A scientific research data security protection method based on multi-model cooperation

Abstract

The application relates to the technical field of scientific research data security protection, and discloses a scientific research data security protection method based on multi-model cooperation, which comprises the following steps: obtaining a user identity attribute declaration and an access request parameter, generating an anonymous identifier and an attribute proof through zero-knowledge proof technology; analyzing the multi-modal composition structure of scientific research data; dispersively embedding the coded anonymous identifier into different modes of the scientific research data; constructing a revocable mapping mechanism between the anonymous identifier and the real identity, and performing sharded storage on the mapping relationship; when a data leakage event is detected, extracting watermark information from the leaked data to reconstruct the anonymous identifier; executing a multi-party authorization revocation protocol, and reconstructing the identity mapping relationship after authorization is obtained; through the innovative combination of zero-knowledge proof, multi-modal watermark and a revocable anonymous architecture, the application realizes the effective balance between data tracking and leakage accountability while protecting the privacy of scientific research data visitors.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of scientific research data security protection, more particularly, it relates to a scientific research data security protection method based on multi-model cooperation. BACKGROUND

[0002] In the process of scientific research data sharing, there are several key technical problems: on the one hand, the protection of the identity privacy of data access requires allowing anonymous or minimal identity information disclosure; on the other hand, in order to prevent data leakage, identity association is required for data distribution in order to track and hold accountable. At the same time, the watermark technology should have unforgeability and robustness, be able to resist various attacks, and need to support the integrity protection of cross-modal data. The existing technology is difficult to balance between these seemingly contradictory demands, often sacrificing privacy for accountability, or protecting privacy but weakening accountability. Specifically, the existing scientific research data sharing technology has three core contradictions:

[0003] The contradiction between identity verification and privacy protection: verifying the access qualification of the user usually requires collecting complete identity information, resulting in excessive exposure of privacy; the contradiction between anonymous access and responsibility tracing: allowing anonymous access protects privacy but makes it difficult to trace when a leak occurs, and recording identity loses privacy protection; the contradiction between watermark robustness and data usability: single-modal watermark is easy to be destroyed, and multi-modal watermark may affect data use.

[0004] In addition, the identity authentication system, watermark embedding system and accountability mechanism in the existing technology are independent of each other, lack unified architecture design, and cannot establish a data sharing mechanism that respects privacy and effectively holds accountable. SUMMARY

[0005] The present application provides a scientific research data security protection method based on multi-model cooperation, which solves the contradiction between identity verification and privacy protection, anonymous access and responsibility tracing, watermark robustness and data usability in related technologies, and the technical problem of lack of unified architecture design of independent identity authentication system, watermark embedding system and accountability mechanism.

[0006] The present application provides a scientific research data security protection method based on multi-model cooperation, which includes:

[0007] Obtain user identity attribute declaration and access request parameters, generate anonymous identifier and attribute proof through zero-knowledge proof technology;

[0008] Based on the anonymous identifier and the attribute proof, analyze the multi-modal composition structure of the scientific research data, and generate the corresponding watermark embedding strategy accordingly;

[0009] According to the watermark embedding strategy, the encoded anonymous identifier is dispersedly embedded into different modalities of the scientific research data, realizing multi-modal collaborative watermark embedding;

[0010] After completing the watermark embedding, an anonymous identifier is used to build a revocable mapping mechanism between the anonymous identifier and the real identity, and the mapping relationship is stored in fragments;

[0011] When a data leakage event is detected, the watermark information is extracted based on the leaked data to reconstruct the anonymous identifier;

[0012] In combination with the reconstructed anonymous identifier, a multi-party authorized revocation protocol is executed, the identity mapping relationship is reconstructed after authorization is obtained, and the accountability for data leakage behavior is realized.

