Policy information crowdsourcing-oriented multi-level verification system and method

Abstract

The invention discloses a policy information crowdsourcing-oriented multi-level verification system and method, and relates to the technical field of data verification. A data source verification module; a content identification module; a distributed consensus verification module; and an abnormal behavior detection module. According to the policy information crowdsourcing-oriented multi-level verification system and method, the problem of data authenticity verification in government policy information crowdsourcing is solved, a four-level cascade protection framework is adopted as the core, and through four-level verification of data source verification, content identification, distributed consensus verification and abnormal behavior detection, the data authenticity verification efficiency is improved. The authenticity and compliance of the government policy information crowdsourcing data are ensured, the efficient, low-delay and anti-cheating system performance is achieved through algorithm optimization, and finally the core targets of 95% verification accuracy, 90% cheating detection rate and the like are achieved.

Description

Technical Field

[0001] This invention relates to the field of data verification technology, specifically to a multi-level verification system and method for policy information crowdsourcing. Background Technology

[0002] Government policy information crowdsourcing refers to the government outsourcing the tasks of collecting, organizing, and analyzing policy information to the public through internet platforms in the form of projects. The public can participate in the process of collecting and responding to policy information, while the government can obtain more diverse and extensive opinions and suggestions through the crowdsourcing platform. The information submitted by the public needs to be verified through a data verification system to remove false and useless data.

[0003] Existing data verification systems are single-dimensional, meaning they only check one aspect, making them susceptible to falsification. Furthermore, manual review is extremely inefficient, resulting in long policy information disclosure times. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a multi-level verification system and method for policy information crowdsourcing, which solves the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-level verification system for policy information crowdsourcing, wherein the multi-level verification system for policy information crowdsourcing includes the following modules:

[0006] The data receiving module is used to receive policy information submitted by users.

[0007] The data source verification module is used to verify the legality and integrity of the data source and filter out forged or tampered source data.

[0008] The content recognition module is used to perform structured analysis on data content to determine whether it conforms to policy information standards;

[0009] The distributed consensus verification module uses multiple verification nodes to collaboratively determine the authenticity of data, thereby reducing the risk of malicious forgery by a single node through a distributed consensus mechanism.

[0010] The abnormal behavior detection module is used to analyze abnormal characteristics of user submission behavior and detect the risks of bot batch submissions and group collaborative cheating.

[0011] A method, applied to the aforementioned multi-level verification system for policy information crowdsourcing, includes the following steps:

[0012] S1, Data Reception:

[0013] The data receiving module receives policy information submitted by users, which includes, but is not limited to, text and images.

[0014] S2. Data source verification:

[0015] The data source verification module processes the received data information sequentially using a domain name whitelist matching algorithm, a digital signature verification algorithm, and a content integrity verification algorithm;

[0016] S3, Content Recognition:

[0017] The distributed consensus verification module cleans the text content to remove special symbols and standardized punctuation, and uses WordPiece segmentation to convert it into an input sequence with a length ≤ 512 tokens. A text classifier based on the BERT model is built, using a 12-layer Transformer structure to classify and verify the policy information, outputting three labels: "policy information," "non-policy information," and "invalid information." The label with the highest probability is taken as the result. If the label is "non-policy information" or "invalid information," it returns "untrustworthy"; if it is "policy information," it proceeds to the next step.

[0018] Regular expressions are used to extract structured data such as time (e.g., \d{4}-\d{2}-\d{2}) and amount (e.g., \d+(\.\d{2})? 10,000 yuan) from the text. Then, a BiLSTM-CRF-based NER model is used to identify the policy subject (e.g., "Ministry of Finance", "XX Municipal People's Government") and the policy type (e.g., "Notice" or "Measures").

[0019] Check if the policy contains the core elements: the issuing entity, the issuance date, the effective date, and a description of the core content. If the date or entity is missing, the policy is deemed invalid. If key information is missing or incorrect, such as an invalid date format, the policy is returned as "unreliable". If the information is complete, proceed to the next step.

[0020] S4, Distributed Consensus Verification:

[0021] The distributed consensus verification module randomly selects at least 7 verification nodes from the node pool, which contains at least 100 nodes. It ensures that the nodes do not overlap in geographical or historical domain. The master node broadcasts the data to the 7 verification nodes. Each node independently verifies the data locally and outputs a "pass" or "reject" vote. Then, based on the voting weight of each node, it calculates whether the number of passing votes is greater than or equal to 50% of the total votes. 50% is the passing threshold, which can be adjusted according to actual needs. If it exceeds the threshold, it proceeds to the next step; otherwise, it returns "untrustworthy."

[0022] S5. Abnormal Behavior Detection:

[0023] The abnormal behavior detection module collects the submission timestamps of users participating in crowdsourcing over the past 30 days to form a time interval sequence for time series anomaly detection. It uses a dynamic time warping algorithm to calculate the distance between the user's time series and "normal user patterns" such as random intervals and high-frequency submissions not in the early morning. The smaller the distance, the more normal it is. A distance threshold is set, such as >0.8. Based on historical data training, if the distance exceeds the threshold, it is judged as a time anomaly, such as the robot's fixed-interval submission.

