Campus second-hand commodity transaction credible traceability evaluation method and system

Abstract

The application provides a campus second-hand commodity transaction credible traceability evaluation method and system, relates to the electric digital data processing technical field, and fundamentally improves the accuracy and foresight of the credibility evaluation through the closed-loop linkage control system constructed from multi-dimensional feature deep quantification to individual-group correlation comparison, and then to adaptive strategy dynamic decision-making; The beneficial effects brought about include: 1) the initial trust evaluation ability of new users without historical data is enhanced; 2) the identification accuracy of hidden fraud behaviors with "mode anomaly" characteristics using high credit identity as a cover is improved; 3) the dynamic self-adaptation of risk intervention strategy is realized, which can intelligently adjust the control strength according to the overall risk situation of the platform, and achieves a better balance between ensuring platform safety and optimizing user experience.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically to a reliable traceability assessment method and system for second-hand goods transactions on campus. Background Technology

[0002] With the deepening of the digital economy, peer-to-peer (P2P) information exchange platforms based on user-generated content (UGC) are becoming increasingly popular, covering multiple fields such as social networking, e-commerce, and local services. In these platforms, the credibility of information is the cornerstone of maintaining a healthy ecosystem and user activity. Therefore, developing computational methods capable of accurately, efficiently, and automatically assessing the credibility of massive, heterogeneous, and dynamically generated user information has become a cutting-edge technological trend of common concern in the fields of data science and cyberspace governance, and is of great significance for improving platform operational efficiency and protecting user rights.

[0003] In specific application scenarios such as secondhand goods transactions on campus, existing technologies face the following main challenges in achieving efficient and accurate credibility assessment:

[0004] 1. Traditional credibility assessment methods heavily rely on users' historical behavioral data, such as transaction records and historical reviews. This model struggles to form an effective initial trust judgment for newly registered users or users with occasional transactions who lack historical data. Furthermore, this assessment method, which relies on aggregated historical data, is inherently lagging, and its risk warning capability for carefully designed fraudulent activities suddenly committed using long-accumulated high-credit identities needs improvement.

[0005] 2. Disjointed Analysis of Content and Source: Existing technologies treat the review of published content and the credit assessment of the publisher as two separate processes. For example, content review may employ keyword filtering and image similarity comparison, while source assessment analyzes historical transaction ratings. This disjointed analytical framework makes it difficult to capture deep anomalies in the correlation between "source" and "content." For instance, a seasoned user who should be posting about high-quality digital products (a trusted source) might suddenly post about items completely unrelated to their field of expertise with unusually high prices (abnormal content). The risks inherent in this "source-content" mismatch are easily overlooked under this disjointed analytical model.

[0006] 3. Existing risk intervention measures are mostly based on fixed rule threshold systems. For example, Chinese patent CN120316182A discloses a data analysis method that, while achieving the integration and analysis of multi-source data to support operational decisions, its risk response logic (such as fault detection and dynamic pricing) is essentially still based on preset rule responses. Such rigid systems lack adaptability when facing complex and ever-changing fraud methods and dynamically changing overall platform risk situations. The system cannot automatically adjust the tightness of its intervention strategy according to the global risk level, making it difficult to achieve a dynamic balance between ensuring security and optimizing user experience.

[0007] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to provide a reliable traceability assessment method and system for second-hand goods transactions on campus, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A method for credible traceability assessment of secondhand goods transactions on campus, comprising the following steps:

[0011] S1: Receive and verify a trusted entity credential based on zero-knowledge proof associated with the user's identity, wherein the trusted entity credential is used to prove that the user belongs to a preset authoritative entity;

[0012] S2: Obtain multimodal product information to be evaluated published by the user;

[0013] S3: Based on the data correlation within the multimodal product information, calculate the individual information consistency index to quantify the degree of internal logical self-consistency;

[0014] S4: Based on the trusted entity credentials, determine the group behavior baseline associated with the authoritative entity, and in conjunction with the individual information consistency index, calculate the credential content deviation degree, which characterizes the degree of deviation between the user's behavior and the behavior of the group to which they belong;

[0015] S5: Based on the deviation of the voucher content, generate a credibility assessment result for the multimodal product information and the user.

[0016] A campus secondhand goods transaction credibility traceability assessment system, the system being used to execute the campus secondhand goods transaction credibility traceability assessment method, comprising:

[0017] Credential verification and data acquisition module: used to receive and verify trusted entity credentials based on zero-knowledge proofs associated with the user's identity, wherein the trusted entity credentials are used to prove that the user belongs to a preset authoritative entity;

[0018] Obtain multimodal product information to be evaluated, published by the user;

[0019] Multidimensional feature deep analysis module: used to calculate an individual information consistency index that quantifies the degree of internal logical self-consistency based on the data correlation within the multimodal product information;

[0020] Individual-Group Deviation Measurement Module: Based on the trusted entity credentials, it determines the baseline of group behavior associated with the authoritative entity, and in conjunction with the individual information consistency index, calculates the credential content deviation degree, which characterizes the degree of deviation between user behavior and the behavior of their group.

[0021] Reinforcement learning adaptive decision-making module: used to generate a credibility assessment result for the multimodal product information and the user based on the deviation degree of the voucher content.

[0022] Compared with existing technologies, the beneficial effects of this invention are: by introducing trusted entity credentials based on zero-knowledge proofs, and while verifying their basic validity, further quantitatively analyzing the maturity and continuity of their credential lifecycle in distributed timestamp records; a more in-depth dynamic source profile is established for all users, transcending the platform's internal behavior. Simultaneously, by introducing the calculation of entropy in the content generation process, potential fraudulent intent can be captured from the micro-behavioral level of information creation, realizing a shift from static content review to dynamic intent prediction.

[0023] By constructing a voucher content deviation degree to jointly form a multi-dimensional behavioral feature vector, and employing the Mahalanobis distance algorithm to measure the degree of deviation of the individual's behavioral vector from the behavioral baseline of its group in the multi-dimensional covariance space, this method can accurately identify concealed fraudulent behaviors that appear normal in a single dimension but exhibit abnormal combination patterns, achieving a fundamental shift from single-information decision-making to multi-information collaborative decision-making.

[0024] By introducing a reinforcement learning-driven adaptive risk management mechanism, the calculated voucher content deviation, user dynamic trust score, and overall platform risk situation are used as decision states. The reinforcement learning model dynamically and adaptively selects and triggers the optimal risk intervention strategy, enabling risk management to evolve from a "rule-based" passive response to a "learning-based" predictive and adaptive proactive management, ensuring the robustness and advancement of the assessment system in long-term operation. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall application process of the present invention;

[0026] Figure 2 This is a schematic diagram illustrating the execution logic of the "Individual Information Consistency Index" of the present invention;

[0027] Figure 3 This is a schematic diagram illustrating the execution logic of the "Comprehensive Credibility of Credentials" in this invention;

[0028] Figure 4 This is a schematic diagram illustrating the execution logic of steps S4 and S5 of the present invention. Detailed Implementation

[0029] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0030] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0031] Example 1:

[0032] Please see Figures 1 to 4 The present invention provides a technical solution:

[0033] A method for credible traceability assessment of secondhand goods transactions on campus includes the following steps:

[0034] S1: Receive and verify a trusted entity credential based on zero-knowledge proof associated with the user's identity, wherein the trusted entity credential is used to prove that the user belongs to a preset authoritative entity;

[0035] S2: Obtain multimodal product information to be evaluated published by the user;

[0036] S3: Based on the data correlation within the multimodal product information, calculate the individual information consistency index to quantify the degree of internal logical self-consistency;

[0037] S4: Based on the trusted entity credentials, determine the group behavior baseline associated with the authoritative entity, and in conjunction with the individual information consistency index, calculate the credential content deviation degree, which characterizes the degree of deviation between the user's behavior and the behavior of the group to which they belong;

[0038] S5: Based on the deviation of the voucher content, generate a credibility assessment result for the multimodal product information and the user.