[0013] Further, the steps of obtaining user identity attribute declaration and access request parameter, generating anonymous identifier and attribute proof through zero-knowledge proof technology, specifically include: preprocessing user identity attribute data, including normalizing numerical value type attribute, encoding conversion of category type attribute, and converting time type attribute to relative time stamp; generating verifiable attribute proof through zero-knowledge proof generation algorithm without leaking specific attribute value; generating unique anonymous identifier through hash chain algorithm.

[0014] Further, the step of analyzing the multi-modal composition structure of the scientific research data and generating the corresponding watermark embedding strategy, specifically includes: performing modal-specific preprocessing on each type of data to achieve consistency in dimension for different modal data; identifying feature regions of each modal using a multi-modal content analysis algorithm; calculating the amount of information that can be embedded for each modal using an adaptive multi-modal watermark capacity evaluation algorithm; and using Reed Solomon error correction coding algorithm to encode and fragment the identity information.

[0015] Further, in the step of multi-modal collaborative watermark embedding, the following watermark algorithms are used for different modalities: for text modal, semantic preservation watermark algorithm is used; for image modal, DCT domain watermark algorithm is used; for audio modal, spread spectrum watermark algorithm is used; for numerical data, least significant bit replacement algorithm is used.

[0016] Further, the step of building a revocable mapping mechanism between the anonymous identifier and the real identity, specifically includes:

[0017] The identity mapping relationship is divided by Shamir secret sharing algorithm;

[0018] A threshold encryption algorithm is used to ensure that the complete mapping relationship can only be reconstructed when a specific condition is met, wherein the specific condition means that only when no less than a preset threshold number of authorized nodes jointly submit the identity mapping relationship fragments they hold, the complete identity mapping relationship can be reconstructed through threshold decryption algorithm;

[0019] The identity mapping relationship is stored in different authorized nodes in a distributed manner.

[0020] Further, the step of reconstructing the anonymous identifier based on the extracted watermark information from the leaked data specifically comprises: using corresponding watermark extraction algorithms to extract watermark information from each modality in parallel; recovering damaged identity information fragments through error correction decoding algorithms; using a multi-modal cross-validation mechanism to integrate the watermark information extracted from each modality, so that the anonymous identifier can be reconstructed even if some modal watermarks are severely damaged or missing.

[0021] Further, the step of executing the multi-party authorized revocation protocol specifically comprises: confirming the legality of the revocation request through a threshold signature verification algorithm; recovering the complete mapping relationship from the fragments using a secret reconstruction algorithm based on the Lagrange interpolation principle; recording the entire revocation process through an audit log generation algorithm.

[0022] Further, in the step of reconstructing the anonymous identifier based on the extracted watermark information from the leaked data, to further enhance the data security monitoring capability, the step of monitoring the data access behavior in real time and updating the risk score is further included, specifically comprising: based on the real-time access log stream of the user, the behavior monitoring rule set and the historical behavior baseline model, the access mode is analyzed in real time through a sliding window algorithm; abnormal behavior detection is performed using the IsolationForest algorithm; the user risk score is dynamically updated by combining the historical score, the weight of the abnormal behavior type, the abnormality degree and the time decay factor through a user risk score dynamic updating algorithm.

[0023] Further, in the step of executing the multi-party authorized revocation protocol and reconstructing the identity mapping relationship after obtaining the authorization, to form a complete security closed loop, the step of generating a chain of accountability evidence and updating the protection strategy is further included, specifically comprising: based on the watermark extraction result, the identity mapping reconstruction result, the access behavior log and the context information of the leakage event, a complete leakage propagation path is constructed through an evidence correlation analysis algorithm; the responsible party is determined based on multi-source evidence using a responsibility identification algorithm; the security rules are updated according to the leakage mode through a protection strategy self-adaptive optimization algorithm.

[0024] The application provides a computer storage medium, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the above-mentioned scientific research data security protection method based on multi-model cooperation.

