A multi-source cross-platform fusion-based anti-fraud early warning system

Abstract

The application provides an anti-fraud early warning system based on multi-source cross-platform fusion, belongs to the field of network security and privacy computing, and aims to solve the problems of privacy set intersection matching failure and early warning false alarm caused by data noise. The scheme cleanses data and calculates cleansing distortion degree through a multi-modal normalization engine; generates a fuzzy feature vector containing multiple variants through a fuzzy feature mapper to avoid hash avalanche effect; performs privacy set intersection in the ciphertext domain through an encryption calculation unit; and finally generates an early warning by a risk decision module in combination with variant weight and time decay factor for weighted scoring. The application upgrades the precise matching to fuzzy query, significantly improves the early warning robustness and accuracy, and realizes risk quantification and privacy security.

Description

Technical Field

[0001] This application relates to the fields of network security and privacy computing technology, and in particular to a system and method for anti-fraud early warning in a multi-party data fusion scenario. Background Technology

[0002] With the increasing prevalence of telecommunications fraud and other illegal activities, the integration of data from multiple sources, including finance, communications, and the internet, for joint anti-fraud early warning has become a crucial technological approach. To comply with relevant regulations on data security and personal privacy protection, the industry typically employs privacy computing techniques such as privacy set intersection. This technology allows participating parties to calculate the intersection of their data sets without directly exposing their original data; for example, to determine whether a target user is on a blacklist.

[0003] In the current anti-fraud risk control system, the core challenge lies in the 'multi-source heterogeneity' and 'cross-platform isolation' characteristics of data assets. On the one hand, high-value risk data is scattered across different cross-platform entities such as financial institutions, telecommunications operators, internet platforms, and government departments, forming physical data silos. On the other hand, data from different sources (multi-source) differs significantly in collection standards, storage formats, and business definitions (for example, the same user's identifier may be a real-name account on the bank's end, but a virtual ID on social media, and there are inconsistencies in format). Existing cross-platform collaboration often adopts simple blacklist distribution or hash collision modes, which not only fails to meet the real-time and privacy compliance requirements of data fusion between different platforms, but more importantly, when faced with the format noise prevalent in multi-source heterogeneous data, the lack of a unified fusion and normalization mechanism easily leads to the failure of cross-platform joint prevention and control mechanisms.

[0004] However, existing privacy set intersection techniques based on precise hash matching have a serious drawback in practical applications. Due to user input habits, optical character recognition errors, or adversarial interference deliberately constructed by fraudsters (such as using special symbols, homophones, or variant characters), the identity identifiers of the same user (such as mobile phone numbers and ID card numbers) often show slight differences across different data sources. This phenomenon can be called "data noise." Standard hash functions are extremely sensitive to input content; a change in any bit can lead to a significant difference in the output hash value, resulting in the so-called "hash avalanche effect." This makes privacy set intersection protocols that rely on precise hash value comparison completely ineffective in scenarios with data noise, directly causing a large number of fraud risks to be missed due to matching failures, resulting in a large number of false negatives and severely impacting the effectiveness of anti-fraud systems' early warning capabilities. Furthermore, many existing data cleaning tools rely on static, fixed rule bases, making it difficult to cope with constantly evolving new adversarial noise, and they lack quantitative evaluation mechanisms for errors that may be introduced into the cleaning process itself, thus failing to effectively solve the above problems. Summary of the Invention

[0005] The purpose of this application is to provide an anti-fraud early warning system based on multi-source cross-platform fusion, which aims to solve the technical problem in the prior art that when using privacy set intersection to perform multi-source data matching, data noise such as inconsistent identity identifier formats, input errors or adversarial interference will cause an avalanche effect after hash operation, resulting in matching failure and missed early warnings.

[0006] To achieve the above objectives, this application provides an anti-fraud early warning system based on multi-source cross-platform fusion, characterized by comprising: a multimodal normalization engine configured to receive heterogeneous original identity data from multiple sources, clean the original identity data based on a preset rule base, generate normalized standard data, and calculate a cleaning distortion index reflecting the change in information content before and after cleaning; a fuzzy feature mapper configured to generate a set of variants containing multiple variants satisfying preset edit distance conditions based on the normalized standard data when the cleaning distortion index meets a preset threshold, and perform hash processing on the variant set to generate a fuzzy feature vector to be queried; an encryption calculation unit configured to send the fuzzy feature vector to be queried to a data source node through a secure channel, and perform a privacy set intersection operation in the ciphertext domain to obtain a ciphertext overlap result; and a risk decision module configured to calculate a final risk confidence score based on the ciphertext overlap result, preset variant weights and time decay factors, through a weighted risk scoring model, and generate an early warning instruction based on the score.

