Specifically, the process includes:

[0017] The encrypted multi-modal content features are imported into the SM4 decryption engine, and the pre-allocated security key is used to initialize the counter sequence. The ciphertext block and the counter are subjected to bitwise XOR operation, and the timestamp validity of the HMAC signature is verified synchronously. Finally, the decrypted multi-modal content features are output.

[0018] The decrypted multi-modal content features are input into an SHA-3 hash generator, the text semantic vector and the visual fingerprint are spliced to generate a main content abstract value, the abstract value is bound with the biological feature information of the publisher and the timestamp of the Beidou satellite timing system to form an initial evidence storage unit;

[0019] When the content modification behavior is detected, a branch fingerprint generation protocol is triggered, the new content feature vector after modification is compared with the original feature vector bit by bit through an XOR difference engine, the difference bits are marked and a difference sequence is generated;

[0020] The hash value corresponding to the original feature vector is spliced with the difference sequence, and an incremental hash value is generated through SHA-3 operation, which is used as a branch identification core and is encapsulated with the biological feature information of the modifier and the Beidou timestamp to form a branch fingerprint;

[0021] After the fingerprint generation is completed, all generated fingerprints are stored by using a Merkle Patricia tree structure, the parent node of the target fingerprint is retrieved, a new pointer is appended in the PPA to improve the Merkle Patricia tree structure, and adjacent single nodes are combined into extended nodes through a Radix tree optimization path compression algorithm, and finally a blockchain fingerprint library is obtained.

[0022] In the scheme, a dynamic directed acyclic graph is constructed based on the branch fingerprint data stored in the blockchain fingerprint library, and the dynamic directed acyclic graph is used for anomaly propagation detection and tracing to obtain anomaly propagation detection information, which specifically includes:

[0023] The blockchain fingerprint library is obtained, the branch fingerprint structure and the corresponding parent node pointer are extracted from the blockchain fingerprint library, each branch fingerprint is used as a graph node and the parent node pointer is used as a directed edge by using an adjacency list data structure, and an initial directed acyclic graph is constructed;

[0024] The initial directed acyclic graph is synchronized to a graph database Neo4j in real time, a unique index is created for each node by using a Cypher query language, a propagation relationship modeling is performed, and a dynamic directed acyclic graph is generated;

[0025] When a new branch fingerprint is written into the blockchain fingerprint library, the parent pointer chain is traced back to the root node, the new node and the edge are inserted after verifying that the path is acyclic, and the out-edge list of the adjacency list and the in-degree statistics are updated;

[0026] The dynamic directed acyclic graph is imported into a pre-trained graph attention network for anomaly propagation detection and tracing, the node feature vector is received through an input layer, and a propagation entropy feature extraction layer is set in the graph attention network;

[0027] In the propagation entropy feature extraction layer, the feature vectors of all parent nodes of each target node are aggregated, the propagation probability weight of the parent nodes to the target node is calculated through an attention mechanism, and then the propagation entropy value of the target node is calculated to represent the chaotic degree of the propagation path;

[0028] Based on the preset sliding window statistical entropy value change rate, when the absolute change rate of the entropy value in the continuous three windows exceeds the preset threshold, the propagation cluster centered on the target node is marked as an abnormal cluster, and the parent node chain is traversed in reverse through an adjacency list to locate the initial propagation source node, and abnormal propagation detection information is obtained.

[0029] In the scheme, the target user is subjected to multi-dimensional credibility assessment according to the decrypted multi-modal content feature and the abnormal propagation detection information, and credibility assessment information is generated, specifically including:

[0030] The decrypted multi-modal content feature and the abnormal propagation detection information are obtained, and a multi-dimensional credibility assessment mechanism is established based on a content authenticity dimension, a propagation anomaly dimension and a user credit dimension;

[0031] Based on the decrypted multi-modal content feature, content authenticity assessment is performed, text data is input into a RoBERTa model, semantic contradiction fragments are identified through a self-attention layer, and a contradiction probability value is output; for image data, an error level analysis algorithm is introduced to calculate pixel abnormal dispersion, and a tampering index is obtained;

[0032] For video data, a SyncNet model is used to detect audio-visual synchronization error and generate an abnormal coefficient; the contradiction probability value, the tampering index and the abnormal coefficient are linearly weighted and fused to obtain a content authenticity score;

[0033] According to the abnormal propagation detection information, propagation anomaly assessment is performed, the propagation entropy change rate and the abnormal node are extracted from the abnormal propagation monitoring information, the number of nodes marked as abnormal in the current propagation path is counted, the abnormal node proportion is multiplied by the propagation entropy change rate to generate a path abnormality coefficient, which is used as a propagation anomaly score;

[0034] For the user credit dimension, the evaluation results of six dimensions of content quality, propagation behavior compliance, social influence, historical reporting accuracy, community contribution value and activity level are weighted and fused to generate a user credit score;

[0035] The content authenticity score, the propagation anomaly score and the user credit score are weighted using a preset weighting fusion coefficient to generate a comprehensive credibility index, and credibility assessment information is obtained.

