An ai full-link privacy protection decision system and method

Abstract

This invention discloses an AI end-to-end privacy protection decision-making system based on standardized layout and dual-path isolation, comprising: a standardized end-side device with a built-in security chip, generating standardized electronic documents and anonymized electronic seals containing irreversible large tokens; a data desensitization unit, used to receive the anonymized electronic seals, strip all link metadata, and generate random anonymous IDs; an AI decision-making unit, configured to perform AI inference based on the large tokens, generate decision results, and apply differential privacy noise to the decision results before returning them to the data desensitization unit; and a result relay platform, configured to temporarily store decision results and support user-initiated retrieval, with records immediately destroyed after retrieval; wherein, the data desensitization unit forwards the anonymous IDs and encrypted large tokens to the AI ​​decision-making unit, receives the decision results from the AI ​​decision-making unit, and pushes them to the result relay platform, never directly returning them to the user.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data security technology, and specifically to a privacy protection decision-making system. Background Technology

[0002] With the deep application of AI big data models in sensitive fields such as healthcare, finance, and government affairs, how to protect the privacy of individuals and organizations while utilizing the value of data has become a core bottleneck for industrial development. Hu Jinqian disclosed a comprehensive integrated big token data management system and method in patent application CN202511774541.4, proposing to aggregate and encapsulate digital asset packages into a single, verifiable big token data structure.

[0003] Existing technologies mainly fall into two unrelated categories, each with its own fatal flaws: The first category is post-processing anonymization based on unstructured data. This type of solution typically relies on OCR technology to perform real-time parsing and anonymization of uploaded non-standard files such as PDFs and images before AI processing. However, this solution faces three major challenges: enormous computational consumption, high risk of anonymization omissions, and lack of verifiability. The second category is anonymous communication solutions based on link isolation. This type of solution focuses on hiding network layer information such as the user's IP address, but ignores the two key issues of unanonymized source data and exposure of the return link, failing to form a truly meaningful "end-to-end" privacy protection loop. Therefore, there is an urgent need for an innovative technical solution that deeply integrates data format standardization with dual isolation of communication paths.

[0004] It should be noted that the above description of the technical background is only for the purpose of providing a clear and complete explanation of the technical solutions of the present invention and facilitating understanding by those skilled in the art. It should not be assumed that the above technical solutions are known to those skilled in the art simply because they have been described in the background section of this invention. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an AI end-to-end privacy protection decision system and method based on layout standardization and dual-path isolation.

[0006] This application discloses an AI end-to-end privacy protection decision-making system based on standardized format and dual-path isolation. It comprises: a standardized end-side device with a built-in security chip, configured to generate standardized electronic documents according to a predefined electronic document format standard and generate an anonymized electronic seal including an irreversible large token through the security chip; a data desensitization unit, used to receive the anonymized electronic seal, strip all link metadata, and generate a random anonymous ID (AID); an AI decision-making unit, configured to perform AI inference based on the large token, generate decision results, apply differential privacy noise to the decision results, and return them to the data desensitization unit; and a result relay platform, configured to temporarily store the decision results and support user-initiated retrieval using a one-time password or QR code, with immediate destruction of the records after retrieval; wherein the data desensitization unit forwards the anonymous ID and encrypted large token to the AI ​​decision-making unit, and the data desensitization unit receives the decision results from the AI ​​decision-making unit and pushes them to the result relay platform, never directly returning them to the user.

[0007] Furthermore, the above system satisfies the upper bound constraint of full-link mutual information: MI(S;R)≤E to quantify the degree of anonymization achieved by the system; where S represents the original set of subject identity information, R represents the final decision result returned to the user, MI(;) is Shannon mutual information, and E is a preset minimal positive number used to characterize the upper limit of information leakage allowed by the system.

[0008] Furthermore, the aforementioned standardized edge device also includes: a standard setting engine, used to predefine the unified format, field naming conventions, and physical coordinates of electronic documents; a document generator, used to generate formatted electronic documents based on the aforementioned standard templates and forcibly inject a machine-readable standard version number into the document metadata; and an anonymized electronic seal module, built into the aforementioned standardized edge device, used to accurately locate and extract all organizational information fields using a pre-built organizational information location map, and aggregate them into an irreversible semantic large token.