[0007] Optionally, the system further includes an intelligent normalization agent module, configured to invoke a large language model to perform semantic processing on the original identity data when the cleaning distortion index exceeds a preset threshold or the risk confidence score is within a preset fuzzy range.

[0008] Optionally, the intelligent normalization proxy module further includes a dynamic rule generation unit, configured to generate regular expressions for recognizing similar noise patterns based on the processing results of the large language model, and update them to the rule base of the multimodal normalization engine.

[0009] Optionally, the system further includes a feedback closed-loop unit configured to adjust the preset variant weights of different variant types in the risk decision module in reverse based on the false alarm or missed alarm results fed back by the external business system.

[0010] Optionally, the wash-distortion metric is the ratio of the Levinstein edit distance between the original identity data and the normalized standard data to the length of the original identity data. For example, the wash-distortion metric... It can be calculated using the following formula:

[0011]

[0012] in, This is the original identity data. To normalize standard data, This represents the Levinstein edit distance function. Indicates the length of the string.

[0013] Optionally, the weighted risk scoring model is used to calculate the risk confidence score by multiplying the corresponding variant weights of all variants of the ciphertext overlap results by a time decay factor, summing the results, and then inputting the sum into a Sigmoid activation function. For example, the weighted risk scoring model... It can be calculated using the following formula:

[0014]

[0015] in, This represents the risk confidence score. It is the Sigmoid activation function. For the number of variants, For the first Confidence weights for each variant This is an indicator function that indicates whether the variant matches the intersection. The current system time. The entry time of the blacklist data. This is the time decay coefficient.

[0016] Optionally, the variant set includes at least one of the following: exact match variants, variants based on homophones or similar-looking characters, and variants that satisfy the preset edit distance condition; wherein, the preset edit distance condition is: the Levenstein edit distance between the generated variant and the normalized standard data is less than or equal to K; for numeric data, K is 1 or 2; for text data, K is within 20% of the length of the normalized standard data.

[0017] Optionally, the encryption calculation unit employs an RSA-based blind signature or a privacy set intersection protocol based on unintentional transmission.

[0018] Optionally, at least one of the multimodal normalization engine and the cryptographic computing unit is deployed within the trusted execution environment of the initiating node.

[0019] Optionally, the secure channel employs a Transport Layer Security Protocol version 1.3 or a higher encrypted transmission protocol.

[0020] Compared with the prior art, this application has the following beneficial effects:

[0021] 1. Overcoming the bottleneck of multi-source heterogeneous data fusion. This application effectively solves the problem of 'difficult data fusion' caused by inconsistent data collection standards and format noise interference between different platforms through multimodal normalization and fuzzy feature mapping mechanisms, and achieves efficient and accurate fusion of multi-source data under the premise of privacy protection.

[0022] 2. Achieving Risk Quantification and Controllability. This application quantifies the risks introduced by the data cleaning process to the original information by calculating the cleaning distortion index during the data preprocessing stage. When the distortion is too high, subsequent high-cost encryption calculations can be stopped, which not only avoids erroneous matches that may be caused by "over-cleaning" but also saves computing resources.

[0023] 3. Ensuring data privacy and security for all parties. The entire matching process in this application is completed in the encrypted domain under the guidance of the encrypted computing unit, ensuring that the original sensitive data of each participating party is not leaked during cross-platform data fusion, thus meeting stringent data security and compliance requirements.

[0024] 4. Improve the timeliness and accuracy of early warning decisions. This application introduces a time decay factor into the risk decision-making module, which can dynamically reduce the risk weight of outdated blacklist data, effectively solving the problem of "historical blacklists mistakenly affecting current users" and making the early warning results closer to the current actual risk situation; at the same time, it introduces variant weights, which enable weighted scoring of matching results with different credibility levels, further improving the accuracy of decision-making. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the architecture of an anti-fraud early warning system based on multi-source cross-platform fusion, provided for an embodiment of this application.

[0026] Figure 2 A flowchart illustrating an anti-fraud early warning method based on multi-source cross-platform fusion provided in this application embodiment.

[0027] Figure 3 This is a timing diagram of the signaling interaction between the initiating node and the data source node in one embodiment of this application.