[0024] Furthermore, in step S2, the specific process of the domain name whitelist matching algorithm is as follows:

[0025] A predefined set of trusted government domain names, such as *.gov.cn and *.gov.com.cn, is stored in a distributed cache such as Redis and supports dynamic updates.

[0026] Parse the source URL from the metadata of user-submitted data, such as:

[0027] https: / / www.mof.gov.cn / zhengce / , extract the subdomain such as mof.gov.cn;

[0028] Using pre-compiled regular expressions such as ^[a-zA-Z0-9-]+\.gov\.cn$, the extracted domain names are matched to determine whether they are in the trusted domain name database, with a time complexity of O(1);

[0029] If the match fails, return "untrusted" directly; if the match succeeds, proceed to the next step.

[0030] The specific process of the digital signature verification algorithm is as follows:

[0031] Request an SSL certificate from the server corresponding to the source URL, and extract the public key (256-bit elliptic curve public key) and digital signature from the certificate.

[0032] The signature validity is verified using a 256-bit elliptic curve cryptography algorithm, and the algorithm flow is as follows:

[0033] Input: Certificate public key, raw data hash value, server signature;

[0034] Calculation: Verify whether the signature was generated by the corresponding private key using elliptic curve equations;

[0035] If signature verification fails, such as due to an expired certificate or forged signature, it will return "Untrusted"; if successful, proceed to the next step.

[0036] The specific process of the content integrity verification algorithm is as follows:

[0037] The SHA-256 algorithm is used to calculate the hash value of the text or image content submitted by the user. If it is text content, it is hashed directly, while if it is an image, the pixel features are extracted and then hashed.

[0038] Retrieve historical snapshots of the content corresponding to the source URL from the system database and extract the historical hash value;

[0039] Compare the current hash value with the historical snapshot hash value to detect whether the content has been tampered with;

[0040] If the hash values ​​do not match, it indicates that the content has been tampered with, and the system returns "untrusted"; if they match, the system passes.

[0041] Furthermore, the equation of the elliptic curve is as follows:

[0042] The server generates a signature (r, s) on the original data hash value h, where r and s are calculated based on the private key d and the curve equation. The private key d is a randomly selected integer, and the public key Q is a point on the curve obtained by the elliptic curve dot product.

[0043] Q = d × G, where G is the base point of the curve and is a predefined generator;

[0044] During verification, the system uses the public key Q, signature (r, s), and hash value h to perform point operations using a curve equation:

[0045] Calculate u1 = h × s -1 (mod.n), u2=r×s -1 (mod.n) yields point P = u1×G + u2×Q, where (mod.n) is the modulo operation;

[0046] If the x-coordinate of P is equal to r, then the signature is valid, meaning the signature was generated by the corresponding private key.

[0047] Furthermore, in step S3, if the data contains images such as scanned copies of policy documents, the text content and image content are separated, and features are extracted separately. The text uses BERT features, and the images use ResNet50 features.

[0048] The text and image features are weighted using an attention mechanism, such as 0.7 for text and 0.3 for images. The weighting is dynamically adjusted based on the complexity of the content. This verifies the consistency between the text description and the image content and outputs a consistency score. If the consistency score is less than 0.6 (which is the threshold), the result is "untrustworthy"; if it is greater than or equal to 0.6, the result is "pass".

[0049] Furthermore, the content complexity index is calculated as follows:

[0050] Text complexity C text :

[0051]

[0052] Keywords are matched using a predefined policy dictionary, and numerical values ​​are extracted using the regular expression \d+(\.\d+)?.

[0053] Image complexity C img :

[0054]

[0055] The more text blocks and the more complex the edges, the richer the image information.

[0056] The dynamic weight formula is as follows:

[0057] The base weights are 0.7 for text and 0.3 for images, which are dynamically adjusted according to complexity. It is necessary to ensure that text α + image β = 1.

[0058] Where max is the maximum complexity value in the system's historical data, used to normalize the complexity of the current mode to the range of 0-1, ensuring the rationality and stability of the weight adjustment;

[0059] If the complexity of text is much higher than that of images, such as plain text policies, α can approach 0.9; if the complexity of images is high, such as scanned copies of stamped documents, β can approach 0.5.

[0060] Semantic consistency verification:

[0061] Weighted text features: T weighted =α+T

[0062] Weighted image features: I weighted =β+I

[0063] Cosine similarity calculation:

[0064] The range is [-1, 1], and the closer it is to 1, the more consistent the semantics.

[0065] Fusion Score:

[0066] Final semantic score: score semantic = 0.6×sim+0.4×P, where P is the consistency probability, which is obtained through the output of the multimodal pre-trained model;

[0067] Numerical consistency verification:

[0068] Text numerical extraction: Numerical values ​​are extracted using the regular expressions \d{4}-\d{2}-\d{2} (time) and \d+\.? \d*ten thousand yuan (amount), forming a set N. text ;

[0069] Image numerical extraction:

[0070] Text-based numerical values ​​are extracted from the OCR recognition results, following the same process as text regularization.

[0071] Chart-based numerical data: A chart parsing model is used to identify bar charts and tables, extracting numerical data to form a set N. img ;

[0072] Consistency judgment:

[0073]

[0074] Among them, the higher the percentage of intersection, the stronger the numerical consistency;

[0075] Scoring Fusion Rule: The final consistency score combines the semantic score and the numerical score, and is weighted using a logistic regression model. The weighting formula is as follows:

[0076] The final score is calculated as w1 × semantic score + w2 × numerical score + b, where w1 and w2 are the weights learned by the model, i.e., w1 + w2 = 1, and b is the bias term.