[0009] Furthermore, the aforementioned anonymized electronic seal module also includes: generating proof that the source of the aforementioned large token is legitimate and conforms to the format standard; generating a one-time extraction password P and its hash value H(P) bound to this request, and generating a QR code including an anonymous ID placeholder and H(P); and immediately physically overwriting the memory of the aforementioned security chip after the anonymized electronic seal is generated, completely destroying the original file, the main fingerprint, and the temporary key.

[0010] Furthermore, the aforementioned data desensitization unit also includes: removing all network layer metadata; generating a random anonymous ID (AID) that is independent of the subject and encrypting the storage mapping relationship using a periodically rotating key; introducing random time delays and disrupting time correlation; and ensuring that the upload path entry node and the return path exit node are logically and physically completely isolated with no data overlap.

[0011] Furthermore, the aforementioned data desensitization unit includes a timing disturbance controller configured to generate random variables based on hardware random numbers, so that the forwarding delay of each request conforms to an exponential distribution, thereby eliminating the linear correlation between upload time and result return time.

[0012] Furthermore, the aforementioned security chip generates a large token by performing semantic compression and irreversible mapping operations using a semantic aggregation function T=F(X). Here, X=[X1,X2,X3,…] is an organization attribute vector precisely extracted from a standardized file, and F is a non-injective and irreversible semantic aggregation function. Its output T is a large token that satisfies the conditional entropy H(X|T)=H(X), indicating that T does not carry any reversible information about X.

[0013] Furthermore, the differential privacy noise applied by the aforementioned AI decision-making unit satisfies the following randomization mechanism M for any adjacent datasets D, D': Da(M(D)∣∣M(D'))≤B. Here, M is the AI ​​decision-making mechanism, Da(.∣∣.) is the Rényi divergence, a>1 is the order, and B is the privacy loss.

[0014] Furthermore, the format standards of the aforementioned standardized electronic documents are compatible with the Extensible Markup Language Schema (XML Schema) of the World Intellectual Property Organization (WIPO) ST.36 / ST.96 standard, or the Open Fixed Layout Document (OFD) of the People's Republic of China National Standard GB / T 33190-2016 "Electronic Document Storage and Exchange Format".

[0015] This application also discloses an AI end-to-end privacy protection decision-making method based on format standardization and dual-path isolation, including the following steps: S1, on a standardized end-device, a standardized electronic document is generated according to a predefined electronic document format standard; S2, using a built-in security chip, organizational information is accurately extracted based on an organizational information location map to generate an anonymized electronic seal including an irreversible large token, and its conditional entropy constraint is verified; S3, the above-mentioned anonymized electronic seal is uploaded to a data desensitization unit, which removes link metadata, generates a random anonymous ID, applies an exponential distribution delay, and ensures physical isolation between the upload and return paths; S4, the AI ​​decision-making unit performs AI inference based on the above-mentioned large token, generates a decision result, and applies Reni differential privacy noise; S5, the data desensitization unit pushes the above-mentioned decision result to a result relay platform; S6, the user actively retrieves the decision result from the result relay platform with a one-time password, and the platform immediately destroys the record after retrieval; wherein, the entire process ensures final mutual information.

[0016] The beneficial effects of this invention are as follows.

[0017] End-to-end dead loop protection: For the first time, "format standardization of data source" and "dual-path isolation of communication link" are organically integrated.

[0018] Extremely efficient privacy protection: Eliminates the need for expensive OCR, reduces computing power consumption by 60% to 80%, and increases processing speed by more than 5 times.

[0019] Complete subject anonymization: theoretically guaranteed by the mathematically formalized upper bound of mutual information MI(S;R)≤E and conditional entropy H(X|T)=H(X).

[0020] Strong verifiability and compliance: It complies with the stringent requirements of the EU's General Data Protection Regulation (GDPR) and other regulations regarding "anonymized" data.

[0021] High engineering feasibility: This demonstrates that mechanisms such as caching and QR code extraction balance security and user experience. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of an AI end-to-end privacy protection decision system based on layout standardization and dual-path isolation in one embodiment of the present invention.

[0023] Figure 2 This is a flowchart of an AI end-to-end privacy protection decision-making method based on layout standardization and dual-path isolation in one embodiment of the present invention.