[0025] The application has the beneficial effects that: in the contradiction between identity verification and privacy protection, the application adopts attribute verification technology based on zero-knowledge proof, so that the user can prove that he or she meets the access qualification condition without revealing specific identity information, the "verification but not disclosure" mechanism effectively avoids the problem of excessive exposure of privacy caused by complete identity collection in the traditional method, and the coordination and unification of identity verification and privacy protection are realized;

[0026] In the contradiction between anonymous access and responsibility tracing, the application constructs a revocable anonymous watermark architecture, the mapping relationship between the anonymous identifier and the real identity is protected and divided by using Shamir secret sharing and threshold encryption technology, so that the user can access data in an anonymous manner under normal circumstances, and the anonymity can be revoked for responsibility tracing only after the leakage and authorization of multiple parties, the mechanism separates the identity information management and the responsibility tracing authority, forms an effective check and balance mechanism, and realizes the dynamic balance between anonymous protection and responsibility tracing;

[0027] In the contradiction between watermark robustness and data usability, the application realizes the consideration of high robustness of watermark and high usability of data by a multi-modal collaborative watermark mechanism, disperses and embeds the user identity information into different data modalities after Reed-Solomon encoding, and adopts the most suitable watermark algorithm (such as semantic preservation watermark for text, DCT domain watermark for image, spread spectrum watermark for audio and least significant bit replacement for numerical data) for different modalities, so that even if part of the modalities are damaged, the system can still recover the identity information from other modalities through cross-validation, and the reliability of the responsibility evidence is improved;

[0028] The application also effectively solves the problem that the identity authentication system, the watermark embedding system and the responsibility tracing mechanism are independent in the prior art, and constructs a complete and safe closed loop from access verification, data distribution, behavior monitoring to leakage responsibility tracing, so that each link forms an organic and unified protection system, the anonymous identifier generated by the zero-knowledge proof is directly used for watermark embedding, and the anonymous identifier extracted from the watermark can be mapped to the actual identity through multi-party authorization, so that the whole process is seamlessly connected, and a unified architecture design which protects privacy and effectively traces responsibility is formed. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is a flowchart of a scientific research data security protection method based on multi-model cooperation in the application;

[0030] Figure 2 is a column chart of the watermark capacity and robustness coefficient of different data modalities;

[0031] Figure 3 is a relationship diagram of the mapping relationship between the anonymous identifier and the real identity and the multi-party authorization mechanism;

[0032] Figure 4 is a line graph of the success rate of watermark extraction of different modal data when facing various intensity attacks;

[0033] Figure 5 is a scatter plot of the relationship between different types of abnormal behavior and risk scores;

[0034] Figure 6 is a Sankey diagram of the complete process from data access to responsibility tracing. DETAILED DESCRIPTION

[0035] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that discussions of these implementations are merely provided to enable those skilled in the art to better understand so as to be able to implement the subject matter described herein, and variations of elements discussed can be made by one skilled in the art. Individual examples of the subject matter described herein can omit, substitute, or add various procedures or components as appropriate, and not all implementations of the subject matter described herein will necessarily accomplish tasks in the same order or with the same components. Additionally, features described as taking place in a particular example can also occur in other examples unless otherwise noted.

[0036] A scientific research data security protection method based on multi-model cooperation is disclosed in at least one embodiment of the present application, as shown in Figure 1 includes:

[0037] Step 1, obtain user identity attribute declaration and access request parameters, generate anonymous identifier and attribute proof through zero-knowledge proof technology;

[0038] In this step, first, the identity attribute declaration submitted by the user (represented in JSON format), the access request parameter (containing the required permission level), and the attribute certificate issued by the trusted agency (in digital signature format) are obtained. Then, the zero-knowledge proof generation algorithm is used to process the above input data, to generate a verifiable attribute proof without revealing the specific attribute value, and at the same time, a unique anonymous identifier is generated through the hash chain algorithm. Finally, the zero-knowledge proof certificate (containing the attribute satisfaction proof), the unique anonymous identifier (256-bit hash value), and the access qualification verification result (Boolean value) are output.