[0077] Furthermore, in step S4, the voting weight of each node is linked to the node's reputation score. The higher the reputation score, the greater the impact on the result. The initial reputation score of a new node is 0.5, with a maximum score of 1.0. The initial score of historical nodes is calculated based on "historical accuracy × 0.6 + response speed × 0.4", where a response speed ≤ 1 second is recorded as 1.0, and > 5 seconds is recorded as 0.

[0078] After each verification, the results are updated using the exponential decay formula:

[0079] The new score new_score = 0.9 × historical score old_score + 0.1 × current_result, where current_result is the matching identifier between the node's local judgment result and the final global consensus result. A value of 1 indicates that the verification is correct, and a value of 0 indicates that it is incorrect.

[0080] Weighting rules:

[0081] Node voting weight = its own reputation score ÷ the sum of the reputation scores of all nodes. High-reputation nodes, such as those with a historical accuracy rate ≥ 0.9, have a significantly higher weight than low-reputation nodes, such as those with a historical accuracy rate < 0.6.

[0082] Furthermore, in step S5, the abnormal behavior detection module also performs content similarity detection, the specific process of which is as follows:

[0083] Tokenize the text submitted by the user, remove stop words such as "的" and "是", generate a bag-of-words set such as {补贴, 2023, 政策,...}, use the MinHash algorithm to reduce the dimensionality of text features from 1000 dimensions to 128 dimensions, and quickly calculate the Jaccard similarity with more than 100,000 policy information in the system's historical database: Similarity = number of intersection words ÷ number of union words. If the similarity ≥ 0.85 and 0.85 is the preset threshold, it is determined as duplicate content such as a duplicate submission with minor modifications.

[0084] Further, in the step S5, the abnormal behavior detection module also performs user relationship network analysis, and the specific process is as follows:

[0085] Construct a user-user association graph based on the characteristics of user IP address, device ID, and submission content correlation. Use the Node2Vec algorithm to transform user nodes into 128-dimensional vectors. The higher the vector similarity, the closer the user association. Calculate the cosine similarity of the target user and "known malicious users". If ≥ 0.75, it is determined as network anomaly, and the closer to 1, the more similar the user characteristics are to malicious users, so as to judge whether it is gang cheating. 0.75 is the preset threshold.

[0086] Further, in the step S5, based on the calculated time anomaly score S1, content repetition score S2, and network anomaly score S3, perform weighting and summarize the risk score:

[0087] Based on the weights trained from historical data, risk score = 0.4 × S1 + 0.3 × S2 + 0.3 × S3;

[0088] If the risk score ≤ the threshold, it passes, and the final result is "trusted"; otherwise, return "untrusted".

[0089] The present invention provides a multi-level verification system and method for crowdsourcing of policy information, having the following beneficial effects:

[0090] 1. The multi-level verification system and method for crowdsourcing of policy information solve the problem of data authenticity verification in government policy information crowdsourcing. The core adopts a four-tier cascading protection architecture, and verifies layer by layer through four levels: data source verification, content recognition, distributed consensus verification, and abnormal behavior detection. It not only ensures the authenticity and compliance of government policy information crowdsourcing data, but also realizes the system performance of high efficiency, low latency, and anti-cheating through algorithm optimization, and finally achieves core goals such as a verification accuracy rate of 95% and a cheating detection rate of 90%.

[0091] 2. The multi-level verification system and method for policy information crowdsourcing, in the distributed consensus verification module, weights nodes by reputation, giving higher weight to nodes with historically high accuracy rates, thus increasing their weight in the final vote and their influence on the result. This weakens interference from malicious and low-performance nodes, thereby improving data authenticity and reducing the risk of malicious data forgery. Attached Figure Description

[0092] Figure 1 This is a schematic diagram of the steps of the present invention;

[0093] Figure 2-5 This is a table showing the comparison of the performance indicators of each module of the present invention with those of existing technologies, as well as an analysis of their advantages. Detailed Implementation

[0094] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0095] like Figure 1 As shown, the present invention provides a technical solution: a multi-level verification system for policy information crowdsourcing, the multi-level verification system for policy information crowdsourcing comprising the following modules:

[0096] The data receiving module is used to receive policy information submitted by users.

[0097] The data source verification module is used to verify the legality and integrity of the data source and filter out forged or tampered source data.

[0098] The content recognition module is used to perform structured analysis on the data content to determine whether it conforms to the policy information standard, which is prepared in advance.

[0099] The distributed consensus verification module uses multiple verification nodes to collaboratively determine the authenticity of data, thereby reducing the risk of malicious forgery by a single node through a distributed consensus mechanism.

[0100] The abnormal behavior detection module is used to analyze abnormal characteristics of user submission behavior and detect the risks of bot batch submissions and group collaborative cheating.

[0101] A method, applied to the aforementioned multi-level verification system for policy information crowdsourcing, includes the following steps:

[0102] S1, Data Reception:

[0103] The data receiving module receives policy information submitted by users, which includes, but is not limited to, text and images.