[0028] Figure 4 This is a closed-loop diagram illustrating the evolution of intelligent agents and rules in one embodiment of this application.

[0029] The main reference numerals in the attached diagrams are explained as follows: 10 - Initiating node; 20 - Data source node; 110 - Multimodal normalization engine; 120 - Fuzzy feature mapper; 130 - Encrypted calculation unit; 140 - Risk decision module; 150 - Intelligent normalization agent module; 151 - Dynamic rule generation unit; 301 - PSI query request; 302 - PSI query response; 400 - Large language model; 500 - Business feedback; S201 - Data reception step; S202 - Normalization and distortion calculation step; S203 - Distortion judgment step; S204 - Fuzzy query vector generation step; S205 - PSI execution step; S206 - Risk score calculation step; S207 - Early warning generation step. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0031] Example 1

[0032] This application discloses an anti-fraud early warning system based on multi-source cross-platform fusion, which aims to provide a technical solution that can still accurately and securely perform multi-source data fusion early warning even in the presence of data noise.

[0033] Please see Figure 1 The diagram illustrates the architecture of an anti-fraud early warning system based on multi-source cross-platform fusion, as provided in an embodiment of this application.

[0034] The system can be deployed in a distributed network environment and mainly includes an initiating node 10 and one or more data source nodes 20. The initiating node 10 can be a business party that needs to perform risk inquiries, such as a bank or payment platform; while the data source node 20 can be a party that holds risk data (such as a blacklist), such as the National Anti-Fraud Center or telecommunications operators.

[0035] Based on the aforementioned distributed architecture, this system achieves cross-platform integration in a physically isolated environment. Specifically, 'integration' does not refer to the physical aggregation of raw data, but rather to the logical-level feature alignment and risk element fusion achieved within the encrypted domain through a secure channel established by the encrypted computing unit 130. The feature vector generated by the fuzzy feature mapper 120 of the initiating node can cross platform boundaries and perform matching operations within the trusted execution environment of the data source node 20. This architecture not only breaks down physical barriers between platforms, achieving value fusion of multi-party data, but also ensures, through an anti-noise mechanism, that matching losses caused by inconsistent data standards are minimized during cross-platform interaction.

[0036] In one embodiment of this application, the core functional modules of the system can be mainly deployed on the initiating node 10. Specifically, the initiating node 10 is equipped with a multimodal normalization engine 110, a fuzzy feature mapper 120, an encryption computing unit 130, and a risk decision module 140. Correspondingly, the data source node 20 is also equipped with a cooperating encryption computing unit 130 to collaboratively complete the encryption computing task.

[0037] The multimodal normalization engine 110 serves as the system's entry point, receiving raw identity data in various formats from different business systems or data sources. In the multi-source access scenario of this embodiment, the raw identity data encompasses heterogeneous data formats across industries, including but not limited to: structured financial transaction messages (such as bank card numbers and transaction serial numbers), semi-structured communication signaling data (such as device IMEI and IP addresses), and unstructured user input text (such as address information and remarks). Due to differences in the collection terminals and business specifications of different data sources, this multi-source data often carries the format characteristics of specific platforms. This system uses the multimodal normalization engine to clean and align heterogeneous data, laying a unified data benchmark for subsequent fusion computation.

[0038] It should be noted that this data may contain various "noise," such as extra spaces, mixed full-width and half-width characters, special symbols, and even user input errors. The engine 110 internally maintains a configurable and extensible rule base. Upon receiving the raw data, the engine can apply rules from the rule base to clean and standardize the data, thereby generating normalized standard data in a uniform format. As an example, the rule base may include, but is not limited to: rules for removing all whitespace characters, rules for converting all letters to uppercase or lowercase, and rules for converting full-width characters to their corresponding half-width characters.

[0039] Specifically, after cleaning, the multimodal normalization engine 110 can also calculate a cleaning distortion index. This index is used to quantify the degree of modification to the original information during the cleaning process, thereby assessing the risk of potential errors introduced by the cleaning operation. In a specific implementation of this application, the cleaning distortion index can be defined as the ratio of the Levenstein edit distance between the original identity data and the normalized standard data to the length of the original identity data. Specifically, this cleaning distortion index... It can be calculated using the following formula:

[0040]

[0041] in, Represents original identity data. This represents the normalized standard data generated after cleaning. The function is used to calculate the minimum number of editing operations (insertion, deletion, or replacement) required between two strings. The function returns the length of the string. This ratio provides a standardized measure of the relative extent to which the original data was modified.