[0036] In the scheme, the preset smart contract hierarchical response mechanism judges whether regulation is needed through the credibility assessment information, and introduces a jury mechanism to optimize controversial regulation cases, specifically including:

[0037] Obtain the credibility evaluation information, judge whether it is necessary to carry out regulation and control in combination with the preset smart contract hierarchical response mechanism, when the credibility evaluation information is greater than a type threshold, it is marked as green content, and bandwidth priority and recommendation weighting are given;

[0038] When the credibility evaluation information is less than a type threshold and greater than a type threshold, it is marked as yellow content, and the propagation rate is limited;When the credibility evaluation information is less than a type threshold, it is marked as red content, the propagation path is frozen, the abnormal source node is located and the investigation report is generated;

[0039] When the regulation and control strategy is executed, the report data submitted by the user end is input into the adversarial sample generation module based on DBSCAN density clustering, the Euclidean distance is measured to measure the feature similarity of the report case, the high conflict samples connected in density are aggregated into an adversarial training cluster, and the generated adversarial sample is input into the federal learning framework to share model parameter updates in an encrypted state;

[0040] The jury mechanism is introduced to optimize the dispute regulation and control case, and a number of users with user credit scores greater than a preset user credit score threshold are randomly selected to form an arbitration committee according to the user credit scores output by the user credit dimension;

[0041] The controversial content is anonymized through zero-knowledge proof technology, three rounds of voting are conducted to confirm the final disposal scheme, and the adversarial sample is generated according to the controversial regulation and control case characteristics, the judgment conclusion and the disposal scheme, so as to drive the federal learning framework to enhance the judgment ability of fuzzy cases in the next iteration.

[0042] The second aspect of the application provides a social network dynamic credibility evaluation system based on blockchain fingerprints, characterized in that the system comprises a content feature extraction module, a blockchain hash generation and storage module, a content propagation and branch management module, a multi-modal artificial intelligence evaluation module, a dynamic credit evaluation module, an intelligent contract dynamic regulation and control module, and a closed-loop feedback and optimization module:

[0043] The content feature extraction module is used to parse and extract key structured feature vectors from multi-modal content data;

[0044] The blockchain hash generation and storage module processes the feature vectors and content fingerprints into unique and tamper-proof hash values and anchors them to the distributed ledger;

[0045] The content propagation and branch management module is used to dynamically track and record the propagation path and derived version of the content, and perform propagation anomaly detection;

[0046] The multi-modal artificial intelligence evaluation module automatically detects the content authenticity and tampering traces by using a preset model.

[0047] The dynamic reputation evaluation module is used for dynamically calculating and updating the real-time reputation score of a user or a node, and judging the credibility of the propagated content.

[0048] The smart contract dynamic regulation module is used for performing priority scheduling, rate limiting or safety fuse control strategy of content propagation.

[0049] The closed-loop feedback and optimization module generates adversarial training samples using user feedback and dispute arbitration data to drive the federated learning mechanism to continuously update the multi-modal evaluation model parameters.

[0050] The application discloses a social network dynamic credibility evaluation method and system based on a blockchain fingerprint, a content feature extraction module processes text features and analyzes multimedia content; a blockchain hash generation module adopts a SHA-3 algorithm and a Beidou time service to construct a double-channel verification; a data storage module realizes efficient storage based on an improved Merkle Patricia tree; a content propagation management module uses a dynamic directed acyclic graph to track information variation; a dynamic reputation scoring module realizes minute-level user portrait updating and reputation evaluation through six-dimensional features; a smart contract module establishes a three-level response mechanism to realize precise regulation; and a closed-loop optimization module continuously improves system performance through federated learning and a DAO mechanism. The problems of traceability difficulty, single evaluation and regulation lag in traditional social network information governance are effectively solved, and dynamic credibility management of the whole process from content production to feedback is realized. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the drawings needed to be used in the embodiments or examples will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0052] Figure 1 A social network dynamic credibility evaluation method based on a blockchain fingerprint is provided for an embodiment of the present application.

[0053] Figure 2 A social network dynamic credibility evaluation system block diagram based on a blockchain fingerprint is provided for an embodiment of the present application.

[0054] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the drawings. DETAILED DESCRIPTION

[0055] In order to enable a more clear understanding of the above-mentioned objects, features and advantages of the present application, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0056] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, however, the present application can also be implemented in other manners different from those described herein, and therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0057] Figure 1 A social network dynamic credibility evaluation method based on a blockchain fingerprint is provided for an embodiment of the present application, and a flowchart of the method is shown in the figure.

[0058] As Figure 1 shown, the present application provides a social network dynamic credibility evaluation method based on a blockchain fingerprint, and a flowchart of the method is shown in the figure, which includes:

[0059] S102, from the multi-modal content data captured in real time from the target social network platform, extracting features from the obtained multi-modal content data, and transmitting the encrypted multi-modal content features to the next processing unit after encryption processing;

[0060] S104, generating a blockchain multi-branch fingerprint according to the received encrypted multi-modal content features, and storing the fingerprint using an improved Merkle Patricia tree structure to obtain a blockchain fingerprint library;

[0061] S106, constructing a dynamic directed acyclic graph based on the branch fingerprint data stored in the blockchain fingerprint library, and using the dynamic directed acyclic graph for anomaly propagation detection and tracing to obtain anomaly propagation detection information;

[0062] S108, performing multi-dimensional credibility evaluation on the target user according to the decrypted multi-modal content features and the anomaly propagation detection information, and generating credibility evaluation information;

[0063] S110, based on a preset smart contract hierarchical response mechanism, determining whether regulation is needed through the credibility evaluation information, and introducing a jury mechanism to optimize controversial regulation cases.