[0039] Before processing the input data, the system will preprocess the user identity attribute data, including: normalizing numerical value type attributes, mapping numerical values of different ranges to the [0, 1] interval, encoding categorical attributes, using One-hot Encoding to convert non-numerical data into numerical vectors; converting time class attributes into relative time stamps. These preprocessing ensures that attribute data of different types and dimensions can be uniformly processed in the zero-knowledge proof system.

[0040] It should be noted that the above zero-knowledge proof generation algorithm is based on the Schnorr protocol. The zero-knowledge proof generation algorithm receives the identity attribute and related credentials of the user as input, outputs a zero-knowledge proof certificate, and enables the verifier to confirm that the user has the required attributes without knowing the specific attribute values.

[0041] In the present embodiment, the main parameters of the Schnorr protocol are set as follows: a 2048-bit prime number is selected as the main parameter, a 256-bit prime number is selected as the sub-parameter, and the sub-parameter can divide the main parameter minus one, and the order of the generator is equal to the sub-parameter. The security parameter (challenge value) in the protocol is limited in a larger numerical range, and the random number (commitment value) also has a definite value interval. In the process of generating the proof, the system sets a verification threshold, only when the verification score is higher than 0.95, it is considered that the attribute of the user meets the access condition.

[0042] The multi-round hash chain algorithm in the present embodiment is an improvement on the conventional hash chain algorithm. By introducing a multi-round iteration and timestamp combination mechanism, the security of the anonymous identifier is enhanced, so that the identifier is both unique and cannot be reversely derived from the original identity of the user. The multi-round hash chain algorithm receives user attribute information, a timestamp and a random salt value as input, and outputs a 256-bit anonymous identifier hash value.

[0043] In the present embodiment, the number of iterations of the multi-round hash chain algorithm is set to ten thousand. Each iteration uses the SHA-256 hash function for calculation. The timestamp used to generate the hash chain has a precision of milliseconds, and the length of the random salt value is 128 bits. In specific implementation, first, the hash value of the user attribute is used as the initial input, then in each iteration, the hash result of the last iteration is concatenated with the current timestamp and random salt value, and then the SHA-256 hash operation is performed on the concatenated result. After ten thousand iterations, the final hash value is the anonymous identifier. In the whole process, the string concatenation operation is used to combine the hash value, the timestamp and the salt value into a whole input, ensuring that each round of input has uniqueness and unpredictability.

[0044] Step 2, based on the anonymous identifier and the attribute proof, analyzing the multi-modal composition structure of the scientific research data, and generating the corresponding watermark embedding strategy accordingly;

[0045] In this step, first, the scientific research data to be shared (including text, image, audio, numerical value and other modalities), data structure meta-information (JSON format describes the distribution of each modality) and the anonymous identifier generated in step 1 are obtained. Then, the multi-modal content analysis algorithm is used to identify the feature area of each modality, the watermark capacity evaluation algorithm is used to calculate the amount of information that can be embedded in each modality, and the error correction coding algorithm is used to encode and fragment the identity information. Finally, output the multi-modal feature map (identify the embeddable area of each modality), the identity information encoding fragment set (each fragment contains part of the identity information and the check code), and the watermark embedding strategy table (specify the fragment index embedded in each modality).

[0046] Before performing multi-modal content analysis, the system first performs modality-specific preprocessing on each type of data: Term Frequency-Inverse Document Frequency processing is performed on text data to convert text into numerical feature vectors; image data is normalized to scale pixel values to the [0, 1] interval; frequency domain transformation and energy normalization are performed on audio signals; numerical data is standardized to convert to a distribution with a mean of 0 and a variance of 1. These preprocessing operations ensure the consistency of different modal data in dimension, allowing the multi-modal capacity evaluation algorithm to accurately calculate and compare the watermark carrying capacity of different modalities.