[0039] Further explanation: Verifying the trusted entity credential specifically includes: using a preset public key associated with the authoritative entity, performing a zero-knowledge proof verification algorithm on the trusted entity credential to confirm the credential validity status of the trusted entity credential without obtaining the user's specific identity information;

[0040] After verifying the trusted entity credentials, the process further includes:

[0041] Based on the distributed timestamp records of the trusted entity credentials, calculate the credential maturity, which characterizes its historical length, and the credential continuity, which characterizes the stability of its activity frequency.

[0042] The overall credibility of a voucher is obtained by nonlinearly fusing the voucher validity status, voucher maturity, and voucher continuity.

[0043] Further explanation: The specific indicators for calculating individual information consistency include:

[0044] Calculate the image-text concept alignment degree, which represents the semantic matching degree between the image and the text;

[0045] Calculate the price-value deviation coefficient, which represents the degree of deviation between the listed price and the predicted fair value of a commodity.

[0046] Calculate the entropy of the content generation process that generates abnormal behaviors during the creation of quantitative information.

[0047] Further explanation: The entropy of the content generation process includes: based on the behavioral logs of users when creating multimodal product information, jointly quantifying the source attributes of the image information and the stability of the editing process of the text information;

[0048] The text-image concept alignment, price-value deviation coefficient, and content generation process entropy are used as inputs, and nonlinear fusion is performed through a fuzzy logic reasoning system to obtain the individual information consistency index.

[0049] Further explanation: The fuzzy logic-based reasoning system includes a fuzzy rule base, which automatically optimizes and generates rules for the fuzzy rule base by training machine learning on historically labeled fraudulent and normal sample data;

[0050] By analyzing the metadata of user-uploaded images and logging editing behavior during the text input process, and applying Shannon entropy calculation logic, the entropy of the content generation process is obtained.

[0051] The following are specific implementation instructions for the above content:

[0052] The core technical feature of this embodiment lies in its multi-dimensional trust quantification method, which combines credential lifecycle analysis with content generation process entropy. This method, through zero-knowledge proof cryptography, verifies the fundamental validity of the user's trusted entity credentials while protecting user privacy. It also introduces a quantitative analysis of the maturity and continuity of the credential within a traceable distributed timestamp record, thereby conducting an in-depth assessment of the source's trustworthiness. Simultaneously, this method not only evaluates the static content consistency of multimodal product information published by users in terms of image-text semantics and price value, but also introduces the analysis of micro-behavioral data during the content generation process to quantify its content generation process entropy, revealing the potential intent behind information dissemination. Finally, through a non-linear multi-parameter fusion mechanism, the multi-dimensional evaluation results of source trustworthiness and content trustworthiness are integrated to output a comprehensive trust quantification index. In this embodiment, the key parameters are defined as follows:

[0053] The validity status of the voucher is indicated by the following parameter symbol: It is a binary logical value used to characterize whether the trusted entity credentials provided by the user have passed cryptographic verification; the specific way to obtain it is by executing the output of the zero-knowledge proof verification algorithm based on the zk-SNARKs protocol; if the verification algorithm returns "true", the parameter is determined to be the value 1; if it returns "false", it is determined to be the value 0.

[0054] Certificate maturity, its parameter symbol is Its value range is a normalized value in the interval [0, 1], used to quantify the time span of a trusted entity's certificate since its first issuance, reflecting its historical length. Its acquisition and determination method is as follows: The calculation of this parameter relies on an immutable distributed ledger of certificate activities associated with the authoritative entity, which records the first issuance timestamp of each certificate; its calculation logic is: obtain the time difference between the current timestamp and the first issuance timestamp of the certificate to obtain the certificate's duration. Then, the certificate's duration is input into a preset, monotonically increasing normalization function for processing. This normalization function is a variant of the logistic growth function, expressed as: the value of certificate maturity is equal to 1 divided by "1 plus the exponent of the natural constant e", and the product of the certificate's duration (negative maturity) and a preset time scale factor;

[0055] The time scale factor is a positive number used to adjust the speed at which vouchers "mature." When the voucher's duration equals the time scale factor, the voucher maturity reaches 0.73; in this campus transaction scenario, the time scale factor is set to 180 days. This means that vouchers with a duration of six months are considered to have reached a high level of maturity. This parameter can be calibrated and adjusted by the platform operator based on actual business data.

[0056] Document continuity, its parameter symbol is Its value range is a normalized value in the range [0, 1], used to quantify the stability and persistence of the activity frequency of a voucher over a period of time. Its acquisition and determination method is as follows: the computing device obtains all activity timestamps of the voucher within a preset evaluation window from the voucher activity distributed ledger, forming a timestamp sequence; in this embodiment, the preset evaluation window is set to 365 days.

[0057] The difference between all adjacent timestamps in the timestamp sequence is calculated to obtain a sequence composed of multiple time intervals. If the sequence has fewer than two elements, stability cannot be calculated, and the maximum silence duration is used directly. The arithmetic mean and standard deviation of this time interval sequence are calculated; then, the coefficient of variation (CV) is obtained by dividing the standard deviation by the mean. The smaller the CV value, the more stable the activity frequency. The silence duration (in days) is obtained by taking the difference between the current timestamp and the latest timestamp in the timestamp sequence; document continuity. The value of is calculated using an exponential decay function, expressed as: the exponent of the natural constant e, which is negative, representing the sum of the activity frequency stability penalty and the silence duration penalty. The activity frequency stability penalty equals the sum of the coefficient of variation (CV) and the preset stability weighting factor. The product of the two factors. The silence duration penalty term is equal to "the silence duration divided by the normalized silence duration obtained from the preset evaluation window, and then multiplied by the preset silence weight factor". The product of [the product of the two]. In this embodiment, the weighting factor is... and Both are set to 0.5 to balance the severity of punishment for the two types of negative behaviors, and the specific values ​​are determined through historical data analysis.

[0058] Price-value deviation coefficient, its parameter sign is Its value range is a normalized value in the range [0, 1], used to quantify the deviation between the listed price of a commodity and its fair market value; its acquisition and determination method is: based on a pre-trained price regression model, which is obtained through supervised learning training on massive historical transaction data; in this embodiment, the price regression model is a gradient-boosting-decision-tree model; the calculation logic is: extracting structured features from multimodal commodity information, including but not limited to commodity category, brand, model, newness of declaration, geographical location of publication, etc., and using these features as inputs to feed into the price regression model to obtain the predicted fair value; then, obtaining the actual listed price of the commodity. The "price-value deviation coefficient" is equal to the absolute value of the difference between the actual listed price and the predicted fair value, divided by the predicted fair value. This ratio is further processed using a peak-shaving function; specifically, for all ratios greater than or equal to 1, the final value is set to 1 to map the calculation result to the [0, 1] interval. In this embodiment, for the selected second-hand mobile phone, its structured features are input into the price regression model, resulting in a predicted fair value of 2000 yuan. The user's actual listed price is 2500 yuan. The absolute value of the difference is 500 yuan, and dividing this value by the predicted fair value yields an initial deviation of 0.25. Since 0.25 is within the [0, 1] interval, the final determined price-value deviation coefficient is 0.25.