[0024] The reference numerals in the above figures are as follows:

[0025] The system comprises: a decision-making system 10, a standardized end-side device 110, a data anonymization unit 120, an AI decision-making unit 130, a result transfer platform 140, a security chip 111, a timing disturbance controller 121, a standard setting engine 112, a document generator 113, an anonymized electronic seal module 114, and steps S1 to S6. Detailed Implementation

[0026] To better understand this invention, the following embodiments are provided in conjunction with the accompanying drawings. It should be understood that the embodiments of this invention are for illustrative purposes only and not for limiting the invention; the scope of protection of this invention is defined solely by the claims. The embodiments provided are merely preferred embodiments and are not intended to limit the invention in any way. Those skilled in the art can make changes, equivalent substitutions, or modifications based on the content of this invention to form different implementations. However, any changes and modifications, and any equivalent substitutions made to the method of this invention without departing from the inventive concept are within the scope of protection of this invention.

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, and / or combinations thereof.

[0029] First, please refer to Figure 1 , Figure 1 This is a schematic diagram of an AI end-to-end privacy protection decision system 10 based on layout standardization and dual-path isolation in one embodiment of the present invention, as shown below. Figure 1 As shown, the AI ​​end-to-end privacy protection decision system 10 based on layout standardization and dual-path isolation of the present invention includes: a standardized end-side device 110, a data desensitization unit 120, an AI decision unit 130, and a result transfer platform 140.

[0030] The standardized edge device 110 has a built-in security chip 111. The standardized edge device 110 is configured to generate standardized electronic documents according to a predefined electronic document format standard, and to generate an anonymous electronic seal including an irreversible large token through the security chip 111. In one embodiment of the present invention, the large token can be an anonymous semantic fingerprint, but the present invention is not limited thereto.

[0031] The above-mentioned standardized electronic document formatting standards are compatible with the Extensible Markup Language Schema (XML Schema) of the World Intellectual Property Organization (WIPO) ST.36 / ST.96 standard, or the Open Fixed Layout Document (OFD) of the People's Republic of China National Standard GB / T 33190-2016 "Electronic Document Storage and Exchange Format".

[0032] The aforementioned data anonymization unit 120 consists of multiple distributed nodes. The aforementioned data anonymization unit 120 is configured to receive the aforementioned anonymized electronic seal, strip all link metadata, generate a random anonymous ID, and forward the aforementioned anonymous ID and the encrypted large token to the aforementioned AI decision-making unit 130.

[0033] In one embodiment, the encryption algorithm may be such as the Advanced Encryption Standard (AES-256), but this is not a limitation of the present invention. Any encryption algorithm that conforms to the spirit of the present invention falls within the scope of the present invention.

[0034] Meanwhile, the aforementioned data anonymization unit 120 receives the decision results from the aforementioned AI decision-making unit 130 and pushes them to the result relay platform 140, never directly returning them to the user.

[0035] For example, in this embodiment of the invention, the data desensitization unit 120 can be used to: remove all network layer metadata; generate a random anonymous ID that is independent of the subject and use a periodically rotating key to encrypt and store the mapping relationship; introduce random time delays to destroy time correlation; and ensure that the upload path entry node and the return path exit node are completely isolated logically and physically, with no data overlap.

[0036] In one embodiment of the present invention, the data desensitization unit 120 includes a timing perturbation controller 121, configured to generate random variables based on hardware random numbers, so that the forwarding delay of each request conforms to an exponential distribution, thereby eliminating the linear correlation between upload time and result return time. Because the exponential distribution has a memoryless property, this means that no matter how long the data has been waiting in the data desensitization unit 120, the probability distribution of being sent out in the next second is constant. Based on the timing decoupling principle of the memoryless nature of the exponential distribution, in the present invention, the timing perturbation controller 121 built into the data desensitization unit 120 introduces a random delay conforming to an exponential distribution, mathematically severing the temporal causal chain between request upload and result return.

[0037] The AI ​​decision-making unit 130 is configured to perform AI inference based on the large token, generate decision results, apply differential privacy noise to the decision results, and then return them to the data desensitization unit 120. This ensures that the decision results can have statistical ambiguity, making it impossible to precisely distinguish which specific entity generated them even if the decision results are obtained.