[0104] S2. Data source verification:

[0105] The data source verification module processes the received data information sequentially using a domain name whitelist matching algorithm, a digital signature verification algorithm, and a content integrity verification algorithm;

[0106] The specific process of the domain name whitelist matching algorithm is as follows:

[0107] A predefined set of trusted government domain names, such as *.gov.cn and *.gov.com.cn, is stored in a distributed cache such as Redis and supports dynamic updates.

[0108] Parse the source URL from the metadata of user-submitted data, such as:

[0109] https: / / www.mof.gov.cn / zhengce / , extract the subdomain such as mof.gov.cn;

[0110] Using pre-compiled regular expressions such as ^[a-zA-Z0-9-]+\.gov\.cn$, the extracted domain names are matched to determine whether they are in the trusted domain name database, with a time complexity of O(1);

[0111] If the match fails, return "untrusted" directly; if the match succeeds, proceed to the next step.

[0112] The specific process of the digital signature verification algorithm is as follows:

[0113] Request an SSL certificate from the server corresponding to the source URL, and extract the public key (256-bit elliptic curve public key) and digital signature from the certificate.

[0114] The signature validity is verified using a 256-bit elliptic curve cryptography algorithm, and the algorithm flow is as follows:

[0115] Input: Certificate public key, raw data hash value, server signature;

[0116] Calculation: Verify whether the signature was generated by the corresponding private key using elliptic curve equations;

[0117] A point operation is performed using the public key, the hash value of the original data, and the server signature. If the calculation result matches the coordinates of the curve point, the signature is valid; otherwise (e.g., certificate expired, signature forged), "untrusted" is returned. If valid, proceed to the next step.

[0118] The specific process of the content integrity verification algorithm is as follows:

[0119] The SHA-256 algorithm is used to calculate the hash value of the text or image content submitted by the user. If it is text content, it is hashed directly, while if it is an image, pixel features (such as the mean of RGB channels and edge features) are extracted and then hashed.

[0120] Retrieve historical snapshots of the content corresponding to the source URL from the system database and extract the historical hash value;

[0121] Compare the current hash value with the historical snapshot hash value to detect whether the content has been tampered with;

[0122] If the hash values ​​do not match, it indicates that the content has been tampered with, and the system returns "untrusted"; if they match, the system passes.

[0123] The equation of the elliptic curve is as follows:

[0124] The server generates a signature (r, s) on the original data hash value h, where r and s are calculated based on the private key d and the curve equation. The private key d is a randomly selected integer, and the public key Q is a point on the curve obtained by the elliptic curve dot product.

[0125] Q = d × G, where G is the base point of the curve and is a predefined generator;

[0126] During verification, the system uses the public key Q, signature (r, s), and hash value h to perform point operations using a curve equation:

[0127] Calculate u1 = h × s -1 (mod.n), u2=r×s -1 (mod.n) yields point P = u1×G + u2×Q, where (mod.n) is the modulo operation;

[0128] If the x-coordinate of P is equal to r, then the signature is valid, meaning the signature was generated by the corresponding private key.

[0129] S3, Content Recognition:

[0130] The distributed consensus verification module cleans the text content to remove special symbols and standardized punctuation, and uses WordPiece segmentation to convert it into an input sequence with a length ≤ 512 tokens. A text classifier based on the BERT model is built, using a 12-layer Transformer structure to classify and verify the policy information, outputting three labels: "policy information," "non-policy information," and "invalid information." The label with the highest probability is taken as the result. If the label is "non-policy information" or "invalid information," it returns "untrustworthy"; if it is "policy information," it proceeds to the next step.

[0131] Regular expressions are used to extract structured data such as time (e.g., \d{4}-\d{2}-\d{2}) and amount (e.g., \d+(\.\d{2})? 10,000 yuan) from the text. Then, a BiLSTM-CRF-based NER model is used to identify the policy subject (e.g., "Ministry of Finance", "XX Municipal People's Government") and the policy type (e.g., "Notice" or "Measures").

[0132] Check if the policy contains the core elements: the issuing entity, the issuance date, the effective date, and a description of the core content. If the date or entity is missing, the policy is deemed invalid. If key information is missing or incorrect, such as an invalid date format, the policy is returned as "unreliable". If the information is complete, proceed to the next step.

[0133] If the data contains images such as scanned copies of policy documents, the text content and image content are separated, and features are extracted separately. BERT features are used for the text, and ResNet50 features are used for the images.

[0134] The text and image features are weighted using an attention mechanism, such as 0.7 for text and 0.3 for images. The weighting is dynamically adjusted based on the complexity of the content. The consistency between the text description and the image content is verified, and a consistency score is output. If the consistency score is <0.6 (0.6 is the threshold), it is returned as "untrustworthy"; if it is ≥0.6, it is passed.

[0135] The content complexity index is calculated as follows:

[0136] Text complexity C text :

[0137]

[0138] Keywords are matched using a predefined policy dictionary, and numerical values ​​are extracted using the regular expression \d+(\.\d+)?.

[0139] Image complexity C img :

[0140]

[0141] The more text blocks there are and the more complex the edges, the richer the image information. Among them, OCR (Optical Character Recognition) technology is used to process the image and recognize the text content in the image. These recognized text areas form "OCR text blocks".