[0059] Entropy in the content generation process, with parameter symbol as Its value range is a normalized value in the range [0, 1], used to quantify the abnormality or uncertainty of user behavior during the creation of product information; its acquisition and determination method is: based on user micro-behavior logs collected by the front-end device, and quantified by a calculation model based on the concept of Shannon-Entropy in information theory; the front-end device in this embodiment includes a mobile app or a web browser; the calculation model in this embodiment comprehensively evaluates the behavior in the following two sub-dimensions:

[0060] Image source uncertainty: Analyze the EXIF ​​metadata of uploaded images; if an image contains multiple editing software modification records or its creation time and upload time interval are abnormal or camera device information is missing, it is classified as a "high uncertainty" source; conversely, if it is a direct output from the device and has no modification records, it is classified as a "low uncertainty" source; according to the preset rule base, different probability values ​​are assigned to different source states;

[0061] Text input instability: Record the total number of characters, the number of characters deleted or modified, the total input time, and the number of pauses when the user inputs descriptive text, and calculate the "editing rate" and "input flow level"; where "editing rate" represents "number of characters deleted or modified / total number of characters"; "input flow level" represents "total number of characters / total input time";

[0062] Substituting the states and probabilities of the aforementioned behavioral events into the Shannon entropy calculation logic, we multiply the probability of each event state by its base-2 logarithm, then sum all the results and take the negative value. Finally, the calculated entropy value is mapped to the [0, 1] interval using the max-min normalization method, thus obtaining the content generation process entropy. The maximum and minimum entropy values ​​are obtained through offline statistical calibration of a large number of normal user behavior samples.

[0063] In this embodiment, a user uploaded an image that had been processed by editing software three times, with a probability of 0.7 indicating high uncertainty. The user also entered a 500-word description, of which 150 words were modified, with a probability of 0.6 indicating a high editing modification rate. Based on the probability distribution of these two event states, the calculated entropy value is 1.3. The example demonstrates that offline calibration reveals the entropy range for a normal user to be [0.2, 1.5]. Therefore, using the max-min normalization method, the final entropy for the content generation process is determined to be (1.3-0.2) / (1.5-0.2), resulting in 0.85.

[0064] Image-text concept alignment, its parameter symbol is: The acquisition method involves obtaining a pre-trained multimodal model from the publicly available machine learning model library Hugging-FaceHub, deploying the model and its associated image transformer and text segmenter processor on a computing device, uniformly adjusting the size of user-uploaded product images to 224×224 pixels, converting the image pixel values ​​from the range [0, 255] to tensor format, and finally normalizing the image tensors based on the mean and standard deviation used during model pre-training. Preprocessing is then performed on the user-input product description text. This preprocessing includes: calling the text segmenter to convert the text string into a sequence of integer IDs; then truncating or padding the sequence to achieve the model's fixed input length of 77 words; inputting the preprocessed image tensor and text ID sequence into the deployed CLIP model's image encoder and text encoder, respectively, performing forward propagation calculations to obtain 512-dimensional image feature vectors and 512-dimensional text feature vectors; and finally performing L2 norm normalization on the obtained image and text feature vectors. Calculate the dot product between the two normalized vectors. Since the vectors have been normalized, the result of their dot product is the cosine similarity between the two vectors, with a range of [-1, 1]. Perform a linear transformation on the cosine similarity to map it to the [0, 1] interval required in this embodiment: Image-text concept alignment. The value is equal to "the cosine similarity value plus 1, and then the result is divided by 2".

[0065] Individual information consistency index, its parameter symbol is Its value range is a comprehensive index in the range [0, 1]. The higher the value, the greater the internal contradiction of the information and the lower the credibility. It is obtained by nonlinearly fusing three sub-dimensional indicators: the alignment degree of graphic and textual concepts, the price-value deviation coefficient, and the entropy of the content generation process.

[0066] The reliability of the comprehensive voucher is indicated by the parameter symbol: Its value range is a normalized value in the interval [0, 1], used to comprehensively evaluate the validity, historical length, and activity stability of trusted entity credentials; it measures the validity status of the credentials. Voucher maturity and document continuity The credibility of the comprehensive certificate is obtained by non-linear fusion of these three sub-dimensional indicators. The value is equal to the "credential validity status". The value multiplied by the certificate maturity level The value is used as a preset maturity weighting factor. The result of the exponentiation operation and the continuity of the voucher The value is assigned a preset continuity weight factor. The product of the results of exponentiation. Maturity weighting factor in this embodiment. and continuity weighting factor Both are positive numbers, and their sum is 1, used to adjust the relative importance of maturity and continuity in the overall assessment. This embodiment sets [the following values] in a scenario that emphasizes the historical accumulation of vouchers. Specifically, by performing logistic regression analysis on the characteristics of historical fraudulent and legitimate vouchers, the optimized calibration is obtained from the model's coefficients. This embodiment is executed by a computing device, and the specific steps are as follows:

[0067] 1.1) The initial input includes: a trusted entity certificate submitted by the user front-end that is associated with a product posting behavior, public input, multimodal product information (including images, text and prices), and user micro-behavior logs collected during the posting process.

[0068] 1.2) Execute step S1; the computing device calls the pre-stored verification key and executes the zk-SNARKs verification algorithm on the received trusted entity credential proof and public input. If the output of the verification algorithm is "false", the credential validity status is determined. If the value is 0, the process terminates and a verification failure instruction is returned to the front end. If the value is "true", the validity status of the credential is determined. If the value is 1, the process continues.

[0069] 1.3) After successful credential verification, the computing device uses the unique identifier contained in the public input of the credential; in the calculation process of this embodiment, the credential maturity level is output. and document continuity The validity status of the credentials and the maturity of credentials and document continuity The credibility of integrated credentials Finally, calculate the individual information consistency index. .

[0070] For individual information consistency index Fusion: This embodiment selects a fusion method based on a fuzzy logic inference system. By "fuzzifying" precise input values ​​through membership functions, and then using IF-THEN linguistic rules for inference, it can simulate nonlinear, experience-based decision-making processes. This can more effectively capture the synergistic risk effects arising from the co-occurrence of risks across multiple dimensions. The construction of the fuzzy logic inference system is as follows:

[0071] Alignment of textual and graphical concepts with input parameters Price-Value Deviation Coefficient and content generation process entropy and the consistency index of individual information of output parameters. Each fuzzy set and membership function is defined separately. This example demonstrates the alignment of text and image concepts. Define three fuzzy sets: {low, medium, high}, and use a trapezoidal or Gaussian function to define the membership degree of each value from the [0, 1] interval to these three fuzzy sets; specifically, for each input parameter (image-text concept alignment degree)... Price-Value Deviation Coefficient and content generation process entropy Each fuzzy set {low, medium, high} predefines a specific membership function. (Alignment of text and image concepts) For example, its membership function is defined as follows:

[0072] The "low" fuzzy set is defined by four inflection point parameters [0, 0, 0.2, 0.4]. It represents the alignment degree between the text and image concepts. When the value is between 0 and 0.2, its membership degree is 1; when it is between 0.2 and 0.4, the membership degree decreases linearly from 1 to 0; when it is greater than 0.4, the membership degree is 0.