[0038] The above result relay platform 140 is configured to temporarily store the above decision results and support users to actively retrieve them with a one-time password or QR code, and destroy the records immediately after retrieval.

[0039] It is worth noting that the entire system satisfies the following upper bound constraint on end-to-end mutual information: MI(S;R)≤E; where S represents the original set of subject identity information (such as the user's name, ID number, device unique identifier, etc., which can point to a specific individual), R represents the final decision result returned to the user (such as loan amount, medical diagnosis suggestion, etc.), MI(;) is Shannon mutual information, used to measure how much information about the subject identity S can be obtained from the result R, and E is a preset very small positive number (e.g., E bits), used to quantify the degree of anonymization achieved by the system, representing the upper limit of the amount of information that the system is allowed to leak. The closer E is to 0, the higher the degree of anonymization.

[0040] It is worth noting that the aforementioned standardized end-side device 110 also includes a standard setting engine 112, a document generator 113, and an anonymized electronic seal module 114.

[0041] The aforementioned standardization engine 112 is used to predefine the unified format, field naming conventions, and physical coordinates of electronic documents. The main function of the standardization engine 112 is to establish data standards. For example, the unified format defines the structure of the file (such as Extensible Markup Language XML or Portable Document Format PDF / Archive (PDF / A)) and ensures compliance with international standards (such as WIPO ST.36 / ST.96). The aforementioned field naming convention unifies the naming of information from different sources but with the same meaning (e.g., unifying "Name," "User Name," and "Customer Name" as Identity_Name). The aforementioned physical coordinates are crucial to this invention. The standardization engine 112 predefines the precise spatial location of each sensitive piece of information on the page (e.g., (x:10, y:20) to (x:50, y:25) must be the document number area), which eliminates the need for full-text OCR recognition in subsequent processing, thus significantly saving computational power and improving accuracy.

[0042] The aforementioned document generator 113 is used to generate formatted electronic documents according to the aforementioned standard template and to forcibly inject a machine-readable standard version number into the document metadata. For example, the aforementioned document generator 113 is a processing unit that transforms raw data into "standardized media". The aforementioned template generation is based on the template provided by the aforementioned standard setting engine 112, filling the content input by the user into the corresponding physical space. The aforementioned machine-readable version number is a version tag that is forcibly injected into the document's metadata.

[0043] It is worth noting that the generation and maintenance of the above location map are achieved through automated training and dynamic coordinate mapping. First, the standard setting engine 112 uses a convolutional neural network (CNN) to analyze the physical coordinates of the organization information in the pre-calibrated standard layout to complete the initial construction. Subsequent maintenance is carried out in the security chip 111 by monitoring the file version number and using spatial transformation matrix and translation vector to perform second-level calibration and adaptive updates of the coordinates to ensure that the positioning accuracy is always maintained at less than 0.01mm.

[0044] In one embodiment, when a file enters a subsequent node, the system only needs to read this version number to immediately load the corresponding "coordinate map" for precise desensitization, thus avoiding privacy leaks caused by changes in the layout.

[0045] It is worth noting that the aforementioned anonymized electronic seal module 114 is built into the standardized edge device 110. It is used to accurately locate and extract all organizational information fields using a pre-constructed organizational information location map, aggregating them into an irreversible semantic large token. In this way, all extraction and computation involving original sensitive data are performed in an independent environment protected by hardware, preventing external attackers from extracting plaintext through operating system vulnerabilities. The aforementioned anonymized electronic seal module 114 can accurately locate all organizational information fields using the pre-constructed map, much like a shield. Moreover, this design of the present invention completely changes the inefficient and high-risk traditional AI privacy protection model that relies on "upload first, then identify, then de-identify," achieving true end-to-end security.

[0046] In addition, the aforementioned anonymized electronic seal module 114 can also be used to generate zero-knowledge proofs (ZKPs) that prove the legitimacy of the source of the large token and that it conforms to the format standards. For example, while generating the large token, the anonymized electronic seal module 114 uses cryptographic algorithms (such as zero-knowledge succinct non-interactive argument of knowledge, zk-SNARKs) to construct a mathematical proof that the large token was extracted strictly according to the coordinates and rules defined by the aforementioned standard-setting engine 112, ensuring the correctness of the data structure. In this way, it can prove to the aforementioned AI decision-making unit 130 that this data was indeed generated by the certified standardized edge device 110 security device, and not malicious data forged by hackers.