[0047] It should be noted that the above error correction coding algorithm is implemented using Reed-Solomon encoding. The Reed-Solomon encoding algorithm receives an anonymous identifier (256-bit hash value) as input and outputs an encoded symbol sequence with fault tolerance and its fragments, allowing the system to recover the original information in the case of partial data loss or tampering.

[0048] Feature area identification function in multi-modal content analysis algorithm is defined as the data modality In the feature space extract the region that can carry watermark, the feature area identification function based on information entropy and perceptual importance score returns the region coordinate set suitable for embedding watermark in each modality without affecting data usability.

[0049] Feature area identification function The information entropy calculation in the feature area identification function uses the Shannon entropy method. Specifically, the probability of each feature value appearing in the data is calculated, then all the probabilities are multiplied by their logarithms with base 2, and finally the negative sum of these products is taken to get the entropy value of the region. Information entropy reflects the degree of disorder of feature distribution, the more uniform the probability distribution, the higher the entropy value, indicating that the region is more suitable for embedding watermark. The feature area selection threshold is set to When , the corresponding region is identified as a feature area suitable for embedding watermark. Perceptual importance score The value in the interval is defined as the normalized value, and the calculation formula is:

[0050] ;

[0051] wherein represents the perceptual importance score, is the information entropy weight coefficient, is the edge complexity weight coefficient, is the content saliency weight coefficient, is the information entropy of the feature region , is the edge complexity of the feature region , and is the content saliency of the feature region ; in the embodiment, the information entropy of the feature region , , , , .

[0052] The adaptive multi-modal watermark capacity evaluation algorithm in the embodiment is an innovative algorithm designed for different data modal characteristics. According to the intrinsic characteristics of each modal data (such as the semantic redundancy of text, the complexity of image, the spectral characteristics of audio, and the precision of numerical data), the adaptive multi-modal watermark capacity evaluation algorithm dynamically calculates the amount of watermark information that each modal can carry, and optimizes the allocation strategy of watermark fragments based on the calculation results, thereby maximizing the robustness of the watermark while maintaining the usability of the data. The adaptive multi-modal watermark capacity evaluation algorithm receives data characteristic parameters of various modalities as input, and outputs watermark capacity values of each modality and an optimized watermark fragment allocation scheme.

[0053] The watermark capacity calculation formula of each modality in the adaptive multi-modal watermark capacity evaluation algorithm is as follows:

[0054] Text modality:

[0055] ;

[0056] wherein is the watermark capacity of the text modality, is the text modality capacity coefficient ( ), is the text length, is the semantic redundancy.

[0057] Image modality:

[0058] ;

[0059] wherein is the watermark capacity of the image modality, is the image modal capacity coefficient (set to 0.01), is the image width, is the image height, is the image complexity.

[0060] Audio modal:

[0061] ;

[0062] wherein is the audio modal capacity, is the audio modal capacity coefficient (set to 0.001), is the audio duration (in seconds), is the sampling frequency, is the masking threshold.

[0063] Numerical data:

[0064] ;

[0065] wherein is the numerical modal capacity, is the numerical modal capacity coefficient (set to 0.1), is the number of numerical values, is the available precision bits.

[0066] The watermark fragment allocation optimization employs a linear programming method, with the constraint that the amount of watermark information for each modality does not exceed its capacity upper limit, and the objective function is to maximize the overall watermark robustness score.

[0067] The complete mathematical expression of the watermark fragment allocation optimization is to maximize the objective function:

[0068] ;

[0069] wherein denotes the maximization, denotes the summation symbol, is the robustness coefficient of modality , is the number of watermark bits allocated to modality , is the time decay function of modality , is the expected preservation time.