[0073] The "middle" fuzzy set is defined by four inflection point parameters [0.2, 0.4, 0.6, 0.8]. It represents the alignment degree between the text and image concepts. When the value is between 0.2 and 0.4, the membership degree increases linearly from 0 to 1; when it is between 0.4 and 0.6, the membership degree is 1; and when it is between 0.6 and 0.8, the membership degree decreases linearly from 1 to 0.

[0074] The "high" fuzzy set is defined by four inflection point parameters [0.6, 0.8, 1, 1]. It represents the alignment degree between the text and image concepts. When the value is between 0.6 and 0.8, the membership degree increases linearly from 0 to 1; when it is greater than or equal to 0.8, the membership degree is 1. This embodiment calculates the image-text concept alignment degree. The value is 0.7. Based on the membership function defined above:

[0075] Its membership level of "low" is 0. Its membership level of "medium" is calculated using the linear relationship as (0.8-0.7) / (0.8-0.6), resulting in 0.5. Its membership level of "high" is calculated using the linear relationship as (0.7-0.6) / (0.8-0.6), resulting in 0.5. Therefore, The system interprets both concepts as "medium" and "high" with a level of 0.5. Based on these two activated fuzzy concepts, the fuzzy inference engine will trigger all relevant IF-THEN rules in parallel.

[0076] For the establishment of the fuzzy rule base: a set of "IF-THEN" rules is established by labeling with domain expert knowledge, as follows:

[0077] During the initial system deployment, a basic fuzzy rule base, manually defined by anti-fraud experts based on their experience, will be built-in. This ensures that the system has basic judgment capabilities even during the cold start phase.

[0078] During system operation, samples of product information clearly labeled "fraudulent" or "normal" are continuously collected. Using this sample data, the fuzzy rule base is trained offline or periodically online via an Adaptive Neural Fuzzy Inference System (ANFIS) or a genetic algorithm. Optimization includes: fine-tuning the inflection point parameters of the membership function to better reflect the distribution of real data; adjusting the weight or confidence factor of each rule; and automatically identifying and adding or deleting redundant or inefficient rules in extreme cases.

[0079] Rule 1: IF( (for low) AND ( (for high) THEN ( (For high).

[0080] Rule 2: IF (for high) AND ( (for the middle) THEN ( (For medium to high levels).

[0081] Rule 3: IF( (for high) AND ( (for low) AND ( (for low) THEN ( (Low).

[0082] The weights of the rule base are automatically optimized and adjusted through offline analysis of a large amount of labeled fraudulent / normal sample data, using machine learning genetic algorithms or ant colony optimization algorithms to achieve the best classification results. This embodiment uses the Mamdani fuzzy inference model. For any given set of input values, the system calculates the trigger strength of each rule and obtains a fuzzy output result based on the conclusion of the rule.

[0083] The fuzzy output is converted into a numerical value in the interval [0, 1]. This embodiment uses the centroid-method to calculate the geometric center of the area covered by the output fuzzy set. The x-coordinate of this center is the final individual information consistency index. The specific implementation steps of the center-of-gravity method described above are as follows:

[0084] The output variable is the individual information consistency index. The domain [0, 1] is discretized into N equally spaced sampling points. N = 101 points are selected, each labeled xi = {0, 0.01, 0.02, ..., 1.0}. After fuzzy inference, the union of all triggered rule conclusions is taken, i.e., the maximum membership degree is taken, to obtain the final membership degree value at each sampling point xi, denoted as μ(xi). The set of μ(xi) constitutes the final output fuzzy set graph. The final individual information consistency index... The value is equal to "for all sampling points xi from 0 to 1, calculate the product of its value and the membership degree μ(xi) of that point, then sum all these products together," and then "divide this sum by the sum of the membership degrees μ(xi) of all sampling points." The following is a detailed implementation description of the above: The calculation process of this embodiment ultimately generates two core output indicators: Comprehensive Certificate Credibility. Consistency index with individual information .

[0085] When the credibility of the comprehensive certificate The closer the value is to 1, the better the system's overall performance in judging the authenticity of the user's credentials, historical accumulation, and activity stability, and the higher the prior credibility of the information source.

[0086] When the credibility of the comprehensive certificate The closer the value is to 0, the more serious the defect in the credential and the lower the prior credibility of the source. In this embodiment, "defect" represents verification failure, new registration with no activity, sudden activity after a long period of silence, etc.

[0087] When the individual information consistency index The closer the value is to 1, the more serious the internal contradictions and anomalies the system judges in the product information published by the user in various dimensions, the lower the credibility of the content itself, and the higher the risk of fraud; in this embodiment, "various dimensions" represents multiple dimensions such as image and text matching, pricing logic, and generation behavior.

[0088] When the individual information consistency index The closer a value is to 0, the more self-consistent the product information content is across all dimensions, the more normal the behavioral patterns are, and the higher the credibility of the content itself.

[0089] When other parameters remain unchanged, the certificate maturity The increase will lead to a decrease in the credibility of composite vouchers. Monotonically increasing; characterized by a positive correlation;

[0090] When other parameters remain unchanged, the continuity of the certificate The increase will lead to a decrease in the credibility of composite vouchers. Monotonically increasing; this indicates a positive correlation. The design aims to penalize credentials with anomalous activity patterns; credentials with consistent, stable activity are more reliable than those that are long-term inactive or have irregular activity.

[0091] When other parameters remain unchanged, the alignment of the graphic and textual concepts The increase in this will lead to an increase in the individual information consistency index. Monotonically decreasing; this indicates a negative correlation. (Image and text concept alignment) A higher value indicates a better match between the text and images, and a more consistent content, thus representing an individual information consistency index that reflects risk. The lower. In fuzzy logic reasoning systems, "( (for high) THEN ( The rule of "lower (for example)" ensures this negative correlation trend, which is consistent with the intuitive judgment that "the more relevant the content is, the more credible it is";

[0092] When other parameters remain constant, the price-value deviation coefficient The increase will lead to an increase in the individual information consistency index. Monotonically increasing. This indicates a positive correlation. Price-Value Deviation Coefficient The higher the price, the more significant the deviation between the listed price and the fair value, according to fuzzy logic rules. (for high) THEN ( The high (high) ensures this positive correlation, accurately reflecting the common sense that "the more outrageous the price, the higher the risk".

[0093] When other parameters remain unchanged, the entropy of the content generation process The increase will lead to an increase in the individual information consistency index. Monotonically increasing. This is a positive correlation. Entropy in the content generation process. The abnormality of content creation behavior was quantified. For fraudulent information, its generation process, including "using stolen images online and repeatedly modifying text," exhibits higher uncertainty and complexity in behavioral dynamics, i.e., higher content generation process entropy. This positive correlation design enables this embodiment to capture fraudulent intent at the behavioral level.

[0094] To verify the effectiveness of this embodiment in identifying different types of risk information, six typical user posting scenarios were designed below. These scenarios cover various situations from normal users to different fraud methods, aiming to systematically test the role and beneficial effects of the two core innovations in this embodiment: credential lifecycle analysis and multi-dimensional content consistency analysis. The experiment quantifies the technological progress of this embodiment by comparing the results of the "individual information consistency index" output by the method of this embodiment with the "baseline content risk score" output by the baseline method that only relies on static content analysis.