[0047] The aforementioned anonymized electronic seal module 114 can also be used to generate a one-time extraction password P and its hash value H(P) bound to this request, and to generate a QR code including an anonymous ID placeholder and H(P). For example, the dynamic credential and QR code generation break the unique association of the backhaul link. In one embodiment, the aforementioned anonymized electronic seal module 114 randomly generates a strong password P locally and calculates its one-way hash value H(P). Only H(P) is encapsulated in the anonymous seal and sent out, while the original password P remains with the user (or in the QR code). After the seal is generated, the aforementioned anonymized electronic seal module 114 immediately performs a physical-level overwrite of the security chip memory, completely destroying the original file, the main fingerprint, and the temporary key.

[0048] For example, the aforementioned anonymized electronic seal module 114 generates a QR code containing H(P) and reserves a placeholder space for an "Anonymous ID (AID)". After the data is assigned an AID by the aforementioned data desensitization unit 120, the AID is backfilled to the relay platform. In this way, once the seal is generated and issued, the security chip immediately initiates the destruction process. This is not a logical "deletion mark", but rather the use of random numbers to repeatedly overwrite the memory area storing the original data, the subject fingerprint, and the temporary key (typically conforming to the US National Industrial Security Standard DoD 5220.22-M standard). In this invention, the objects destroyed include: the original file, to prevent plaintext recovery after the device is lost; the subject fingerprint, to destroy the unique attribute feature used to generate large tokens, preventing reverse tracing; and the temporary key, to ensure that even if future encryption technology is cracked, previously sent data cannot be recovered. In this way, this invention achieves "task completed, traces zeroed out". Even if a hacker subsequently compromises the end device, they will not be able to find any privacy clues about the transaction.

[0049] In one embodiment of the present invention, the security chip in the aforementioned standardized end-side device generates a large token through semantic compression and irreversible mapping operations using a semantic aggregation function T=F(X). Here, X=[X1,X2,X3,…] is an organizational attribute vector precisely extracted from the standardized file, and F is a non-injective and irreversible semantic aggregation function whose output T is a large token that satisfies the conditional entropy H(X|T)=H(X). That is, the conditional entropy is equal to the original entropy to indicate that T does not carry any reversible information about X. This means that the possibility of obtaining the subject information X from the token is theoretically reduced to zero.

[0050] In one embodiment of the present invention, the differential privacy noise applied by the AI ​​decision-making unit 130 satisfies Rényi Differential Privacy (RDP), that is, for any adjacent datasets D, D', the randomization mechanism M applied satisfies Da(M(D)||M(D'))≤B. Here, M is the AI ​​decision-making mechanism, Da(.||.) is the Rényi divergence, a>1 is the order, and B is the privacy loss. In the AI ​​decision-making process of the present invention, even if the input data has been anonymized, attackers may still infer the characteristics of the original data by observing subtle changes in the output and using statistical methods. Rényi Differential Privacy (RDP) is a more accurate privacy quantification standard than traditional differential privacy, and is only one embodiment used in the present invention, not a limitation thereof. Those skilled in the art should understand the operation of Rényi Differential Privacy in the present invention, and will not be elaborated further here.

[0051] Please refer to Figure 2 , Figure 2This is a flowchart of an AI end-to-end privacy protection decision-making method based on layout standardization and dual-path isolation, according to an embodiment of the present invention. Figure 2 As shown, the present invention provides an AI end-to-end privacy protection decision-making method based on standardized format and dual-path isolation, comprising the following steps: S1. On a standardized end-device, a standardized electronic document is generated according to a predefined electronic document format standard; S2. Using a built-in security chip, organizational information is accurately extracted based on an organizational information location map, generating an anonymized electronic seal including an irreversible large token, and verifying that it meets conditional entropy constraints; S3. The anonymized electronic seal is uploaded to a data desensitization unit, which removes link metadata, generates a random anonymous ID, applies an exponential distribution delay, and ensures physical isolation between the upload and return paths; S4. The AI ​​decision-making unit performs AI inference based on the large token, generates a decision result, and applies Reni differential privacy noise; S5. The data desensitization unit pushes the decision result to a result relay platform; S6. The user actively retrieves the decision result from the result relay platform using a one-time password, and the platform immediately destroys the record after retrieval; wherein, the entire process ensures final mutual information.