[0142] The dynamic weight formula is as follows:

[0143] The base weights are 0.7 for text and 0.3 for images, which are dynamically adjusted according to complexity. It is necessary to ensure that text α + image β = 1.

[0144] Where max is the maximum complexity value in the system's historical data;

[0145] If the complexity of text is much higher than that of images, such as plain text policies, α can approach 0.9; if the complexity of images is high, such as scanned copies of stamped documents, β can approach 0.5.

[0146] Semantic consistency verification:

[0147] Weighted text features: Tweighted =α+T

[0148] Weighted image features: I weighted =β+I

[0149] Cosine similarity calculation:

[0150] The range is [-1, 1], and the closer it is to 1, the more consistent the semantics.

[0151] Fusion Score:

[0152] Final semantic score: score semantic = 0.6×sim+0.4×P, where P is the consistency probability, which is obtained through the output of the multimodal pre-trained model;

[0153] Numerical consistency verification:

[0154] Text numerical extraction: Numerical values ​​are extracted using the regular expressions \d{4}-\d{2}-\d{2} (time) and \d+\.? \d*ten thousand yuan (amount), forming a set N. text ;

[0155] Image numerical extraction:

[0156] Text-based numerical values ​​are extracted from the OCR recognition results, following the same process as text regularization.

[0157] Chart-based numerical data: A chart parsing model is used to identify bar charts and tables, extracting numerical data to form a set N. img ;

[0158] Consistency judgment:

[0159]

[0160] Among them, the higher the percentage of intersection, the stronger the numerical consistency;

[0161] Scoring Fusion Rule: The final consistency score combines the semantic score and the numerical score, and is weighted using a logistic regression model. The weighting formula is as follows:

[0162] The final score = w1 × semantic score + w2 × numerical score + b, where w1 and w2 are the weights learned by the model, i.e., w1 + w2 = 1, and b is the bias term;

[0163] S4, Distributed Consensus Verification:

[0164] The distributed consensus verification module randomly selects at least 7 verification nodes from the node pool (the 7-node selection is based on "satisfying Byzantine fault tolerance by tolerating a maximum of 2 malicious nodes"). The node pool contains at least 100 nodes (to ensure randomness and resistance to manipulation). The nodes are ensured to have no overlap in geographical location (covering the five regions of East / South / West / North / Central) and historical domain (such as finance, education, and healthcare). The master node broadcasts the data to the 7 verification nodes. Each node independently verifies the data locally and outputs a "pass" or "reject" vote. Then, based on the voting weight of each node, it calculates whether the number of passing votes is greater than or equal to 50% of the total votes. 50% is the passing threshold, which can be adjusted according to actual needs, such as 60% for sensitive policies. If it exceeds the threshold, proceed to the next step; otherwise, return "untrusted".

[0165] The voting weight of each node is linked to its reputation score. The higher the reputation score, the greater the impact on the result. The initial reputation score of a new node is 0.5, with a maximum score of 1.0. The initial score of historical nodes is calculated based on "historical accuracy × 0.6 + response speed × 0.4", where a response speed of ≤1 second is counted as 1.0 and >5 seconds is counted as 0.

[0166] After each verification, the results are updated using the exponential decay formula:

[0167] The new score new_score = 0.9 × historical score old_score + 0.1 × current_result, where current_result is the matching identifier between the node's local judgment result and the final global consensus result. A value of 1 indicates that the verification is correct, and a value of 0 indicates that it is incorrect.

[0168] Weighting rules:

[0169] Node voting weight = its own reputation score ÷ the sum of the reputation scores of all nodes. High-reputation nodes, such as those with a historical accuracy rate ≥ 0.9, have a significantly higher weight than low-reputation nodes, such as those with a historical accuracy rate < 0.6.

[0170] In the distributed consensus verification module, nodes are weighted by reputation, with nodes having a higher historical accuracy rate having a higher weight. This increases their weight in the final vote and thus their influence on the result. This reduces interference from malicious and low-performance nodes, thereby improving data authenticity and reducing the risk of malicious data forgery.

[0171] S5. Abnormal Behavior Detection:

[0172] The abnormal behavior detection module collects the submission timestamps of users participating in crowdsourcing in the past 30 days, forms a time interval sequence for time series anomaly detection, and uses the dynamic time warping algorithm to calculate the distance between the user time series and the "normal user pattern" such as random intervals and non-early morning high-frequency submissions. The smaller the distance, the more normal it is. A distance threshold is set, such as >0.8 (trained based on historical data, 95% of normal users have a distance <0.8). Based on historical data training, if it exceeds the threshold, it is determined as a time anomaly, such as the fixed interval submission of a robot. Among them, the normal mode is defined as "random intervals (fluctuation range ±50%), non-early morning (0:00 - 6:00) submission frequency <10%".