[0095] Baseline content risk score: A risk assessment indicator used for comparison, with a value range of [0, 1]. This indicator is obtained by linearly weighting and fusing the alignment of text and image concepts and the price-value deviation coefficient, and represents the existing technical level of voucher lifecycle analysis, content generation process entropy, and fuzzy logic fusion, which do not include the core innovation points of this embodiment.

[0096]

[0097] Comparing Scenario 1 (normal experienced user) and Scenario 3 (image theft and low-price traffic generation): In Scenario 1, all input parameters are normal. The average value of the "Individual Information Consistency Index" output by this embodiment is 0.09, while the average value of the baseline method is 0.04, both of which can be identified as low risk. However, in Scenario 3, fraudsters use newly registered credentials (average credential maturity 0.205) to publish information with severely mismatched images and text (average image-text alignment 0.425), low prices (average price deviation coefficient 0.775), and a suspicious content generation process (average content generation process entropy 0.865). The average value of the "Individual Information Consistency Index" output by this embodiment reaches 0.91, while the average value of the baseline method is 0.675. The risk score of this embodiment is 34.8% higher than that of the baseline method. The percentage increase is calculated as follows: [(mean risk of this embodiment - mean risk of the baseline method) / mean risk of the baseline method] × 100% = [(0.91 - 0.675) / 0.675] × 100% ≈ 34.8%. This significant increase is due to the introduction of entropy in the content generation process and the fuzzy logic fusion mechanism. The linear weighted model of the baseline method is easily lowered by seemingly normal parameters, while the fuzzy logic rules of this embodiment can more accurately capture the "synergistic effect" of multi-dimensional concurrent risks, thereby making more decisive and accurate high-risk judgments.

[0098] Comparing Scenario 4 (disguised counterfeit goods) and Scenario 6 (covered by high-credibility credentials): these two groups represent typical "high-level" fraud. In Scenario 4, fraudsters use relatively good credentials to publish counterfeit goods with highly consistent images and text and slightly lower prices, but the content generation process is abnormal (mean entropy 0.735). In Scenario 6, fraudsters use almost perfect high-credibility credentials (mean overall credentials credibility 0.965) to publish meticulously crafted content, but the price is abnormal and the generation process is suspicious. For Scenario 6, the baseline method, which only considers static content, gives an average risk score of only 0.375, easily leading to it being judged as low to medium risk and overlooked. However, the method in this embodiment, despite facing high overall credentials credibility, still gives an average "individual information consistency index" of 0.875 through fuzzy logic reasoning, thanks to its keen capture of the entropy of the content generation process (mean 0.81) and the price-value deviation coefficient (mean 0.71). In the most challenging Scenario 6, the risk identification capability of this embodiment is improved by 133.3% compared to the baseline method. The percentage increase is calculated as follows: [(mean risk of this embodiment - mean risk of the baseline method) / mean risk of the baseline method] × 100% = [(0.875 - 0.375) / 0.375] × 100% ≈ 133.3%. This demonstrates that the entropy of the content generation process acts as a risk indicator.

[0099] A specific analysis of Scenario 5 (“Zombie” Credential Fraud): This scenario simulates long-existing credentials (average credential maturity 0.905) that have recently shown no activity (average credential continuity 0.055), suddenly being used to publish seemingly normal information. The credential lifecycle analysis mechanism in this embodiment plays a crucial role here. Despite high credential maturity, the low credential continuity results in an average “comprehensive credential credibility” of only 0.21, prompting the system to immediately mark this source as high-risk. Systems lacking lifecycle analysis capabilities would be misled by its high “maturity.” This powerfully demonstrates the necessity and advancement of combining credential maturity and continuity analysis, enabling this embodiment to effectively prevent the risk of credential theft or “account farming” for fraudulent purposes.

[0100] Further explanation: The group behavior baseline includes statistical distribution parameters of historical information consistency indicators associated with the authoritative entity and product category;

[0101] The calculation of the deviation of the voucher content specifically includes:

[0102] Construct a current user behavior feature vector that includes individual information consistency indicators and comprehensive credential credibility;

[0103] The group behavior baseline is defined as the mean vector and covariance matrix of group behavior features corresponding to the user's group.

[0104] The Mahalanobis distance algorithm is applied to calculate the distance between the current user behavior feature vector and the group behavior baseline, which is used as the deviation of the voucher content.

[0105] Further explanation: The generation of credibility assessment results specifically includes: using the deviation of the credential content as a key input feature to update the user's dynamic trust score; and triggering a preset adaptive review strategy based on a comprehensive assessment of the dynamic trust score and the deviation of the credential content.

[0106] Further explanation: The comprehensive evaluation based on the dynamic trust score and the deviation of the credential content to trigger a preset adaptive review strategy specifically includes:

[0107] The voucher content deviation and the dynamic trust score are used together as state inputs and fed into a pre-trained reinforcement learning model.

[0108] The reinforcement learning model outputs and executes an optimal review policy with the maximum expected utility value based on the state input.

[0109] Further explanation: The state input of the reinforcement learning model further includes the platform risk level, which represents the current overall security status of the platform;

[0110] After executing the optimal review strategy, the reward value is calculated based on the subsequent feedback of the strategy execution results;

[0111] The reward value is used to update the reinforcement learning model to optimize its subsequent policy selection.

[0112] Furthermore, the following specific implementation instructions are provided for the above content:

[0113] The core technical feature of this embodiment lies in: a control method for adaptive risk management based on multidimensional covariance deviation quantification and reinforcement learning-driven approaches. It constructs a multidimensional user behavior feature vector containing individual content consistency and source credential credibility, and uses the Mahalanobis distance algorithm to calculate the credential content deviation degree, which reveals multidimensional correlation anomalies, by measuring the distance between this vector and the corresponding group behavior baseline in the covariance space. Subsequently, the credential content deviation degree is used as a key input, not only to update the user's dynamic trust score but also to form a decision state together with information such as the platform's global risk situation, which is then input into a pre-trained reinforcement learning model. This reinforcement learning model, through a mechanism that maximizes long-term returns, replaces the traditional static rule threshold, dynamically and adaptively selecting and triggering the optimal risk intervention strategy. In this embodiment, the specific implementation details are as follows:

[0114] The mean vector of group behavioral characteristics, with parameter notation as follows: This is represented by a two-dimensional vector, used to characterize the average performance of a specific group in historical behavior, as defined jointly by authoritative entities and product categories. It is determined by statistically calculating relevant data from all completed product release events belonging to the same group in a historical database. Specifically, the first component of the mean vector of the group's behavioral characteristics represents the "consistency index of individual information across all historical events of that group." "The summation is divided by the total number of events to obtain the arithmetic mean; the second component represents the 'comprehensive credibility of the evidence for all historical events of the group'." "After summing, divide by the total number of events to obtain the arithmetic mean."

[0115] The covariance matrix of group behavior characteristics, with parameter notation as follows: This represents a two-dimensional multiplied symmetric matrix used to quantify the dispersion of two dimensions—individual information consistency index and comprehensive evidence credibility—in the historical behavior of a specific group, as well as the linear correlation between them. It is determined as follows: Simultaneously calculated with the mean vector of group behavioral characteristics, the elements on the diagonal of the matrix represent the variances of the individual information consistency index and the comprehensive evidence credibility, respectively. The elements off-diagonal represent the covariance between these two indices. Specifically, for each historical event, the difference between its individual information consistency index and the group mean is calculated, multiplied by the difference between its comprehensive evidence credibility and the group mean, and then the product of these products for all events is summed, finally divided by the total number of events minus 1.