[0042] The fuzzy feature mapper 120 is used to implement the noise-resistant matching function of the scheme in this application. As an optional implementation, the fuzzy feature mapper 120 can receive normalized standard data generated by the multimodal normalization engine 110 and confirm its cleaning distortion index. The preset integrity threshold has not been exceeded (e.g., It is then activated. Understandably, if the distortion exceeds this threshold, it may mean that the quality of the original data is too low and the cleaning process has severely distorted the original information. At this time, the system can stop the subsequent process to avoid generating incorrect matches.

[0043] When the conditions are met, the fuzzy feature mapper 120 can generate a set of variants containing multiple variants that satisfy preset edit distance conditions based on normalized standard data and through a series of preset variant generation strategies. This step aims to cover various minor variations that may exist in the original data. In a specific implementation of this application, the variant set may include at least one of the following types of variants:

[0044] 1. Exact match variant: that is, the normalized standard data itself.

[0045] 2. Variants based on homophones or similar-looking characters: For example, for the "X" in an ID card number, a variant "x" can be generated; for the number "8", a similar-looking letter "B" can be generated.

[0046] 3. Variations within a preset edit distance: For example, all strings that can be generated by adding, deleting, or replacing a single character in normalized standard data.

[0047] In a preferred embodiment of this application, the threshold Strictly limited to Within the range.

[0048] The technical reason for this limitation is that, through experimental verification, the applicant discovered that in anti-fraud blacklist matching scenarios, over 90% of the "data noise" is caused by input errors of 1 to 2 characters (such as OCR recognition errors or typos). If set... The number of variants generated will explode exponentially (e.g., for an 11-digit phone number, when...). At times, the number of potential variants can reach millions, which will lead to a significant increase in the false positive rate of subsequent Bloom filters, and the computational overhead of the Power Probe Injection (PSI) exceeds the timeliness requirements of real-time alerts. Therefore, limiting the edit distance to within 2 is the optimal balance between recall (covering noise) and computational efficiency.

[0049] Therefore, limiting the edit distance to within 2 is the optimal balance between recall (coverage noise) and computational efficiency.

[0050] After generating the variant set, the fuzzy feature mapper 120 can hash each variant element in the set (e.g., using a salted SHA-256 algorithm) to enhance security. It can then encode all hash values ​​into a compact data structure. In this embodiment, a Bloom filter is preferably used to generate a fuzzy feature vector to be queried. It is understood that a Bloom filter is a space-efficient probabilistic data structure that uses a bit array and a set of hash functions to represent a set, efficiently determining whether an element belongs to that set. In this way, multiple queries that would normally require multiple variants can be integrated into a single query on a single Bloom filter vector, significantly improving processing efficiency.

[0051] The encrypted computing unit 130 performs privacy-preserving cross-node computations. Specifically, the encrypted computing unit 130 located at the initiating node 10 can send the fuzzy feature vector to be queried, generated by the fuzzy feature mapper 120, to the data source node 20 via a secure channel. This secure channel ensures the confidentiality and integrity of the data during transmission. Subsequently, the two encrypted computing units 130 on the initiating node 10 and the data source node 20 collaboratively execute a privacy set intersection protocol. This protocol is performed in the ciphertext domain to calculate the overlap between the fuzzy feature vector to be queried (which represents the variant set of the querying party) and the blacklist data set held by the data source node 20. After the computation is completed, the encrypted computing unit 130 of the data source node 20 returns a ciphertext overlap result to the initiating node 10. It should be noted that this result only reveals the number of variants that hit the blacklist, without disclosing the specific variants that hit, nor any other information in the blacklist of the data source node 20, thereby protecting the data privacy of both parties.

[0052] As the system's decision-making center, the risk decision module 140, upon receiving the ciphertext overlap result from the encryption calculation unit 130, can initiate a weighted risk scoring model to calculate the final risk confidence score. This model can comprehensively consider information from multiple dimensions to achieve more accurate risk assessment. Specifically, the model can combine preset variant weights and a time decay factor. The variant weights assign different levels of confidence to different types of variants; for example, exact match variants have the highest weight (e.g., 1.0), while variants involving similar-looking character substitutions have a lower weight (e.g., 0.6), because the latter introduces more uncertainty. The time decay factor is used to reduce the influence of outdated blacklist data, as risk records from over a year ago may have diminished reference value in the present.