[0064] Further, in a preferred embodiment of the present application, the multi-modal content data captured in real time from the target social network platform, the extracted multi-modal content features are encrypted and transmitted to the next processing unit, which specifically includes:

[0065] The distributed crawler technology is used to capture multi-modal content data in real time on a target social network platform, the multi-modal content data including text, images and video files published by users;

[0066] For the captured text content data, non-semantic symbols are filtered through a regular expression, and a wordpiece set is generated by using a word segmentation engine to split sentences into wordpiece sequences;

[0067] The term frequency-inverse document frequency weight of each word in the document is calculated to generate a primary feature representation to represent the keyword distribution feature, and the primary feature representation is input into a BERT model based on a Transformer architecture for deep semantic embedding to obtain a first modal content feature;

[0068] For the captured image and video data, a PHASH perceptual hashing algorithm is used to downsample image or video key frames to a preset pixel specification and convert them into grayscale images, low-frequency coefficients are extracted by discrete cosine transform to generate visual fingerprints, a StyleGAN3 discriminator model is called synchronously, local statistical features of the pixel matrix are analyzed hierarchically, and a confidence probability value of content deep forgery is output to obtain a second modal content feature;

[0069] The first modal content feature and the second modal content feature are combined to form a multi-modal content feature, which is encrypted by an SM4 national encryption algorithm and transmitted to the next processing unit through a special secure channel.

[0070] It should be noted that the distributed crawler system is used to real-time capture multi-modal content data in social network platforms, including original text, uploaded pictures and video files and other diversified carriers published by users. For text content, first, irrelevant non-semantic symbols are removed through regular expression, and then a word segmentation engine is used to accurately disassemble the sentence into a discrete word sequence. The word set is imported as a key input into the TF-IDF algorithm processing flow, and a primary statistical feature vector is generated by calculating the term frequency-inverse document frequency weight, thereby capturing the distribution characteristics of the text keywords. The primary feature representation is then input into the BERT deep learning model based on the Transformer architecture, which learns the context semantic association through the self-attention mechanism, and finally outputs a 768-dimensional deep semantic embedding vector as the first modal content feature. For image and video data, a parallel processing channel is started: on the one hand, the PHASH perceptual hashing algorithm is used to perform a standardized down-sampling operation on the image or video key frame, which is uniformly converted into a 32x32 pixel grayscale image. The low-frequency component coefficients are extracted by discrete cosine transform to generate a compact visual fingerprint sequence. On the other hand, the StyleGAN3 pre-trained discriminator model is simultaneously called, which disassembles the input image pixel matrix layer by layer and extracts local texture statistical features, and then outputs the confidence probability value of the content existing deep fake operation. The above visual fingerprint and fake confidence are jointly encapsulated as the second modal content feature package. Finally, the system vertically concatenates the first modal (text semantic vector) and the second modal (visual fingerprint + fake confidence) features in the feature space to form a complete multi-modal content feature representation. The fusion feature is highly encrypted by the SM4 algorithm, and is transmitted in real time to the downstream processing unit using the platform's pre-set SSL / TLS secure channel, providing core data input for subsequent blockchain notarization and credibility assessment.

[0071] Further, in a preferred embodiment of the present application, the multi-branch fingerprint of the blockchain is generated according to the received encrypted multi-modal content features, and the improved Merkle Patricia tree structure is used for fingerprint storage to obtain a blockchain fingerprint library, which specifically comprises:

[0072] The encrypted multi-modal content features are obtained, and the encrypted multi-modal content features are imported into the SM4 decryption engine, and the pre-allocated security key is used to initialize the counter sequence, and the ciphertext block is XORed with the counter bit by bit, and the timestamp validity of the HMAC signature is verified synchronously, and finally the decrypted multi-modal content features are output;

[0073] The decrypted multi-modal content features are input into the SHA-3 hash generator, and the text semantic vector and the visual fingerprint are concatenated to generate a main content digest value, which is bound with the publisher's biometric information and the timestamp of the Beidou satellite timing system to form an initial notarization unit;

[0074] When the content modification behavior is detected, the branch fingerprint generation protocol is triggered, the modified new content feature vector is compared with the original feature vector bit by bit through the XOR difference engine, the difference bits are marked and the difference sequence is generated;

[0075] The hash value corresponding to the original feature vector is obtained, spliced with the difference sequence, and an incremental hash value is generated through SHA-3 operation as a branch identification core and the biological feature information of the modifier, Beidou timestamp is jointly encapsulated as a branch fingerprint;

[0076] After completing the fingerprint generation, all generated fingerprints are stored by adopting the Merkle Patricia tree structure, the parent node of the target fingerprint is retrieved and a new pointer is appended in the PPA to improve the Merkle Patricia tree structure, and through the Radix tree optimization path compression algorithm, adjacent single nodes are combined into extended nodes, and finally a blockchain fingerprint library is obtained.