[0116] The current user behavior feature vector, with parameter notation as follows: It represents a two-dimensional vector, which is the coordinate representation of the user's current posting behavior in the multi-dimensional feature space; the determination method is as follows: the first component is the individual information consistency index. The value; the second component is the overall credibility of the voucher. The value of .

[0117] Voucher content deviation, its parameter symbol is The value represents a non-negative scalar value with a range of [0, +∞), used to quantify the deviation of the current user's behavior feature vector from the baseline of their group's behavior in a multidimensional statistical space. The larger the value, the more abnormal the behavior. It is determined as follows: using the Mahalanobis-Distance algorithm, specifically by calculating the difference between the current user's behavior feature vector and the mean vector of the group's behavior features to obtain a difference vector; then calculating the inverse matrix of the group's behavior feature covariance matrix; and multiplying the transpose of the difference vector with the inverse matrix obtained in the second step to obtain an intermediate vector; then calculating the dot product of this intermediate vector and the original difference vector; finally, taking the square root of the dot product result to obtain the voucher content deviation degree.

[0118] User dynamic trust score, its parameter symbol is: Its value range is an integer in the range [0, 1000], used to dynamically track and quantify the long-term trustworthiness of users. The determination method is as follows: each user is given an initial trust score of 500, which is updated based on each posting behavior of the user; and an update logic based on punishment and reward is adopted; the new trust score equals the old trust score minus the punishment item. This punishment item is equal to the product of the basic update step size and a non-linear punishment function based on the deviation of the voucher content; the punishment function is set as follows: when the deviation of the voucher content is lower than the safety threshold, the punishment function value is negative to achieve a reward and increase the trust score; when the deviation of the voucher content is higher than the safety threshold, the punishment function value is positive, and its value increases exponentially with the degree to which the deviation of the voucher content exceeds the safety threshold to achieve severe punishment for serious abnormal behavior; in this embodiment, the basic update step size is set to... The safety threshold is .

[0119] Furthermore, this embodiment defines a security threshold. The determination rule is as follows: the security threshold In this embodiment, it is set to 1.0, which is intended to provide users with dynamic trust scores. The update mechanism provides a balance between rewards and penalties; its specific calibration methods include the following:

[0120] Obtain a dataset containing a large number of historical user posting samples that have been labeled as "normal" or "benign" behavior; for each sample in the benign sample set, calculate its voucher content deviation. The value is used to form a deviation degree of the voucher content specifically for benign behavior. The statistical distribution of the values ​​is analyzed, and specific quantiles that can represent typical normal behavior are selected as the safety threshold. In this embodiment, the voucher content deviation corresponding to the median or 60th percentile of the distribution is selected. value.

[0121] Furthermore, the calibration method of the penalty function in this embodiment was determined through the following offline experimental calibration method:

[0122] We collect a large number of logs of historical user-posted events. Each event includes a calculated deviation of the credential content, as well as an event tag confirmed after manual review or user appeal. Event tags include: normal, low quality, general fraud, and serious fraud.

[0123] Using the deviation of the voucher content as the independent variable and the risk level represented by the corresponding event label as the dependent variable, a nonparametric regression method or piecewise function fitting technique is applied to learn the mapping relationship between the deviation and the penalty intensity to be applied. In this embodiment, the nonparametric regression method includes support vector regression (SVR) or Gaussian process regression (GPR). The "risk level" includes a preset penalty score that can be quantified: "normal" corresponds to -5 points, and "serious fraud" corresponds to +100 points. Finally, the parameters of the fitted function model or piecewise function are solidified into the online calculation system as the specific implementation of the nonlinear penalty function.

[0124] The platform risk level, its parameter symbol is: Its value range is a normalized value in the interval [0, 1], used to macroscopically characterize the current overall security status of the platform; platform risk level The backend monitoring system calculates periodically, and in this embodiment, it is set to update hourly. Specifically, it counts the number of events within a past time window where the deviation of the voucher content exceeds the high-risk threshold, and then divides this number by the total number of events within that time window to obtain the high-risk event incidence rate. This high-risk event incidence rate is then used to represent the platform's risk level. In this embodiment, the "past time window" is set to 24 hours; the high-risk threshold is set to... .

[0125] Furthermore, regarding high-risk thresholds The determination rules are set to 3.0 in this embodiment, aiming to define the platform's risk level. The calculation provides a threshold for efficiently and accurately identifying high-risk events. Its specific calibration method includes the following:

[0126] Obtain a labeled dataset containing "high-risk" and "non-high-risk" samples that have been manually or otherwise labeled; perform classification performance analysis. This is done using a range of different voucher content bias levels. The labeled dataset is divided into candidate thresholds, and the precision and recall of the system in identifying high-risk events are calculated for each candidate threshold. The candidate threshold that optimizes the preset evaluation metric is selected as the final high-risk threshold. The preset evaluation metric is the value that maximizes the recall while ensuring a precision higher than 95%.

[0127] The utility value of the intervention strategy, with parameter symbol as follows: It is a core parameter in reinforcement learning models (Q-learning), representing the intervention strategy represented by the "action" when a specific state s is observed. The expected value of the long-term cumulative reward can be obtained; the utility value of this intervention strategy is stored in a Q-table or in the weights of a deep Q-network (DQN), and is learned and updated through a continuous "state-action-reward-new state" cycle; its update logic originates from the Bellman Equation; the calculation logic of the new intervention strategy utility value is: the old intervention strategy utility value plus the learning rate multiplied by the difference between the sum of the immediate reward and the maximum expected future reward, and the old intervention strategy utility value. The immediate reward is determined based on the actual effect of the strategy execution. In this embodiment, successful fraud blocking is +100, false blocking of normal users is -50, and normal posting without intervention is +1. The core calculation process of this embodiment is as follows:

[0128] 2.1) The initial input is the overall credibility of the current user's credentials calculated in the previous steps. Consistency index with individual information This includes the user's identity credentials and the product category information.

[0129] 2.2) Based on the input voucher type and product category, retrieve the corresponding mean vector of group behavior features from the historical behavior database. Covariance matrix of group behavior characteristics .

[0130] 2.3) Assess the credibility of the input comprehensive vouchers. Consistency index with individual information Construct a two-dimensional current user behavior feature vector .

[0131] 2.4) Applying the Mahalanobis distance calculation model, the above feature mean vectors are... Group behavior characteristic covariance matrix and current user behavior feature vector As input, the deviation rate of voucher content is calculated. .

[0132] 2.5) Based on the calculated deviation rate of voucher content The system invokes a penalty and reward-based update logic model to update the user's dynamic trust score. The new score will then be stored back in the user database.

[0133] 2.6) Integrate all the information required for the current decision-making process to construct the state s of the reinforcement learning model. This state s must include at least: the updated user dynamic trust score. The deviation of the voucher content calculated in this study And the current platform risk level obtained from the backend monitoring system. .