[0053] In one specific implementation of this application, the weighted risk scoring model The following formula can be used for calculation:

[0054]

[0055] in, It is the final risk confidence score; It is an activation function, such as the Sigmoid function, which can map any real value to the interval (0, 1), which can be easily understood as a probability; It is the total number of variants in the variant set; It is the first Confidence weights for each variant; It is an indicator function, when the variant Confirmed to have hit the intersection When the time condition is met, its value is 1; otherwise, it is 0. It is the current system timestamp; It is the timestamp of the blacklist data that was hit and entered into the database; It is a positive number, which serves as a time decay coefficient. The larger the value, the faster the risk weight decays over time.

[0056] Risk decision module 140 calculates Then, it will be compared with a preset risk threshold (e.g., 0.75). If the score exceeds the threshold, the system will generate an early warning instruction to notify the downstream business system that there is a high risk and that manual intervention or corresponding risk control measures are required.

[0057] Please refer to the following: Figure 2 and Figure 3 .in, Figure 2 A flowchart illustrating an early warning method provided in an embodiment of this application. Figure 3 This is the signaling interaction timing diagram between the initiating node and the data source node. The following will use a specific scenario as an example to describe the working process of this embodiment: Suppose that the business system of the initiating node 10 (a bank) receives a mobile phone number entered by a user. The original data is "138 1234 5678", which contains a half-width space.

[0058] Step S201: The system receives the raw data.

[0059] Step S202: The multimodal normalization engine 110 starts and applies the "remove non-numeric characters" rule from its rule base to clean "138 1234 5678" into the normalized standard data "13812345678". Simultaneously, the cleaning distortion index is calculated. In this example, the original data length is 12 (including spaces), the standard data length is 11, and the Leavenstein edit distance is 1 (removing one space). Therefore... .

[0060] Step S203: Determine the cleaning distortion index. Since 0.083 is much smaller than the preset threshold of 0.4, the process continues.

[0061] Step S204: The fuzzy feature mapper 120 generates a set of variants based on "13812345678", which may include: {"13812345678", "1381234567B", "13B12345678"}. Subsequently, these three variants are hashed and stored in a Bloom filter to generate the fuzzy feature vector to be queried.

[0062] Step S205: The encryption calculation unit 130 encapsulates this vector in a PSI query request 301 and sends it to the data source node 20 (e.g., an anti-fraud center) via a secure channel. The encryption calculation unit 130 of the data source node 20 performs a privacy set intersection operation and determines that the entry "13812345678" exists in its blacklist database. After the operation is completed, a ciphertext overlap result indicating "hitting 1 element" is encapsulated in a PSI query response 302 and returned.

[0063] Step S206: After receiving the response, the risk decision module 140 of the initiating node 10 calculates the risk score. Assume the weight of the exact match variant "13812345678" is... Furthermore, the blacklist data was entered a week ago, and the calculated time decay factor is 0.95. Since the other two variants were not hit, their indicator function values ​​are 0, therefore the summation term is... After processing with the Sigmoid function, a higher risk confidence score can be obtained, for example... .

[0064] Step S207: Assuming the risk threshold is set to 0.7, since The system will eventually generate an early warning command.

[0065] Through the above process, even if the original input contains format noise, the system provided in this embodiment can successfully match the blacklist, thereby effectively avoiding false negatives caused by hash avalanche effect and realizing a robust and secure early warning function.

[0066] Example 2

[0067] As an optional implementation, this embodiment introduces an enhancement module based on Embodiment 1, designed to handle more complex and adversarial semantic noise. This type of noise is often difficult for traditional regular expression-based rule bases to cover.

[0068] Please refer to it again. Figure 1 In this embodiment, an intelligent normalization proxy module 150 is added to the system architecture. This module is connected to the multimodal normalization engine 110 and the fuzzy feature mapper 120. Under normal circumstances, the intelligent normalization proxy module 150 is in a standby state, and its triggering conditions are usually twofold: first, the cleaning distortion index calculated by the multimodal normalization engine 110. If the value exceeds a preset threshold (e.g., 0.4), it indicates that conventional cleaning methods may have failed or introduce significant distortion; secondly, the risk confidence score ultimately calculated by the risk decision module 140... falls within a preset "fuzzy interval" (e.g., between 0.3 and 0.6), which indicates that the system has low confidence in the matching result and requires higher-order analysis to assist in decision-making.

[0069] Specifically, the core of the intelligent normalization agent module 150 can be an integrated or invoked large language model. The large language model has the capabilities of natural language understanding, context reasoning, and knowledge association, and can process unstructured and highly confusing information.