[0077] It should be noted that the encrypted multi-modal content feature is first input into the national SM4 decryption engine for security analysis, the pre-allocated security key is used to initialize the counter sequence, the ciphertext block is XORed with the dynamic counter bit by bit through the block cipher counter mode to realize real-time decryption, the validity of the additional HMAC signature timestamp is verified synchronously to ensure that the data is not attacked by a man-in-the-middle during transmission, and finally the complete decrypted multi-modal content feature is output. Then input into the SHA-3 hash generator, the text semantic vector and the visual fingerprint are byte-level spliced to calculate, generate a 256-bit main content digest value, and bind it with the publisher's device biometric information (such as FIDO2 authentication fingerprint) and the nanosecond-level timestamp provided by the Beidou satellite timing system to form an inseparable initial evidence unit. When the system detects content modification behavior through the version comparison algorithm, the branch fingerprint generation protocol is automatically triggered. The XOR difference engine compares the modified new content feature vector with the original feature vector bit by bit, marks the difference bits and generates a binary difference sequence; at the same time, the main hash value corresponding to the original content feature is obtained, and the hash value and the difference sequence are spliced into a new input source, and a unique incremental hash value is generated through SHA-3 hash operation. The value is encapsulated as a structured branch fingerprint data packet together with the modifier's biometric features and the Beidou timestamp. All generated fingerprint data are stored in a structured way through the improved Merkle Patricia tree, the core of which is to dynamically append new branch pointers in the parent pointer array (PPA) after retrieving the parent node of the target fingerprint. Through the Radix tree optimization algorithm, the path compression mechanism is enhanced, and adjacent single nodes are combined into composite extended nodes, which significantly reduces the tree height and query complexity, and finally a blockchain fingerprint library supporting efficient version tracing is constructed. The library realizes minute-level integrity verification through the zk-STARKs zero-knowledge proof technology, and provides auditable underlying data support for subsequent credibility evaluation.

[0078] Further, in a preferred embodiment of the present application, a dynamic directed acyclic graph is constructed based on the branch fingerprint data stored in the blockchain fingerprint library, and the dynamic directed acyclic graph is used for anomaly propagation detection and tracing to obtain anomaly propagation detection information, which specifically includes:

[0079] Obtain the blockchain fingerprint library, extract the branch fingerprint structure and the corresponding parent node pointer from the blockchain fingerprint library, use the adjacency list data structure to construct an initial directed acyclic graph by taking each branch fingerprint as a graph node and the parent node pointer as a directed edge;

[0080] Synchronize the initial directed acyclic graph to the graph database Neo4j in real time, create a unique index for each node through the Cypher query language, and perform propagation relationship modeling to generate a dynamic directed acyclic graph;

[0081] When a new branch fingerprint is written into the blockchain fingerprint library, the path is verified to be acyclic by traversing the parent pointer chain back to the root node, and then the new node and edge are inserted, and the out-edge list of the adjacency list and the in-degree statistics are updated;

[0082] The dynamic directed acyclic graph is imported into a pre-trained graph attention network for anomaly propagation detection and tracing, the node feature vector is received through the input layer, and the propagation entropy feature extraction layer is set in the graph attention network;

[0083] In the propagation entropy feature extraction layer, the feature vectors of all parent nodes of each target node are aggregated, the propagation probability weight of the parent node to the target node is calculated through the attention mechanism, and then the propagation entropy value of the target node is calculated to represent the confusion degree of the propagation path;

[0084] Based on the preset sliding window statistical entropy value change rate, when the absolute change rate of the entropy value in the continuous three windows exceeds the preset threshold, the propagation cluster centered on the target node is marked as an abnormal cluster, and the parent node chain is traversed in reverse through the adjacency list to locate the initial propagation source node, and the abnormal propagation detection information is obtained.

[0085] It should be noted that firstly, the branch fingerprint structure and its associated parent node pointer data are extracted from the blockchain fingerprint library, each branch fingerprint is mapped to a graph node using an adjacency list data structure, and the parent node pointer is converted into a directed edge, and an initial propagation directed acyclic graph skeleton is constructed. The adjacency list stores the direct successor relationship of the node through the pointer array, and maintains the in-degree counter to prevent loop generation. The constructed initial graph structure is synchronized in real time to the graph database Neo4j through the Cypher query language, a unique index is created for each node, the propagation relationship is modeled as a "MODIFY_FROM" relationship type, and a timestamp and modifier identity attribute are attached to form a dynamically updated persistent storage topology. When the blockchain fingerprint library writes a new branch fingerprint, an incremental update process is triggered: recursively traverse the parent pointer chain of the new fingerprint back to the root node, verify the loop-free path, and insert the new node in the adjacency list and update the out-edge list and in-degree statistics of the related nodes. Based on the real-time evolving directed acyclic graph, load the pre-trained graph attention network model, and receive the node feature vector containing the content hash, timestamp, and modifier reputation in the input layer; through the built-in propagation entropy feature extraction layer of the model, aggregate the feature vectors of all parent nodes of each target node, calculate the propagation probability weight of the parent nodes to the target node using the multi-head attention mechanism, and then derive the propagation entropy value of the target node to quantify the propagation path chaos degree. Subsequently, the absolute change rate of the propagation entropy is calculated continuously with a 10-minute sliding window, and when the entropy value change rate in the consecutive three time windows exceeds the preset threshold a (a = 0.35), the propagation cluster centered on the target node is automatically marked as an abnormal cluster. Based on the topology relationship stored in the adjacency list, the system recursively traverses the parent pointer chain of the abnormal node in reverse, locates the initial source node in the propagation path, and generates a detection report containing the abnormal cluster identifier, source node address, and propagation path. The final output of the abnormal propagation detection information provides key topological feature input for subsequent three-dimensional credibility assessment, forming a closed-loop logic chain from propagation path analysis to risk assessment.