[0070] The constraints include: the sum of the number of watermark bits allocated to all modalities is equal to the total length of the encoded identity information; the number of watermark bits allocated to each modality cannot exceed the upper limit of the watermark capacity of the modality; the number of watermark bits allocated to each modality cannot be less than the minimum necessary number of bits, which is usually set to 10% of the total number of bits; in addition, the number of watermark bits allocated to each modality must be a non-negative integer.

[0071] 、 、 、 respectively represent the time decay rates of text, audio, image and numerical modalities, reflecting the resistance of each modality to common attacks. In the present embodiment, 、 、 、 The optimization problem is solved by a branch and bound method, with a maximum iteration number of and a convergence precision of .

[0072] Step 3, according to the watermark embedding strategy, the encoded anonymous identifier is dispersedly embedded into different modalities of the scientific research data, realizing multi-modality collaborative watermark embedding;

[0073] In this step, first, the content of each modality of the scientific research data, the identity information encoding fragments generated in step 2, and the watermark embedding strategy table are obtained. Then, for different modalities, the corresponding watermark embedding algorithm is used to embed the identity information fragments into each modality, and the homomorphic encryption algorithm is used to protect the watermark mapping information. Finally, the multi-modality scientific research data embedded with the watermark, the encrypted watermark information distribution mapping table (recording the watermark position and extraction method of each modality) and the watermark embedding strength parameter set are output.

[0074] It should be noted that the present embodiment uses the following specific algorithms for watermark embedding of different modalities:

[0075] For text modality, a semantic-preserving watermark algorithm is used. The semantic-preserving watermark algorithm receives the original text content and the watermark information bit sequence as input, embeds the watermark information through synonym replacement, syntax structure adjustment, etc., and outputs the text content with unchanged semantics but containing watermark.

[0076] For image modality, a DCT domain watermark algorithm is used. The DCT domain watermark algorithm receives the original image data and the watermark information bit sequence as input, embeds the watermark by modifying the medium frequency components of the DCT coefficients, and outputs the image data with basically unchanged visual effect but containing watermark.

[0077] For audio modality, spread spectrum watermarking algorithm is used. The spread spectrum watermarking algorithm receives original audio signal and watermark information bit sequence as input, embeds watermark information into audio signal through spread spectrum sequence and masking threshold calculation, and outputs audio data containing watermark but with unchanged hearing effect.

[0078] For numerical data, least significant bit replacement algorithm is used. The least significant bit replacement algorithm receives original numerical data set and watermark information bit sequence as input, embeds watermark by replacing the least significant bit of numerical value, and outputs numerical data containing watermark with minimal precision impact.

[0079] In addition, the present embodiment uses Paillier homomorphic encryption algorithm to protect watermark information distribution mapping table. The Paillier homomorphic encryption algorithm receives watermark position information as input and outputs encrypted mapping table, ensuring that even if the mapping table is obtained, the specific position of the watermark cannot be known without holding the key. Among them, the encryption mapping function in the Paillier homomorphic encryption algorithm The Paillier homomorphic encryption algorithm is used to encrypt the message, where is the plaintext message, is a random number, and the encryption mapping function ensures that the encrypted data can be homomorphically operated without leaking the original position information.

[0080] In the Paillier homomorphic encryption algorithm, the public key parameter (i.e. the modulus of the Paillier encryption algorithm, which determines the security strength of the encryption system) is set to 2048 bits, is obtained by multiplying two large prime numbers of the same bit length (the first large prime number of the Paillier algorithm) and (the second large prime number of the Paillier algorithm). The parameter (the generator in the Paillier algorithm) needs to ensure that it can meet the correctness and security of the Paillier encryption algorithm. The parameter is and is the least common multiple. The parameter (a non-zero integer in the Paillier algorithm, which participates in key generation) is a non-zero integer. The random number (the random number used in Paillier encryption) needs to be selected within and requires and to be coprime. The plaintext message (the plaintext data to be encrypted) is limited within to ensure the security and correctness of the encryption operation.