[0173] The abnormal behavior detection module also conducts content similarity detection. The specific process is as follows:

[0174] The user-submitted text is tokenized and stop words such as "de", "shi" are removed to generate a bag-of-words set such as {subsidy, 2023, policy,...}. The MinHash algorithm is used to reduce the dimensionality of text features from 1000 dimensions to 128 dimensions, and the Jaccard similarity is quickly calculated with the system historical database (more than 100,000 policy information): Similarity = number of intersection words ÷ number of union words. If the similarity ≥ 0.85, and 0.85 is the preset threshold (the repeated detection rate is 95% and the false positive rate <3% at this threshold), it is determined as duplicate content, such as a slightly modified repeated submission;

[0175] The abnormal behavior detection module also conducts user relationship network analysis. The specific process is as follows:

[0176] Based on the features of user IP addresses, device IDs, and submission content correlation, a user-user association graph is constructed. The Node2Vec algorithm is used to transform user nodes into 128-dimensional vectors. The higher the vector similarity, the closer the user association. The cosine similarity of the target user and the "known malicious users" is calculated. If ≥ 0.75, it is determined as a network anomaly, and the closer it is to 1, the more similar the user characteristics are to the malicious users, so as to judge whether it is a group cheating. 0.75 is the preset threshold (the group detection rate is 90% at this threshold). Among them, the Node2Vec algorithm and the calculation of the cosine similarity of the target user and the "known malicious users" are both existing technologies, so no more elaboration will be made here;

[0177] Based on the calculated time anomaly score S1, content repetition score S2, and network anomaly score S3, they are weighted and aggregated into a risk score:

[0178] Based on the weights trained from historical data, the risk score = 0.4×S1 + 0.3×S2 + 0.3×S3;

[0179] If the risk score is less than or equal to the threshold, the system passes and the final result is "trustworthy"; otherwise, it returns "untrustworthy". The current threshold is set to 0.6. The system adjusts the risk threshold in real time based on the gradient descent algorithm to ensure that the false alarm rate is less than 5%.

[0180] Based on the above description, this invention ensures the authenticity and compliance of policy information through four-layer cascaded verification, achieving a verification accuracy rate of 95% and a cheating detection rate of 90%, realizing high efficiency and low latency, supporting large-scale crowdsourcing scenarios, and enhancing anti-cheating capabilities through mechanisms such as node reputation weighting and dynamic thresholds, thereby reducing the risk of malicious nodes and group manipulation.

[0181] Example:

[0182] Data source verification:

[0183] Technical Implementation:

[0184] Domain name whitelist matching algorithm: Based on regular expressions to match trusted domain names (such as *.gov.cn);

[0185] Digital signature verification algorithm: Employs ECDSA (256-bit elliptic curve cryptography), which is 3 times faster than RSA in verification speed;

[0186] Content integrity verification algorithm: The hash value is calculated using SHA-256 and compared with historical snapshots, achieving a tamper detection accuracy of 99.9%.

[0187] Content-aware:

[0188] Technical Implementation:

[0189] Deep learning classification algorithm: based on BERT model (12-layer Transformer), with a text classification accuracy of 92%;

[0190] Key information extraction algorithm: Combining regular expressions and NER technology, with an F1 score of 0.89. The F1 score is an important metric in machine learning used to evaluate the performance of classification models. It combines the balance between precision and recall, and its calculation formula is as follows:

[0191] F1 = 2 × (precision × recall) / (precision + recall);

[0192] The F1 score ranges from 0 to 1, with 0.89 being relatively high, indicating that the model has achieved a good balance between "accurately identifying positive examples" and "finding as many positive examples as possible";

[0193] Multimodal fusion algorithm: text + image collaborative verification, attention mechanism for weight allocation, consistency detection accuracy of 95%;

[0194] Distributed consensus verification:

[0195] Technical Implementation:

[0196] Randomly select at least 7 verification nodes;

[0197] Node reputation weighting algorithm: historical accuracy × 0.6 + response speed × 0.4, exponential decay update to prevent score manipulation;

[0198] Abnormal behavior detection:

[0199] Technical Implementation:

[0200] Time series anomaly detection: Based on the DTW algorithm, the robot behavior recognition accuracy is 88%;

[0201] Content similarity detection: Jaccard coefficient + MinHash acceleration, duplicate content detection rate of 95%;

[0202] User relationship network analysis: Node2Vec graph embedding (128-dimensional vector), 90% detection rate of group cheating;

[0203] Overall risk score: Time anomaly × 0.4 + Content duplication × 0.3 + Network anomaly × 0.3, false alarm rate < 5%;

[0204] Compared with the performance indicators of the above embodiments and the prior art, specifically as follows: Figures 2-5 As shown, compared with existing technologies, this system and method solve the problem of data authenticity verification in government policy information crowdsourcing. The core adopts a four-layer cascaded protection architecture, which verifies the authenticity and compliance of government policy information crowdsourcing data through four layers of verification: data source verification, content recognition, distributed consensus verification, and abnormal behavior detection. This not only ensures the authenticity and compliance of government policy information crowdsourcing data, but also achieves high efficiency, low latency, and anti-cheating system performance through algorithm optimization, ultimately achieving core objectives such as a verification accuracy rate of 95% and a cheating detection rate of 90%.

[0205] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A multi-level verification system for policy information crowdsourcing, characterized in that: The multi-level verification system for policy information crowdsourcing includes the following modules: The data receiving module is used to receive policy information submitted by users. The data source verification module is used to verify the legality and integrity of the data source and filter out forged or tampered source data. The content recognition module is used to perform structured analysis on data content to determine whether it conforms to policy information standards; The distributed consensus verification module uses multiple verification nodes to collaboratively determine the authenticity of data, thereby reducing the risk of malicious forgery by a single node through a distributed consensus mechanism. The abnormal behavior detection module is used to analyze abnormal characteristics of user submission behavior and detect the risks of bot batch submissions and group collaborative cheating.