[0070] The technical solution of this embodiment is described below through a specific working process. Suppose the optical character recognition system of the initiating end node 10 recognizes an issuing authority name from the reverse side of an ID card screenshot uploaded by a user. The name is mixed with handwritten characters, printed characters, and stains, and the recognition result is "丄每市公安局静女分局", which is data containing serious noise.

[0071] The multimodal normalization engine 110 receives the raw data "丄每市公安局静女分局". It attempts to apply the rule base for cleaning. However, since characters such as "丄", "每", and "女" are very different from "上", "海", and "安" in the standard name "上海市公安局静安分局", the conventional homophone and形近 replacement rules cannot take effect. Therefore, the cleaned result may be little different from the original input, or some incorrect replacements may be made. In either case, the calculated cleaning distortion will be very large, far exceeding the preset threshold of 0.4.

[0072] The system detects After the limit is exceeded, the normal process can be automatically aborted, and the task can be routed to the intelligent normalization agent module 150.

[0073] The intelligent normalization agent module 150 can encapsulate the original text "丄每市公安局静女分局", together with its context information, such as data source, field type, etc., into a structured prompt and send it to a large language model that is pre-deployed or invoked through an application programming interface and fine-tuned with relevant domain knowledge.

[0074] After receiving the prompt, the large language model performs semantic reasoning. It can understand that "丄" is a character with a similar shape to "上", that "每" and "海" may be confused in some handwritten forms, and that "静女" is very likely a misrecognition of "静安" in the context of "公安局…分局". Based on its reasoning ability, the large language model can output a structured result in JSON format, such as: {"standard_text": "Shanghai Public Security Bureau Jing'an Branch", "confidence": 0.95, "correction_logic": ["丄->上 (shape)", "每->海 (ocr_confusion)", "女->安 (ocr_confusion)"], "adversarial_label": "multiple_ocr_errors"}.

[0075] After receiving this result with high confidence, the intelligent normalization proxy module 150 can use the extracted standard text "Shanghai Public Security Bureau Jing'an Branch" as new and credible normalized standard data and pass it to the fuzzy feature mapper 120.

[0076] The subsequent process can refer to the relevant description in Embodiment 1, that is, the fuzzy feature mapper 120 generates a variant set based on this corrected standard data (only exact matching variants may be required at this time), and continues to execute the private set intersection and risk decision-making processes.

[0077] By introducing the intelligent normalization proxy module 150, the system in this embodiment can effectively handle high-order and adversarial semantic noises that cannot be dealt with by traditional rules, thus greatly expanding the early warning coverage of the system and the recognition ability of new fraud means, and significantly improving the intelligence level and robustness of the system.

[0078] Embodiment 3

[0079] Based on Embodiment 1 and Embodiment 2, this embodiment further provides the system with the ability of adaptive evolution, enabling it to learn from processed new cases and self-optimize according to the actual business effects, forming a dynamic and continuously evolving closed-loop system.

[0080] Please refer to Figure 4 , which shows in detail two core closed-loop mechanisms implemented in this embodiment: the "knowledge distillation" closed loop and the "weight evolution" closed loop.

[0081] Among them, a closed-loop can be a "knowledge distillation" closed-loop, aiming to convert the high-order reasoning ability of large language models into low-cost and high-efficiency conventional rules. This is mainly achieved by adding a dynamic rule generation unit 151 inside the intelligent normalization agent module 150, and this mechanism supports the dynamic update of the rule base of the multimodal normalization engine 110.

[0082] A specific working process may include: Continuing the scenario of Embodiment 2, when the large language model 400 successfully identifies and corrects complex noises such as "丄每市公安局静女分局" that contain multiple optical character recognition errors, the intelligent normalization agent module 150 will not only use the correction result for the current query, but also capture this successful processing case and send an instruction to the dynamic rule generation unit 151. The dynamic rule generation unit 151 can further interact with the large language model 400, for example, sending a new prompt: "Given that '丄每市公安局静女分局' should be '上海市公安局静安分局', please generate a set or a regular expression to identify and correct patterns similar to misidentifying '上' as '丄','海' as '每', and '安' as '女'."

[0083] The large language model 400 can generate a series of replacement rules or a complex regular expression according to this instruction. After receiving the newly generated rules, the dynamic rule generation unit 151 will not immediately put them into the production environment, but first use the historical data set to strictly verify and test the new rules in an isolated sandbox environment to evaluate their accuracy, recall rate, and possible side effects (for example, whether they will incorrectly modify normal data). Only when the new rules pass all tests and prove to be safe and effective will they be formally and atomically updated to the rule base of the multimodal normalization engine 110.