[0052] Example 1: Cross-border patent application

[0053] Application scenario: Chinese companies submit invention patent applications to the European Patent Office.

[0054] S1. On the aforementioned standardized terminal device 110, a standardized electronic document is generated according to a predefined electronic document format standard. The corporate legal department operates on the local computer (i.e., the aforementioned standardized terminal device 110). The aforementioned standard setting engine 112 built into the aforementioned standardized terminal device 110 predefines a template conforming to the WIPO ST.36 XML Schema. The aforementioned document generator 113, based on this template, fills in the user input content to generate a standardized XML format application document and injects it into the metadata. <standardversion> ST36_V1.0< / standardversion> .

[0055] S2. Utilizing the built-in security chip, organizational information is precisely extracted based on the "Organizational Information Location Map" to generate an anonymized electronic seal including an irreversible large token. The security chip 111 of the standardized end-side device 110 is activated. The anonymized electronic seal module 114 within it reads the version number from the file metadata and loads the corresponding "Organizational Information Location Map". This map precisely locates the applicant field at coordinates (x:15mm, y:25mm) and extracts "XX Technology Co., Ltd." The anonymized electronic seal module 114 maps this to an irreversible large token T={applicant_tier:"Tier_1",jurisdiction:"CN"} using the semantic aggregation function F. After verifying that the conditional entropy H(X|T)=H(X) holds, the anonymized electronic seal is generated.

[0056] S3. The anonymized electronic seal is uploaded to the data desensitization unit 120. The data desensitization unit 120 strips the link metadata, generates a random anonymous ID, applies an exponentially distributed delay, and ensures physical isolation between the upload and return paths. The anonymous seal is uploaded to the data desensitization unit 120. The link stripping module of the data desensitization unit 120 removes all network layer metadata such as IP addresses and device fingerprints; the anonymous ID generation module creates a random AID A7B3C9; the timing disturbance controller 121 uses hardware random numbers to make the forwarding delay of this request follow an exponential distribution (e.g., a delay of 2.3 minutes). The entire process ensures physical isolation between the upload entry and the return exit.

[0057] S4. The AI ​​decision-making unit performs AI inference based on the aforementioned large token, generates a decision result, and applies Reni differential privacy noise. After a delay, (A7B3C9, Encrypted_T) is sent to the AI ​​decision-making unit 130. After decryption, the AI ​​decision-making unit 130 performs AI inference based on the label in the large token T, generating a decision result. Subsequently, the AI ​​decision-making unit 130 applies Laplace noise satisfying (α=8, β=0.6)-RDP to the result.

[0058] S5. The aforementioned data anonymization unit 120 pushes the decision result to the result relay platform. The noisy result (A7B3C9, {"classification":"G06T"}) is returned to the aforementioned data anonymization unit 120. After receiving it, the aforementioned data anonymization unit 120 immediately pushes it to the aforementioned result relay platform 140, without storing any records itself.

[0059] S6. The user actively retrieves the decision result from the aforementioned result transfer platform 140 using a one-time password. The platform immediately destroys the record after retrieval. Legal personnel receive a one-time retrieval password P generated by the aforementioned standardized end-side device 11 on their mobile phone. They log into the aforementioned result transfer platform 140, enter P, and successfully download the decision result. Upon completion of data transmission, the aforementioned result transfer platform 140 immediately performs physical-level destruction, deleting all related records.

[0060] Example 2: Compliance Review of International Drug Clinical Trial Reports

[0061] Application scenario: CRO companies submit Clinical Study Reports (CSRs) to the U.S. FDA for AI systems to automatically review their format and content compliance.

[0062] S1 and CSR are generated as XML files by the file generator 113 of the standardized end-side device 110 according to the WIPO ST.96 standard, with metadata injected. <standardversion> ST96_MED_V2.1< / standardversion> .