2. A method applied to a multi-level verification system for policy information crowdsourcing as described in claim 1, characterized in that: The method includes the following steps: S1, Data Reception: The data receiving module receives policy information submitted by users, which includes, but is not limited to, text and images. S2. Data source verification: The data source verification module processes the received data information sequentially using a domain name whitelist matching algorithm, a digital signature verification algorithm, and a content integrity verification algorithm; S3, Content Recognition: The distributed consensus verification module cleans the text content to remove special symbols and standardized punctuation, and uses WordPiece segmentation to convert it into an input sequence with a length ≤ 512 tokens. A text classifier based on the BERT model is built, using a 12-layer Transformer structure to classify and verify the policy information, outputting three labels: "policy information," "non-policy information," and "invalid information." The label with the highest probability is taken as the result. If the label is "non-policy information" or "invalid information," it returns "untrustworthy"; if it is "policy information," it proceeds to the next step. Regular expressions are used to extract structured data such as time (e.g., \d{4}-\d{2}-\d{2}) and amount (e.g., \d+(\.\d{2})? 10,000 yuan) from the text. A BiLSTM-CRF-based NER model is then used to identify policy subjects such as "Ministry of Finance" and "XX Municipal People's Government," and policy types such as "notice" and "measures." Check if the policy contains the core elements: the issuing entity, the issuance date, the effective date, and a description of the core content. If the date or entity is missing, the policy is deemed invalid. If key information is missing or incorrect, such as an invalid date format, the policy is returned as "unreliable". If the information is complete, proceed to the next step. S4, Distributed Consensus Verification: The distributed consensus verification module randomly selects at least 7 verification nodes from the node pool, which contains at least 100 nodes. It ensures that the nodes do not overlap in geographical or historical domains. The master node broadcasts the data to the 7 verification nodes. Each node independently verifies the data locally and outputs a "pass" or "reject" vote. Then, based on the voting weight of each node, it calculates whether the number of passing votes is greater than or equal to 50% of the total votes. 50% is the passing threshold, which can be adjusted according to actual needs. If it exceeds the threshold, it proceeds to the next step; otherwise, it returns "untrustworthy." S5. Abnormal Behavior Detection: The abnormal behavior detection module collects the submission timestamps of users participating in crowdsourcing over the past 30 days to form a time interval sequence for time series anomaly detection. It uses a dynamic time warping algorithm to calculate the distance between the user's time series and "normal user patterns" such as random intervals and high-frequency submissions not in the early morning. The smaller the distance, the more normal it is. A distance threshold is set, such as >0.

8. Based on historical data training, if the distance exceeds the threshold, it is judged as a time anomaly, such as the robot's fixed-interval submission.

3. The method according to claim 2, characterized in that: In step S2, the specific process of the domain name whitelist matching algorithm is as follows: A predefined set of trusted government domain names, such as *.gov.cn and *.gov.com.cn, is stored in a distributed cache such as Redis and supports dynamic updates. Parse the source URL from the metadata of user-submitted data, such as: https: / / www.mof.gov.cn / zhengce / , extract the subdomain such as mof.gov.cn; Using pre-compiled regular expressions such as ^[a-zA-Z0-9-]+\.gov\.cn$, the extracted domain names are matched to determine whether they are in the trusted domain name database, with a time complexity of O(1); If the match fails, return "untrusted" directly; if the match succeeds, proceed to the next step. The specific process of the digital signature verification algorithm is as follows: Request an SSL certificate from the server corresponding to the source URL, and extract the public key (256-bit elliptic curve public key) and digital signature from the certificate. The signature validity is verified using a 256-bit elliptic curve cryptography algorithm, and the algorithm flow is as follows: Input: Certificate public key, raw data hash value, server signature; Calculation: Verify whether the signature was generated by the corresponding private key using elliptic curve equations; If signature verification fails, such as due to an expired certificate or forged signature, it will return "Untrusted"; if successful, proceed to the next step. The specific process of the content integrity verification algorithm is as follows: The SHA-256 algorithm is used to calculate the hash value of the text or image content submitted by the user. If it is text content, it is hashed directly, while if it is an image, the pixel features are extracted and then hashed. Retrieve historical snapshots of the content corresponding to the source URL from the system database and extract the historical hash value; Compare the current hash value with the historical snapshot hash value to detect whether the content has been tampered with; If the hash values ​​are inconsistent, it indicates that the content has been tampered with, and the system returns "untrusted". If everyone agrees, it will pass.

4. The method according to claim 3, characterized in that: The equation of the elliptic curve is as follows: The server generates a signature (r, s) on the original data hash value h, where r and s are calculated based on the private key d and the curve equation. The private key d is a randomly selected integer, and the public key Q is a point on the curve obtained by the elliptic curve dot product. Q = d × G, where G is the base point of the curve and is a predefined generator; During verification, the system uses the public key Q, signature (r, s), and hash value h to perform point operations using a curve equation: Calculate u1 = h × s -1 (mod.n), u2=r×s -1 (mod.n) yields point P = u1×G + u2×Q, where (mod.n) is the modulo operation; If the x-coordinate of P is equal to r, then the signature is valid, meaning the signature was generated by the corresponding private key.