[0086] Further, in a preferred embodiment of the present application, the multi-dimensional credibility assessment of the target user according to the decrypted multi-modal content features and the abnormal propagation detection information generates credibility assessment information, specifically including:

[0087] Obtain the decrypted multi-modal content features and the abnormal propagation detection information, and establish a multi-dimensional credibility assessment mechanism based on the content authenticity dimension, the propagation anomaly dimension, and the user reputation dimension;

[0088] Perform content authenticity assessment based on the decrypted multi-modal content features, input text data into the RoBERTa model, identify semantic contradictory fragments through the self-attention layer, and output a contradiction probability value; for image data, introduce an error level analysis algorithm to calculate the pixel abnormal dispersion degree to obtain a tampering index;

[0089] For video data, the audio-lip synchronization error is detected by a SyncNet model and an anomaly coefficient is generated; the contradiction probability value, the tampering index and the anomaly coefficient are linearly weighted and fused to obtain a content authenticity score;

[0090] According to the abnormal propagation detection information, a propagation anomaly evaluation is performed; a propagation entropy change rate and an abnormal node are extracted from the abnormal propagation monitoring information; the number of nodes marked as abnormal in the current propagation path is counted; the abnormal node proportion is multiplied by the propagation entropy change rate to generate a path anomaly coefficient as a propagation anomaly score;

[0091] For the user reputation dimension, the evaluation results of the six dimensions of content quality, propagation behavior compliance, social influence, historical reporting accuracy, community contribution value and activity are weighted and fused to generate a user reputation score;

[0092] The content authenticity score, the propagation anomaly score and the user reputation score are weighted using a preset weighting fusion coefficient to generate a comprehensive credibility index to obtain credibility evaluation information.

[0093] It should be noted that based on the decrypted multi-modal content features and abnormal propagation detection information, a three-dimensional credibility evaluation system is constructed. In the content authenticity dimension, the text data is input into the RoBERTa model, and the semantic contradiction fragments are identified through the self-attention mechanism of the 12-layer Transformer encoder, and the contradiction probability value in the interval of 0-1 is output; after the error level analysis algorithm is applied to the image data, the tampering index is calculated by calculating the dispersion difference between the pixel blocks; the video data is extracted through the SyncNet model to extract the lip shape flow and audio MFCC features, and the audio-lip synchronization error coefficient is output after time alignment. The three groups of indexes are linearly fused into the content authenticity score according to the preset weight (text 0.4 / image 0.3 / video 0.3). The abnormal propagation dimension directly uses the abnormal propagation detection information: the abnormal node density ratio is calculated by extracting the number of abnormal nodes in the propagation path and the total length of the path; the propagation entropy change rate (taken from the mean value of the continuous time window) of the path is synchronously obtained, and the two are multiplied to generate a path abnormality coefficient in the range of 0-10 as the propagation abnormality score. The user reputation dimension integrates six dynamic features: content quality score (taken from the mean value of the AI detection in step one), propagation compliance rate (abnormal propagation cluster proportion), social influence (improved PageRank value), report accuracy (effective report times / total report times), community contribution value (number of high-quality modified contents) and activity (daily interaction frequency). After dynamically adjusting the weight of each dimension by the time decay function, input the S-shaped function to normalize the real-time reputation score in the interval of 0-100. Finally, the three-dimensional scores (content authenticity A, propagation abnormality B, and user reputation C) are weighted according to the preset fusion coefficient: comprehensive credibility index = 0.15xA+ 0.25x(10-B) / 10 + 0.6xC / 100, and the generated credibility evaluation information provides a quantitative decision basis for subsequent smart contract regulation.

[0094] Further, in a preferred embodiment of the present application, the preset smart contract hierarchical response mechanism judges whether regulation is needed based on the credibility evaluation information, and introduces a jury mechanism to optimize controversial regulation cases, specifically including:

[0095] Obtain the credibility evaluation information, and judge whether regulation is needed in combination with the preset smart contract hierarchical response mechanism, when the credibility evaluation information is greater than a type of threshold, it is marked as green content, and given bandwidth priority and recommendation weighting;

[0096] When the credibility evaluation information is less than a type of threshold and greater than a type of threshold, it is marked as yellow content, and the propagation rate is limited; when the credibility evaluation information is less than a type of threshold, it is marked as red content, and the propagation path is frozen, and the abnormal source node is located and the investigation report is generated simultaneously;

[0097] The closed-loop optimization process is started while the regulation strategy is executed. The user-submitted report data is input into the adversarial sample generation module based on DBSCAN density clustering. The Euclidean distance is used to measure the similarity of the report case features. High-conflict samples with connected density are aggregated into adversarial training clusters. The generated adversarial samples are input into the federated learning framework, and the model parameter updates are shared in an encrypted state.