[0063] S2. The aforementioned anonymized electronic seal module 114 of the standardized end-side device 110 loads the ST.96 map and accurately extracts... <study-sponsor name>The field generates a large token: {sponsor_type:"Biotech_SME",region:"APAC"}.

[0064] S3. The anonymous stamp is uploaded to the aforementioned data desensitization unit 120. After metadata is stripped, AID X8Y2Z1 is assigned, and an exponential delay is applied, it is forwarded to the aforementioned AI decision-making unit 130.

[0065] S4. The aforementioned AI decision-making unit 130 matches applicable regulations (such as FDA 21 CFR Part 11) based on the token, outputs a compliance report, and adds RDP noise to obscure specific violation details.

[0066] S5. The results are pushed from the data desensitization unit 120 to the result transfer platform 140.

[0067] S6. The CRO project manager retrieves the report from the result transfer platform 140 using the QR code generated by the standardized end-side device 110, and the result transfer platform 140 then destroys the data.

[0068] The primary advantage of this invention lies in eliminating the privacy risks and efficiency bottlenecks caused by chaotic data formats at the source. By forcibly adopting internationally standardized formats such as WIPO ST.36 / ST.96 to generate electronic documents, unstructured raw data is transformed into structured information that can be accurately parsed by machines. This allows the anonymization module built into the security chip of the edge device to directly, cost-free, and without omission locate and extract sensitive fields, completely eliminating reliance on unreliable technologies such as OCR, and significantly improving the accuracy, efficiency, and verifiability of anonymization.

[0069] Secondly, this invention constructs a truly comprehensive end-to-end anonymity protection system. It achieves both physical and logical isolation of the upload and return paths through a data desensitization unit: during upload, all metadata is stripped and a random delay is applied to disrupt time correlation; during return, the result is never directly connected to the user, but is pushed to an independent relay platform, where the user actively retrieves it using a one-time credential. This architecture completely blocks all potential side-channel attack paths, including those at the network layer, time dimension, and return link, ensuring that even if the AI ​​decision-making unit or communication link is compromised, the original subject's identity cannot be associated, achieving theoretically provable and engineering-practical end-to-end privacy protection.

[0070] The embodiments of the present invention described above can be implemented in various hardware, software codes, or combinations thereof. For example, embodiments of the present invention can also be program code executing the above methods in a Digital Signal Processor (DSP). The present invention can also relate to various functions executed by a computer processor, digital signal processor, microprocessor, or Field Programmable Gate Array (FPGA). The processor described above can be configured to perform specific tasks according to the present invention, which are accomplished by executing machine-readable software code or firmware code defining the specific methods disclosed in the present invention. The software code or firmware code can be developed into different programming languages ​​and different formats or forms. The software code can also be compiled for different target platforms. However, the different code styles, types, and languages ​​of the software code performing tasks according to the present invention and other types of configuration code do not depart from the spirit and scope of the present invention.

[0071] Therefore, those skilled in the art will recognize that although embodiments of the present invention have been shown and described in detail herein, many other variations or modifications conforming to the principles of the present invention can be directly determined or derived from the disclosure of the present invention without departing from the spirit and scope of the invention. Therefore, the scope of the present invention should be understood and recognized as covering all such other variations or modifications.

Claims

1. An AI-based end-to-end privacy protection decision-making system based on layout standardization and dual-path isolation, comprising: A standardized end-side device with a built-in security chip is configured to generate standardized electronic documents according to a predefined electronic document format standard and generate an anonymous electronic seal containing an irreversible large token through the security chip. A data desensitization unit is used to receive the anonymized electronic seal, strip all link metadata, and generate a random anonymous ID; An AI decision-making unit is configured to perform AI reasoning based on the large token, generate decision results, apply differential privacy noise to the decision results, and then return them to the data desensitization unit. A result transfer platform is configured to temporarily store the decision results and support users to actively retrieve them using a one-time password or QR code, and immediately destroy the records after retrieval; The data anonymization unit forwards the anonymous ID and the encrypted large token to the AI ​​decision-making unit. The data anonymization unit receives the decision result from the AI ​​decision-making unit and pushes it to the result relay platform, never directly returning it to the user.