[0134] 2.7) Input the constructed state s into the trained reinforcement learning model. The reinforcement learning model calculates the utility value of each preset intervention strategy. The strategy with the highest utility value is selected as the optimal strategy to be executed in this instance; in this embodiment, "intervention strategy" The example is characterized as {direct approval, delayed release, triggering manual review, temporary ban};

[0135] 2.8) The system executes the optimal strategy selected in the previous step; simultaneously, the event is marked and added to the observation queue; once the final result of the event is confirmed (in this embodiment, the "final result of the event" is manually confirmed as fraud, or no user complaints are received after delayed release), the system calculates an immediate reward value based on the result, and uses this reward value, the current state s, and the executed "intervention strategy" "And the new state, to update the Q value in the reinforcement learning model, thereby completing a learning loop."

[0136] 2.9) The final output of this embodiment consists of two parts: the first part is the updated user dynamic trust score, serving as an assessment result of the user's long-term reputation; the second part is the specific adaptive review strategy triggered by this event, serving as the immediate handling result of multimodal product information and its publisher. The core innovation of this embodiment lies in selecting Mahalanobis distance as the fusion method for calculating the deviation of voucher content; for comprehensive voucher credibility... Users with a value close to 1 tend to have poor content consistency, which is the individual information consistency index. If the value is too high, this incongruous combination will be effectively amplified by Mahalanobis distance, producing a synergistic gain effect; reinforcement learning models, by maximizing long-term rewards, automatically learn the "importance" of different state features in decision-making through continuous interaction and feedback. The following are specific implementation instructions for the above content:

[0137] When the content deviation of the voucher The closer the value is to 0, the more closely the current user's behavior pattern matches the historical average behavior pattern of their group, the more normal the behavior, and the lower the risk of fraud.

[0138] When the content deviation of the voucher As the value increases, the degree to which a user's behavior deviates from its group norms in the multidimensional statistical space also increases, leading to a higher risk of anomalies and fraud.

[0139] The optimal review strategy is a specific action from the discrete action set output by the reinforcement learning model; including but not limited to {direct approval, delayed release, triggering manual review, temporary ban}. These specific strategies are the optimal dynamic decisions made to maximize long-term platform benefits after comprehensively considering the individual risks of current users and the overall security situation of the platform as represented by the platform's risk level.

[0140] When other parameters remain unchanged, the current user behavior feature vector with the mean vector of group behavioral characteristics The larger the Euclidean distance, the greater the calculated deviation of the voucher content. The larger the value, the greater the positive correlation; this is a non-linear positive correlation, determined by the calculation logic of Mahalanobis distance; this embodiment considers the credibility of comprehensive evidence for high numerical values. However, this is accompanied by high values ​​of individual information consistency index. User, their current user behavior feature vector In the feature space, it will fall into a low probability density region. Even if individual indicators are within a reasonable range, Mahalanobis distance can capture this "inconsistent" combination anomaly, thus outputting a higher voucher content deviation. Used to map sophisticated, deceptive fraud detection logic.

[0141] When the covariance matrix of group behavior characteristics A larger covariance, represented by the off-diagonal elements, indicates greater credibility of the overall evidence in the population. Consistency index with individual information The stronger the linear correlation between them, the more the Mahalanobis distance adjusts the coordinate space along the direction of this correlation, amplifying deviations perpendicular to that direction. Application scenarios for the above include: for groups that generally exhibit "high creditworthiness corresponding to low content risk," individual users exhibiting "high creditworthiness corresponding to high content risk" will be judged as having deviated more severely due to violating this strong correlation pattern within the group.

[0142] User dynamic trust score and platform risk level As state inputs to the reinforcement learning model, they collectively influence the selection of the final optimal censorship strategy; when other states remain unchanged, the user's dynamic trust score... When the threshold is lowered, the model tends to select a more stringent censorship strategy.

[0143] Assuming other conditions remain unchanged, the platform's risk level When the risk level rises, the model also tends to choose a more stringent review strategy, even when facing users with moderate risk; this reflects the system's adaptability: when the overall risk of the platform is high, the risk management "threshold" will automatically tighten to deal with systemic risks.

[0144] Furthermore, to verify the effectiveness of this embodiment in identifying covert fraud and achieving adaptive risk management, six representative user posting scenarios were designed. The experiment aimed to compare the performance of the method in this embodiment (outputting "credential content deviation" and having a reinforcement learning model select the "optimal review strategy") with two baseline methods: Baseline Method 1 (Z-score deviation + fixed rules) and Baseline Method 2 (Mahanobis distance deviation + fixed rules). These two baseline methods represent single technical alternatives to the two core innovations of this embodiment (Mahanobis distance and reinforcement learning), respectively. Through comparison, the technical gains brought by each innovation of this embodiment can be clearly quantified.

[0145] The "baseline deviation" is defined as a deviation index used for comparison, based solely on the consistency index of individual information. Calculate the standard deviation of its deviation from the group mean.

[0146] The defined fixed-rule strategy is a strategy selection mechanism based on a fixed threshold. Specifically, if the IF deviation is <1.5, it is passed directly; if the IF deviation is ≥1.5 and <3.0, it is manually reviewed; if the IF deviation is ≥3.0, it is temporarily banned. Specific experimental data is as follows:

[0147]

[0148] Comparing Scenario 1 (normal behavior) and Scenario 3 (abnormal association pattern): Scenario 3 represents advanced fraud: the attacker bases their actions on "comprehensive credential credibility". A mean value of 0.955 is used to characterize credentials with excellent credibility, and an index based on "individual information consistency" has been published. A mean of 0.49 is used to characterize information with moderate content risk. For baseline method 1, since it only considers the consistency index of individual information... The calculated "baseline deviation" was only 1.225, lower than the "manual review" threshold of 1.5, thus incorrectly selecting "directly pass" and overlooking fraud risk. However, the method in this embodiment, through Mahalanobis distance calculation, astutely captures the incongruous combination of "high-reputation credentials" and "medium content risk," which deviates significantly from the strong correlation of "high reputation corresponding to low risk" within the group. Therefore, the calculated credential content deviation is... The deviation rate is 3.475. This embodiment represents the document content deviation rate. The value is 2.84 times that of the "baseline deviation" value, calculated as 3.475 / 1.225. This significant quantitative difference demonstrates that Mahalanobis distance, compared to traditional unidimensional analysis methods, is more effective in identifying covert fraud with "pattern anomalies" that use high-reputation identities as cover.

[0149] Comparing the performance of baseline method 2 and the method of this embodiment in scenarios five and six: Scenario five and scenario six represent exactly the same individual user behavior, and their credential content deviation is... The mean is approximately 2.8, but this varies depending on the macroeconomic risk situation of different platforms. Scenario five is "platform risk level". The high-risk situation represented by "0.8" is Scenario Six, which is the "platform risk level". The baseline method 2, using fixed rules, outputs the same "manual review" strategy in both scenarios. However, the reinforcement learning model in this embodiment selects "manual review" in scenario six (low risk), but in scenario five (high risk), despite facing the same individual risk, it upgrades the intervention strategy to "temporary ban." This comparison powerfully demonstrates the significant advantages of the reinforcement learning-driven adaptive decision-making mechanism in this embodiment. It proves that the system is not merely a passive risk assessor, but a proactive decision-making system with macro-situational awareness. Compared to all existing technologies based on fixed rules, this embodiment can dynamically adjust its risk tolerance according to the overall security status of the platform, automatically tightening control strategies during periods of high risk and relaxing them during stable periods to ensure user experience, achieving optimal dynamic allocation of risk management resources.