5. A method according to claim 2, characterized in that: In step S3, if the data contains images such as scanned copies of policy documents, the text content and image content are separated, and features are extracted separately. The text uses BERT features, and the images use ResNet50 features. Weigh the text features and image features through an attention mechanism. For example, the text accounts for 0.7 and the image accounts for 0.3, which is dynamically adjusted according to the content complexity. Verify the consistency between the text description and the image content, and output a consistency score. If the consistency score < 0.6 (0.6 is the threshold), then return "untrusted"; if ≥ 0.6, it passes.

6. A method according to claim 5, characterized in that: The content complexity index is calculated as follows: Text complexity C text : Keywords are matched through a predefined policy dictionary, and numerical values are extracted through the regular expression \d+(\.\d+)?; Image complexity C img : The more text blocks and the more complex the edges, the richer the image information; The dynamic weight formula is as follows: The basic weights are 0.7 for text and 0.3 for image, which are dynamically adjusted according to the complexity, and it is necessary to ensure that text α + image β = 1; Where max is the maximum complexity value in the system's historical data; If the text complexity is much higher than that of the image, such as a pure text policy, α can be close to 0.

9. If the image complexity is high, such as a scanned copy of a stamped document, β can be close to 0.5; Semantic consistency verification: Weighted text features: T weighted =α+T Weighted image features: I weighted =β+I Cosine similarity calculation: The range is [-1, 1], and the closer to 1, the more consistent the semantics; Fusion score: Final semantic score: score semantic = 0.6×sim+0.4×P, where P is the consistency probability, which is obtained through the output of the multimodal pre-trained model; Numerical consistency verification: Text numerical extraction: Numerical values ​​are extracted using the regular expressions \d{4}-\d{2}-\d{2} (time) and \d+\.? \d*ten thousand yuan (amount), forming a set N. text ; Image numerical extraction: Numeric values in text: Extracted from the OCR recognition results, following the same regular process as the text; Chart-based numerical data: A chart parsing model is used to identify bar charts and tables, extracting numerical data to form a set N. img ; Consistency judgment: Among them, the higher the intersection ratio, the stronger the numerical consistency; Score fusion rule: The final consistency score combines the semantic score and the numerical score, and is weighted through a logistic regression model. The weighting formula is as follows: Final score = w1 × semantic score + w2 × numerical score + b, where w1 and w2 are the weights learned by the model, that is, w1 + w2 = 1, and b is the bias term.

7. A method according to claim 2, characterized in that: In step S4, the voting weight of each node is linked to the node's reputation score. The higher the reputation score of a node, the greater its influence on the result. The initial reputation score of a new node is 0.5, with a full score of 1.

0. The initial score of a historical node is calculated based on "historical accuracy rate × 0.6 + response speed × 0.4". Among them, if the response speed ≤ 1 second, it is recorded as 1.0, and if > 5 seconds, it is recorded as 0; After each verification, it is updated using the exponential decay formula: New score new_score = 0.9 × historical score old_score + 0.1 × current_result, where current_result is the matching identifier between the node's local judgment result and the final global consensus result. If it is 1, it means that this verification is correct; if it is 0, it means it is wrong; Weight allocation rule: Node voting weight = its own reputation score ÷ the total sum of all nodes' reputation scores. High-reputation nodes, such as those with a historical accuracy rate ≥ 0.9, have a significantly higher weight than low-reputation nodes, such as those < 0.

6.

8. A method according to claim 2, characterized in that: In step S5, the abnormal behavior detection module also conducts content similarity detection. The specific process is as follows: The text submitted by the user is tokenized and stop words such as "de" (of), "shi" (is) are removed to generate a bag-of-words set such as {subsidy, 2023, policy,...}. The MinHash algorithm is used to reduce the dimensionality of the text features from 1000 dimensions to 128 dimensions, and quickly calculate the Jaccard similarity with more than 100,000 policy information in the system's historical database: Similarity = number of intersection words ÷ number of union words. If the similarity ≥ 0.85 (0.85 is the preset threshold), it is determined as duplicate content, such as a slightly modified duplicate submission.

9. A method according to claim 8, characterized in that: In step S5, the abnormal behavior detection module also performs user relationship network analysis, the specific process of which is as follows: A user-user association graph is constructed based on the features of user IP address, device ID, and relevance of submitted content. The Node2Vec algorithm is used to transform user nodes into 128-dimensional vectors. The higher the vector similarity, the closer the user association. The cosine similarity of the target user and the "known malicious user" vector is calculated. If it is ≥0.75, it is judged as a network anomaly. The closer it is to 1, the more similar the user features are to the malicious user, thus judging whether it is a group cheating. 0.75 is the preset threshold.

10. A method according to claim 9, characterized in that: In step S5, the calculated time anomaly score S1, content duplication score S2, and network anomaly score S3 are weighted and summed to form a risk score: The risk score, calculated based on weights trained using historical data, is calculated as follows: Risk Score = 0.4 × S1 + 0.3 × S2 + 0.3 × S3. If the risk score is less than or equal to the threshold, the application passes and the final result is "trustworthy"; otherwise, it returns "untrustworthy".

Images