[0098] The jury mechanism is introduced to optimize controversial cases. According to the user reputation score output by the user reputation dimension, a number of users with user reputation scores greater than the preset user reputation score threshold are randomly selected to form an arbitration committee.

[0099] The controversial content is anonymized through zero-knowledge proof technology. Three rounds of voting are conducted to confirm the final disposal scheme. The adversarial samples are generated based on the characteristics of the controversial regulation cases, the ruling conclusion, and the disposal scheme to drive the federated learning framework to enhance the judgment ability for ambiguous cases in the next iteration.

[0100] It should be noted that the credibility evaluation information triggers the intelligent contract hierarchical response mechanism. When the index is higher than 80, it is automatically marked as green content, the bandwidth priority allocation algorithm is activated (the transmission weight is increased by 30%), and the recommendation exposure weight factor is increased. When the index is between 40 and 80, it is marked as yellow content, and the propagation rate limiter is started, which delays the forwarding to 3 times the baseline value. When the index is detected to be lower than 40, it is immediately marked as red content, the whole network propagation path is frozen, and the traceability engine is called synchronously. The initial abnormal source node is located through reverse traversal of the adjacency list, and an investigation report containing the propagation path map is generated. During the execution of the regulation strategy, the closed-loop optimization process is started synchronously. The user report data is input into the DBSCAN density clustering module after feature engineering conversion. The sample similarity is calculated through the Euclidean distance. High-conflict cases with connected density within the ε radius are aggregated into adversarial training clusters. The cluster data is input into the federated learning framework after desensitization. Each participant (such as a social platform) exchanges model parameter updates in an encrypted state through homomorphic encryption technology. Every 24 hours, the FedAvg algorithm is used to aggregate the global model weight. For critical cases (such as CRED scores between 35 and 45), the jury optimization mechanism is activated: according to the CRED score output by the dynamic reputation evaluation module, 21 users with scores higher than 90 are randomly selected from the whole network to form an arbitration committee. The controversial content is anonymized through zk-SNARK zero-knowledge proof technology. The committee members conduct three rounds of blind voting, and the final disposal scheme is confirmed through Byzantine fault tolerance consensus. The feature vector, ruling conclusion, and disposal label of this controversial case are collectively encapsulated as a reinforcement sample, driving the federated learning framework to enhance the judgment ability for ambiguous scenarios in the next iteration, forming a closed-loop optimization path from dispute disposal to model evolution.

[0101] Figure 2A social network dynamic credibility evaluation system based on a blockchain fingerprint is provided for an embodiment of the application, and the system comprises a content feature extraction module, a blockchain hash generation and storage module, a content propagation and branch management module, a multi-modal artificial intelligence evaluation module, a dynamic reputation evaluation module, an intelligent contract dynamic regulation module, and a closed-loop feedback and optimization module.

[0102] The content feature extraction module is configured to parse and extract a key structured feature vector from multi-modal content data.

[0103] The blockchain hash generation and storage module processes the feature vector and the content fingerprint into a unique and tamper-proof hash value and anchors the hash value to a distributed ledger.

[0104] The content propagation and branch management module is configured to dynamically track and record the propagation path and derivative versions of the content and perform propagation anomaly detection.

[0105] The multi-modal artificial intelligence evaluation module automatically detects the authenticity and tampering traces of the content by using a preset model.

[0106] The dynamic reputation evaluation module is configured to dynamically calculate and update the real-time reputation score of a user or a node and determine the credibility of the propagated content.

[0107] The intelligent contract dynamic regulation module is configured to perform priority scheduling, rate limiting, or safety fuse control strategies for content propagation.

[0108] The closed-loop feedback and optimization module generates adversarial training samples using user feedback and dispute arbitration data to drive a federated learning mechanism to continuously update multi-modal evaluation model parameters.

[0109] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only illustrative, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0110] The units described as separate components above can or can not be physically separate, and the components shown as units can or can not be physical units; they can be located in one place or distributed on multiple network units; and part or all of the units can be selected according to actual needs to achieve the purposes of the embodiments.

[0111] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in the form of hardware or in the form of hardware plus software functional units.

[0112] Those skilled in the art can understand that all or part of the steps of the above method embodiments can be completed by program instruction related hardware, and the foregoing program can be stored in a computer readable storage medium, and the program executes the steps including the above method embodiments when executed; and the foregoing storage medium includes mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic discs or optical discs, and various storage medium that can store program codes.

[0113] Alternatively, the integrated units of the present application, if implemented in the form of software functional modules and sold or used as independent products, can also be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present application can be embodied in the form of software products, and the computer software products are stored in a storage medium, including a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the methods described in the embodiments of the present application. The foregoing storage medium includes mobile storage devices, ROMs, RAMs, magnetic discs or optical discs, and various storage medium that can store program codes.