[0150] Furthermore, this embodiment sets the following interval division criteria; these criteria are based on statistical analysis of massive historical data, specifically the deviation of voucher content between samples marked as fraudulent and normal. The value distribution was fitted and combined with expert experience in platform risk management strategies to formulate them;

[0151]

[0152] Example 2:

[0153] A campus secondhand goods transaction credibility traceability assessment system, the system being used to execute the campus secondhand goods transaction credibility traceability assessment method, comprising:

[0154] Credential verification and data acquisition module: used to receive and verify trusted entity credentials based on zero-knowledge proofs associated with the user's identity, wherein the trusted entity credentials are used to prove that the user belongs to a preset authoritative entity;

[0155] Obtain multimodal product information to be evaluated, published by the user;

[0156] Multidimensional feature deep analysis module: used to calculate an individual information consistency index that quantifies the degree of internal logical self-consistency based on the data correlation within the multimodal product information;

[0157] Individual-Group Deviation Measurement Module: Based on the trusted entity credentials, it determines the baseline of group behavior associated with the authoritative entity, and calculates the credential content deviation degree, which characterizes the degree of deviation between user behavior and the behavior of their group, in conjunction with the individual information consistency index.

[0158] Reinforcement learning adaptive decision-making module: used to generate a credibility assessment result for the multimodal product information and the user based on the deviation degree of the voucher content.

[0159] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization.

[0160] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for evaluating the credibility and traceability of secondhand goods transactions on campus, characterized in that, The specific steps include: S1: Receive and verify a trusted entity credential based on zero-knowledge proof associated with the user's identity, wherein the trusted entity credential is used to prove that the user belongs to a preset authoritative entity; Confirm the validity status of the trusted entity certificate; based on the distributed timestamp record of the trusted entity certificate, calculate the certificate maturity characterizing its historical length and the certificate continuity characterizing its activity frequency stability. A comprehensive document credibility is obtained by nonlinearly fusing document validity status, document maturity, and document continuity. S2: Obtain multimodal product information to be evaluated published by the user; S3: Based on the data correlation within the multimodal product information, calculate the individual information consistency index to quantify the degree of internal logical self-consistency; S4: Based on the trusted entity credential, determine the group behavior baseline associated with the authoritative entity, and in conjunction with the individual information consistency index, calculate the credential content deviation degree, which characterizes the degree of deviation between user behavior and the behavior of their group; the calculation of the credential content deviation degree specifically includes: Construct a current user behavior feature vector that includes individual information consistency indicators and comprehensive credential credibility; The group behavior baseline is defined as the mean vector and covariance matrix of group behavior features corresponding to the user's group. The Mahalanobis distance algorithm is applied to calculate the distance between the current user behavior feature vector and the group behavior baseline, which is used as the deviation of the voucher content. S5: Based on the deviation of the voucher content, generate a credibility assessment result for the multimodal product information and the user.

2. The method for credible traceability assessment of secondhand goods transactions on campus according to claim 1, characterized in that: Verifying the trusted entity credential specifically includes: using a preset public key associated with the authoritative entity to perform a zero-knowledge proof verification algorithm on the trusted entity credential, and confirming the credential validity status of the trusted entity credential without obtaining the user's specific identity information.

3. The method for credible traceability assessment of second-hand goods transactions on campus according to claim 2, characterized in that: The calculation of the individual information consistency index specifically includes: Calculate the image-text concept alignment degree, which represents the semantic matching degree between the image and the text; Calculate the price-value deviation coefficient, which represents the degree of deviation between the listed price and the predicted fair value of a commodity. Calculate the entropy of the content generation process that generates abnormal behaviors during the creation of quantitative information.

4. The method for credible traceability assessment of secondhand goods transactions on campus according to claim 3, characterized in that: The content generation process entropy includes: based on the behavior logs of users when creating multimodal product information, jointly quantifying the source attributes of the image information and the stability of the editing process of the text information; The text-image concept alignment, price-value deviation coefficient, and content generation process entropy are used as inputs, and nonlinear fusion is performed through a fuzzy logic reasoning system to obtain the individual information consistency index.

5. The method for credible traceability assessment of secondhand goods transactions on campus according to claim 4, characterized in that: The fuzzy logic-based reasoning system includes a fuzzy rule base, which automatically optimizes and generates rules for the fuzzy rule base by training machine learning on historically labeled fraudulent and normal sample data. By analyzing the metadata of user-uploaded images and logging editing behavior during the text input process, and applying Shannon entropy calculation logic, the entropy of the content generation process is obtained.

6. The method for credible traceability assessment of secondhand goods transactions on campus according to claim 5, characterized in that: The greater the credibility of the comprehensive credential, the better the characterization system judges the authenticity, historical accumulation and activity stability of the user's credential, and the higher the prior credibility of the information source. The higher the consistency index of individual information, the more serious the internal contradictions and anomalies in the product information published by the user are in all dimensions, the lower the credibility of the content itself, and the higher the risk of fraud.

7. The method for credible traceability assessment of secondhand goods transactions on campus according to claim 6, characterized in that: The group behavior baseline includes statistical distribution parameters of historical information consistency indicators associated with the authoritative entity and product category.

8. The method for credible traceability assessment of secondhand goods transactions on campus according to claim 7, characterized in that: The specific steps for generating a credibility assessment result include: using the deviation of the voucher content as a key input feature to update the user's dynamic trust score; and triggering a preset adaptive review strategy based on a comprehensive assessment of the dynamic trust score and the deviation of the voucher content. Define a pre-defined adaptive review strategy, specifically including: The credential content deviation and the dynamic trust score are used together as state inputs and fed into a pre-trained reinforcement learning model. The reinforcement learning model then outputs and executes the optimal review strategy with the maximum expected utility value based on the state inputs.

9. The method for credible traceability assessment of secondhand goods transactions on campus according to claim 8, characterized in that: The state input of the reinforcement learning model further includes the platform risk level, which represents the current overall security status of the platform. After executing the optimal review strategy, the reward value is calculated based on the subsequent feedback of the strategy execution results; The reward value is used to update the reinforcement learning model to optimize its subsequent policy selection; The smaller the deviation of the voucher content, the more closely the current user's behavior pattern matches the historical average behavior pattern of its group, the more normal the behavior, and the lower the risk of fraud. The optimal censorship strategy is a specific action from the discrete action set output by the reinforcement learning model.

10. A reliable traceability and evaluation system for secondhand goods transactions on campus, characterized in that: The system is used to execute the campus second-hand goods transaction credibility traceability assessment method according to any one of claims 1-9, including: Credential verification and data acquisition module: used to receive and verify trusted entity credentials based on zero-knowledge proofs associated with the user's identity, wherein the trusted entity credentials are used to prove that the user belongs to a preset authoritative entity; Obtain multimodal product information to be evaluated, published by the user; Multidimensional feature deep analysis module: used to calculate an individual information consistency index that quantifies the degree of internal logical self-consistency based on the data correlation within the multimodal product information; Individual-Group Deviation Measurement Module: Based on the trusted entity credentials, it determines the baseline of group behavior associated with the authoritative entity, and calculates the credential content deviation degree, which characterizes the degree of deviation between user behavior and the behavior of their group, in conjunction with the individual information consistency index. Reinforcement learning adaptive decision-making module: used to generate a credibility assessment result for the multimodal product information and the user based on the deviation degree of the voucher content.

Images