[0084] Through this "knowledge distillation" closed-loop, the system can solidify the knowledge obtained from a high-cost large language model call into low-cost and high-efficiency regular expressions. In the future, when the system encounters the same type of noise pattern again, the multimodal normalization engine 110 can directly use the new rules for quick processing without having to call the large language model 400 again, thus achieving the self-enhancement of the system's capabilities and the continuous optimization of the operating efficiency.

[0085] Another closed-loop can be a "weight evolution" closed-loop, aiming to enable the scoring model of the risk decision module 140 to self-correct according to real-world feedback. This requires the system to be able to receive feedback from the external business system 500.

[0086] The process may include: Suppose that in a certain alert, the system successfully matches a variant of a "similar substitution" (e.g., replacing '8' with 'B' in a phone number) and calculates a high risk score, ultimately issuing an alert. However, after manual verification or customer complaint, the downstream business system confirms that this alert was a false alarm.

[0087] The business feedback 500 (including the warning ID and a "false alarm" label) is sent back to the system. Upon receiving this signal, a feedback processing logic within the system (which can be considered part of the risk decision-making module 140 or an independent feedback loop unit) traces the decision path of this warning and finds that it was caused by a match of the "shape-similar substitution" variant. To reduce the likelihood of similar false alarms in the future, the system will activate a parameter fine-tuning mechanism. Specifically, the system will slightly reduce the confidence weight of variants like "shape-similar substitution". For example, if the original weight is 0.6, the system can adjust it to 0.58 based on a preset learning rate.

[0088] Correspondingly, if the system misses a report due to failing to match a noisy blacklist, and this missed report event is subsequently discovered and reported back to the system, the system can also trace the cause. If it is found that the total score did not reach the threshold because the weight of a variant (e.g., a substitution of an uncommon homophone) was too low, the system can appropriately increase the weight of that variant type.

[0089] Through this "weight evolution" closed loop, the system can utilize real business feedback to continuously and automatically optimize the parameters in its risk scoring model, making its decision-making logic increasingly closer to actual risk scenarios. In this way, it can dynamically seek the optimal balance between false positive rate and false negative rate in a constantly changing environment, thereby achieving adaptive improvement in system performance.

[0090] Example 4

[0091] This embodiment describes a security enhancement scheme based on the system architecture of Embodiment 1, which aims to provide a higher level of security for the system in order to address the security risks that the initiating node itself may have.

[0092] In this embodiment, more stringent requirements are imposed on the system's deployment environment and the encryption protocol used to achieve end-to-end, hardware-based trust assurance.

[0093] Specifically, regarding the deployment of core modules, this embodiment explicitly requires that the modules on the initiating node 10 that process the most sensitive data and execute the most core algorithms be deployed within a trusted execution environment provided by the server's central processing unit. It is understood that a trusted execution environment is a secure area based on hardware isolation technology, such as Intel's Software Protection Extensions (SPE) or AMD's Secure Encryption Virtualization (SEVM) technology. It ensures that the code executed and the data processed within it are protected from theft and tampering by the operating system, virtual machine manager, or even physical attacks.

[0094] As a preferred implementation, at least one, or even both, of the multimodal normalization engine 110 and the cryptographic computing unit 130 of the initiating node 10 should be encapsulated and run in a secure “enclave” created by a trusted execution environment.

[0095] Its workflow may include:

[0096] When an alert task is initiated, raw identity data, which may contain user privacy information, is securely loaded from an external, untrusted environment into a trusted execution environment.

[0097] Within the protected “black box” of the Trusted Execution Environment, the multimodal normalization engine 110 performs all cleaning, normalization, and distortion calculation operations.

[0098] Next, also within the trusted execution environment, the fuzzy feature mapper 120 (if its logic is also deployed therein) or is invoked by a program within the trusted execution environment to perform variant generation and construct the Bloom filter vector to be queried.

[0099] The encrypted computing unit 130 loads its private key, salt value and other sensitive encrypted materials inside the trusted execution environment, and executes all encryption and computing steps belonging to the initiator in the privacy set intersection protocol.

[0100] All data that needs to communicate with the outside (i.e., data source node 20) is sent directly from within the trusted execution environment through a controlled interface.