[0114] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for dynamic trustworthiness assessment of social networks based on blockchain fingerprints, characterized in that, include: Multimodal content data captured in real time from the target social network platform is processed by feature extraction. The extracted multimodal content features are then encrypted and transmitted to the next processing unit. Based on the received encrypted multimodal content features, a blockchain multi-branch fingerprint is generated, and an improved Merkle Patricia tree structure is used for fingerprint storage to obtain a blockchain fingerprint database. A dynamic directed acyclic graph is constructed based on the branch fingerprint data stored in the blockchain fingerprint database. The dynamic directed acyclic graph is then used for anomaly propagation detection and source tracing to obtain anomaly propagation detection information. Based on the decrypted multimodal content features and the anomaly propagation detection information, a multidimensional credibility assessment of the target user is performed, generating credibility assessment information. Based on the preset smart contract hierarchical response mechanism, the credibility assessment information is used to determine whether regulation is needed, and a jury mechanism is introduced to optimize dispute regulation cases. Specifically, the step of performing a multi-dimensional credibility assessment of the target user based on the decrypted multimodal content features and the anomaly propagation detection information, and generating credibility assessment information, includes: Obtain the decrypted multimodal content features and the abnormal propagation detection information, and establish a multidimensional credibility assessment mechanism based on the content authenticity dimension, propagation abnormality dimension, and user reputation dimension; Content authenticity is evaluated based on the decrypted multimodal content features. Text data is input into the RoBERTa model, and semantically contradictory segments are identified through a self-attention layer, outputting the contradiction probability value. For image data, an error level analysis algorithm is introduced to calculate pixel anomaly dispersion and obtain the tampering index. For video data, the SyncNet model is used to detect lip-sync errors and generate anomaly coefficients; the contradiction probability value, tampering index and anomaly coefficients are linearly weighted and fused to obtain the content authenticity score; Based on the abnormal propagation detection information, a propagation anomaly assessment is performed. The propagation entropy change rate and abnormal nodes are extracted from the abnormal propagation detection information. The number of nodes marked as abnormal in the current propagation path is counted. The proportion of abnormal nodes is multiplied by the propagation entropy change rate to generate a path anomaly coefficient, which is used as the propagation anomaly score. For the user reputation dimension, the evaluation results of six dimensions—content quality, compliance of dissemination behavior, social influence, historical accuracy of reports, community contribution value, and activity level—are weighted and integrated to generate a user reputation score. The content authenticity score, dissemination anomaly score, and user reputation score are weighted using a preset weighted fusion coefficient to generate a comprehensive credibility index, thus obtaining credibility assessment information.

2. The method for dynamic trustworthiness assessment of social networks based on blockchain fingerprints according to claim 1, characterized in that, The process of capturing multimodal content data in real time from the target social network platform, extracting features from the acquired multimodal content data, encrypting the extracted multimodal content features, and then transmitting them to the next processing unit includes: Distributed crawling technology is used to capture multimodal content data in real time on the target social network platform. The multimodal content data includes text posted by users, uploaded images, and video files. For the captured text content data, non-semantic symbols are filtered out by regularization expressions, and then the word segmentation engine is used to decompose the sentence into a word sequence to generate a word set. The word set is used as the input of the TF-IDF algorithm for initial feature processing. Calculate the term frequency-inverse document frequency weight of each word in the document to generate a primary feature representation to characterize the keyword distribution features. Input the primary feature representation into the BERT model based on the Transformer architecture for deep semantic embedding to obtain the first modality content features. For the captured image and video data, the PHASH perceptual hash algorithm is used to downsample the key frames of the image or video to a preset pixel specification and convert them into grayscale images. Low-frequency coefficients are extracted through discrete cosine transform to generate visual fingerprints. The StyleGAN3 discriminator model is called simultaneously to analyze the local statistical features of the pixel matrix in layers and output the confidence probability value of the content being deepfake to obtain the second modality content features. The first modal content features and the second modal content features are combined to form multimodal content features, which are then encrypted using the SM4 national cryptographic algorithm and transmitted to the next processing unit via a dedicated secure channel.

3. The method for dynamic trustworthiness assessment of social networks based on blockchain fingerprints according to claim 1, characterized in that, The process of generating a blockchain multi-branch fingerprint based on the received encrypted multimodal content features and storing the fingerprints using an improved Merkle Patricia tree structure to obtain a blockchain fingerprint database specifically includes: The encrypted multimodal content features are obtained, and the encrypted multimodal content features are imported into the national cryptographic SM4 decryption engine. The counter sequence is initialized using the pre-allocated security key, and the ciphertext block and the counter are XORed bit by bit. The validity of the timestamp of the HMAC signature is verified synchronously, and finally the decrypted multimodal content features are output. The decrypted multimodal content features are input into the SHA-3 hash generator, and the text semantic vector and visual fingerprint are concatenated to generate the main content summary value. The summary value is then bound to the publisher's biometric information and the timestamp of the Beidou satellite timing system in a triplet to form an initial evidence storage unit. When content modification is detected, the branch fingerprint generation protocol is triggered. The XOR difference engine compares the modified new content feature vector with the original feature vector bit by bit, marks the difference bits, and generates a difference sequence. Obtain the hash value corresponding to the original feature vector, concatenate it with the difference sequence, generate an incremental hash value through SHA-3 operation, and encapsulate it together with the biometric information of the modifier and the Beidou timestamp as the branch identifier core to form a branch fingerprint; After fingerprint generation is completed, all generated fingerprints are stored using a Merkle Patricia tree structure. The Merkle Patricia tree structure is improved by retrieving the parent node of the target fingerprint and appending a new pointer to the PPA. The path compression algorithm is optimized using a Radix tree to merge adjacent single child nodes into extended nodes, ultimately resulting in a blockchain fingerprint database.