2. The AI ​​end-to-end privacy protection decision-making system as described in claim 1, characterized in that: The format of the standardized electronic documents is compatible with the Extensible Markup Language (XML) framework of the World Intellectual Property Organization (WIPO) ST.36 / ST.96 standard, or the open format document of the People's Republic of China national standard.

3. The AI ​​end-to-end privacy protection decision-making system as described in claim 1, characterized in that, The data desensitization unit, Used to remove all network layer metadata; Used to generate a random, anonymous ID that is independent of the subject, and to encrypt and store the mapping relationship using a periodically rotating key; Used to introduce random time delays and disrupt time correlations; The upload path entry node and the return path exit node are logically and physically completely separated, with no data overlap.

4. The AI ​​end-to-end privacy protection decision-making system as described in claim 3, characterized in that, The data desensitization unit includes a timing disturbance controller configured to generate random variables based on hardware random numbers, so that the forwarding delay of each request conforms to an exponential distribution, thereby eliminating the linear correlation between upload time and result return time.

5. The AI ​​end-to-end privacy protection decision-making system as described in claim 1, characterized in that, The standardized end-side equipment also includes: A standards-setting engine is used to predefine the uniform format, field naming conventions, and physical coordinates of electronic documents; A file generator is used to generate formatted electronic files based on the standard template and to force the injection of a machine-readable standard version number into the file metadata; An anonymous electronic seal module is used to accurately locate and extract all organizational information fields using a pre-built organizational information location map, and aggregate them into an irreversible semantic big token.

6. The AI ​​end-to-end privacy protection decision-making system as described in claim 5, characterized in that: The anonymized electronic seal module Used to generate proof that the source of the large token is legitimate and conforms to the format standards; Used to generate a one-time extraction password and its hash value bound to this request, and to generate a QR code containing an anonymous ID placeholder and a hash value; This is used to physically overwrite the memory of the security chip immediately after the seal is generated, completely destroying the original file, the main fingerprint, and the temporary key.

7. The AI ​​end-to-end privacy protection decision-making system as described in any one of claims 1 to 6, characterized in that, The security chip generates a large token by performing semantic compression and irreversible mapping operations using a semantic aggregation function T=F(X); where X=[X1,X2,X3,…] is an organization attribute vector precisely extracted from a standardized file; F is a non-injective and irreversible semantic aggregation function; its output T is a large token and satisfies the conditional entropy H(X|T)=H(X), indicating that T does not carry any reversible information about X.

8. The AI ​​end-to-end privacy protection decision-making system as described in any one of claims 1 to 6, characterized in that, The system satisfies the upper bound constraint of full-link mutual information MI(S;R)≤E to quantify the degree of anonymization achieved by the system; where S represents the original set of subject identity information, R represents the final decision result returned to the user, MI(;) is the Shannon mutual information, and E is a preset minimal positive number used to characterize the upper limit of information leakage allowed by the system.

9. The AI ​​end-to-end privacy protection decision-making system as described in any one of claims 1 to 6, characterized in that, The differential privacy noise applied by the AI ​​decision center satisfies the following randomization mechanism M for any adjacent datasets D and D': Da(M(D)∣∣M(D'))≤B; where M is the AI ​​decision mechanism, Da(.∣∣.) is the Raney divergence, a>1 is the order, and B is the privacy loss.

10. An AI end-to-end privacy protection decision-making method based on layout standardization and dual-path isolation, comprising the following steps: S1. Generate standardized electronic documents on standardized terminal devices according to predefined electronic document format standards; S2. Using the built-in security chip, accurately extract organizational information based on the organizational information location map, generate an anonymous electronic seal containing an irreversible large token, and verify that it meets the conditional entropy constraint. S3. Upload the anonymized electronic seal to the data desensitization unit, whereby the data desensitization unit strips the link metadata, generates a random anonymous ID, applies an exponential distribution delay, and ensures physical isolation between the upload and return paths. S4. The AI ​​decision center performs AI reasoning based on the large token, generates decision results, and applies Reni differential privacy noise. S5. The data anonymization unit pushes the decision results to the result transfer platform; S6. Users can actively retrieve decision results from the result transfer platform using a one-time password. After retrieval, the platform immediately destroys the record. The entire process ensures the final mutual information.

Images