[0101] With this deployment method, even if the server operating system of the initiating node 10 or other applications deployed on the same machine are compromised by malware, attackers will not be able to snoop on or tamper with the raw data being processed inside the trusted execution environment, the intermediate variant sets, the keys used for encryption, and the logic of the core algorithm, thereby greatly enhancing the data confidentiality and integrity of the system during the computing process.

[0102] Furthermore, regarding communication security between nodes, this embodiment explicitly stipulates that the secure channel between the initiating node 10 and the data source node 20 must employ an industry-recognized strong encryption transmission protocol, preferably Transport Layer Security (TLS) version 1.3 or higher. Compared to older versions, TLS version 1.3 discards outdated encryption algorithms, simplifies the handshake process, and improves security and performance. It can effectively resist network attacks such as eavesdropping, tampering, and message forgery, ensuring the security of PSI query request 301 and PSI query response 302 during wide area network transmission.

[0103] Furthermore, regarding the specific selection of the privacy set intersection protocol, this embodiment further limits the type of protocol adopted by the cryptographic computing unit 130. To achieve a higher level of security (e.g., security under a semi-honest or malicious model), a privacy set intersection protocol based on unintentional transmission or a privacy set intersection protocol based on RSA blind signatures can be selected. These protocols are cryptographically proven secure, ensuring that throughout the interaction, the initiating party learns no additional information about the data source set besides the intersection result, and the data source party learns no additional information about the initiating party's queried set besides necessary information such as the set size.

[0104] In summary, this embodiment constructs a full-link security enhancement scheme from computation to communication by placing the core computation in a trusted execution environment, adopting the latest transport layer encryption protocol, and selecting a strongly secure privacy computing protocol, providing the highest level of privacy protection and anti-attack capability for the entire early warning system.

[0105] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A fraud prevention and early warning system based on multi-source cross-platform fusion, characterized in that, include: A multimodal normalization engine is configured to receive multi-source heterogeneous original identity data with different structural formats from different data source nodes, clean the original identity data based on a preset rule base, generate normalized standard data, and calculate a cleaning distortion index that reflects the change in information content before and after cleaning. A fuzzy feature mapper is configured to generate a set of multiple variants that meet a preset edit distance condition based on the normalized standard data when the cleaning distortion index meets a preset threshold, and to perform hash processing on the variant set to generate a fuzzy feature vector to be queried. An encrypted computing unit is configured to send the fuzzy feature vector to be queried to the data source node through a secure channel, and perform a privacy set intersection operation in the ciphertext domain to obtain the ciphertext overlap result. And a risk decision module, configured to calculate the final risk confidence score based on the ciphertext overlap result, preset variant weights and time decay factors through a weighted risk scoring model, and generate an early warning instruction based on the score; The cleaning distortion index is the ratio of the Levinstein edit distance between the original identity data and the normalized standard data to the length of the original identity data.

2. The system according to claim 1, characterized in that, It also includes an intelligent normalization agent module, configured to invoke a large language model to perform semantic processing on the original identity data when the cleaning distortion index exceeds a preset threshold or the risk confidence score is in a preset fuzzy range.

3. The system according to claim 2, characterized in that, The intelligent normalization proxy module also includes a dynamic rule generation unit, configured to generate regular expressions for recognizing similar noise patterns based on the processing results of the large language model, and update them to the rule base of the multimodal normalization engine.

4. The system according to claim 1, characterized in that, It also includes a feedback closed-loop unit, configured to adjust the preset variant weights of different variant types in the risk decision module in reverse based on the false alarm or missed alarm results fed back by the external business system.

5. The system according to claim 1, characterized in that, The weighted risk scoring model is used to calculate the risk confidence score by multiplying the corresponding variant weights of all variants of the ciphertext overlap results by a time decay factor, summing the results, and then inputting the sum into a Sigmoid activation function.

6. The system according to claim 1, characterized in that, The preset edit distance condition is as follows: the Levenstein edit distance between the generated variant and the normalized standard data is less than or equal to K; for numeric data, K is 1 or 2; for text data, K is within 20% of the length of the normalized standard data.

7. The system according to claim 1, characterized in that, The encrypted computing unit employs either an RSA blind signature-based or a privacy set intersection protocol based on unintentional transmission.

8. The system according to claim 1, characterized in that, At least one of the multimodal normalization engine and the cryptographic computing unit is deployed within the trusted execution environment of the initiating node.

9. The system according to claim 1, characterized in that, The secure channel uses a transport layer security protocol 1.3 or a higher version of encrypted transport protocol.

Images