4. The method for dynamic trustworthiness assessment of social networks based on blockchain fingerprints according to claim 1, characterized in that, The process of constructing a dynamic directed acyclic graph based on the branch fingerprint data stored in the blockchain fingerprint database, and using this dynamic directed acyclic graph for anomaly propagation detection and source tracing to obtain anomaly propagation detection information, specifically includes: Obtain the blockchain fingerprint database, extract the branch fingerprint structure and the corresponding parent node pointer from the blockchain fingerprint database, and use the adjacency list data structure to treat each branch fingerprint as a graph node and the parent node pointer as a directed edge to construct an initial directed acyclic graph. The initial directed acyclic graph is synchronized to the graph database Neo4j in real time. A unique index is created for each node using the Cypher query language, and propagation relationship modeling is performed to generate a dynamic directed acyclic graph. When a new branch fingerprint is written into the blockchain fingerprint database, the root node is traced back through the parent pointer chain. After verifying that the path is acyclic, the new node and edge are inserted, and the outgoing edge list and in-degree statistics of the adjacency list are updated. The dynamic directed acyclic graph is imported into a pre-trained graph attention network for anomaly propagation detection and source tracing. The node feature vectors are received through the input layer, and a propagation entropy feature extraction layer is set in the graph attention network. In the propagation entropy feature extraction layer, the feature vectors of all parent nodes of each target node are aggregated, and the propagation probability weights from the parent node to the target node are calculated through an attention mechanism. Then, the propagation entropy value of the target node is calculated to characterize the degree of disorder in the propagation path. Based on the pre-defined sliding window statistical entropy change rate, when the absolute change rate of entropy value exceeds the pre-defined threshold within three consecutive windows, the propagation cluster centered on the target node is marked as an abnormal cluster. The initial propagation source node is located by traversing the parent node chain in reverse through the adjacency list, and abnormal propagation detection information is obtained.

5. The method for dynamic trustworthiness assessment of social networks based on blockchain fingerprints according to claim 1, characterized in that, The pre-defined smart contract-based hierarchical response mechanism determines whether intervention is needed based on the credibility assessment information, and introduces a jury mechanism to optimize dispute intervention cases, specifically including: The credibility assessment information is obtained, and the need for regulation is determined by combining the preset smart contract hierarchical response mechanism. When the credibility assessment information is greater than a certain threshold, it is marked as green content and given bandwidth priority and recommendation weighting. When the credibility assessment information is less than the first-class threshold and greater than the second-class threshold, it is marked as yellow content and its propagation rate is limited; when the credibility assessment information is less than the second-class threshold, it is marked as red content, the propagation path is frozen, and the abnormal source node is located and an investigation report is generated simultaneously. While implementing the control strategy, a closed-loop optimization process is initiated. The report data submitted by the user is input into the adversarial example generation module based on DBSCAN density clustering. The similarity of the report case features is measured by Euclidean distance. Density-connected high-conflict samples are aggregated into adversarial training clusters. The generated adversarial examples are input into the federated learning framework, and the model parameter updates are shared in an encrypted state. The jury mechanism is introduced to optimize dispute resolution cases. Based on the user credit score output from the user credit dimension, a number of users whose user credit scores are greater than the preset user credit score threshold are randomly selected to form an arbitration committee. The disputed content is anonymized using zero-knowledge proof technology, and a three-round voting process is conducted to confirm the final resolution. Adversarial examples are generated based on the characteristics of the disputed case, the resolution conclusion, and the resolution plan, so as to drive the federated learning framework to enhance its ability to judge fuzzy cases in the next iteration.

6. A dynamic trustworthiness assessment system for social networks based on blockchain fingerprints, characterized in that, The system includes a content feature extraction module, a blockchain hash generation and storage module, a content dissemination and branch management module, a multimodal artificial intelligence evaluation module, a dynamic reputation evaluation module, a smart contract dynamic control module, and a closed-loop feedback and optimization module. The content feature extraction module is used to parse and extract key structured feature vectors from multimodal content data; The blockchain hash generation and storage module processes the feature vector and content fingerprint into a unique and immutable hash value, and anchors it to the distributed ledger. The content propagation and branch management module is used to dynamically track and record the propagation path and derivative versions of content, and to detect propagation anomalies. The multimodal artificial intelligence evaluation module uses a preset model to automatically perform in-depth detection of the authenticity and tampering traces of the content; The dynamic reputation assessment module is used to dynamically calculate and update the real-time reputation score of users or nodes, and to determine the credibility of the disseminated content. The smart contract dynamic control module is used to execute priority scheduling, rate limiting, or security circuit breaker control strategies for content propagation. The closed-loop feedback and optimization module uses user feedback and dispute arbitration data to generate adversarial training samples, driving the federated learning mechanism to continuously iterate and update the parameters of the multimodal evaluation model.

Images