[0081] AsFigure 2 As shown, the watermark capacity and robustness coefficient of different data modalities are compared. The core parameters of the multi-modal watermark allocation strategy in step 200 in the patent are reflected, the differences in the ability of different modal data to carry watermarks are embodied, and the roles of each modality in security protection are intuitively displayed.

[0082] Step 4, after completing the watermark embedding, an anonymous identifier is used to build a revocable mapping mechanism between the anonymous identifier and the real identity, and the mapping relationship is stored in fragments;

[0083] In this step, first, the anonymous identifier generated in step 1, the encrypted digest of the user's real identity, and the revocation key generated by the system are obtained. Then, the identity mapping relationship is divided into multiple fragments through a secret sharing algorithm, and a threshold encryption algorithm is used to ensure that the complete mapping relationship can only be reconstructed when certain conditions are met. Finally, the identity mapping relationship fragment set (distributed in different authorized nodes), the revocation authorization key share (distributed to each authorized party), and the mapping relationship metadata (including creation time and expiration conditions) are outputted, wherein the certain conditions are: only when no less than a preset threshold number of authorized nodes jointly submit the identity mapping relationship fragments they hold, can the complete identity mapping relationship be reconstructed through a threshold decryption algorithm.

[0084] It should be noted that the above secret sharing algorithm is implemented by using the Shamir secret sharing scheme. The Shamir secret sharing algorithm receives the mapping relationship between the anonymous identifier and the user's real identity as input, and outputs multiple shares, so that only a sufficient number of shares can be collected to reconstruct the original secret.

[0085] In the Shamir secret sharing scheme, the threshold parameter is set to , wherein is the total number of shares, represents the ceiling function. In this embodiment, is valued at , so , that is, at least shares are needed to reconstruct the secret. The polynomial degree is set to , which ensures that any shares cannot obtain any information about the secret. The polynomial coefficients ( ) are randomly selected in the finite field , wherein the prime number has a bit length of 2 times the secret length, and is at least 256 bits.

[0086] The calculation method of each share is: first, a polynomial is constructed with the secret to be shared as the constant term and the rest of the coefficients as random numbers. Then, a unique identifier is assigned to each share (the identifier is within a specified range). The identifier is substituted into the polynomial and the modulus is taken in the finite field to obtain the corresponding share value. In this way, only when a sufficient number of different shares are collected, the original secret can be restored through polynomial interpolation.

[0087] The present embodiment combines threshold encryption technology and time lock encryption technology to form an enhanced revocable mapping protection mechanism. The enhanced revocable mapping protection mechanism not only requires a predetermined threshold number of authorized parties to cooperate to decrypt the mapping relationship, but also sets a time constraint, so that even if enough authorized parties are obtained, the decryption cannot be completed before the predetermined time arrives, thereby providing double security protection. The enhanced revocable mapping protection mechanism receives a secret mapping relationship and a security parameter as input, and outputs an encrypted mapping relationship and a dispersed decryption key share.

[0088] In the enhanced revocable mapping protection mechanism, the threshold value is set as follows: the total number of authorized parties is multiplied by two-thirds and rounded up to obtain the minimum number of authorized parties required for cooperation. The time lock constraint serves as a difficulty parameter, representing the minimum computation time required for unlocking operations. In the present embodiment, the computational complexity of the time lock is set to one million consecutive square operations. The time lock function encrypts the mapping relationship plaintext, and only after a sufficient amount of computation is performed can the mapping relationship be unlocked. The specific implementation is based on the RSA algorithm, using a 2048-bit modulus. The logic of the decryption threshold verification is as follows: only when the number of valid decryption shares collected is not less than the threshold value , the system allows decryption operations, otherwise the decryption request will be rejected.

[0089] As shown in Figure 3 , the mapping relationship between the anonymous identifier and the real identity is displayed, as well as the multi-party authorization mechanism. For the revocable mapping mechanism corresponding to step 4, the role of secret sharing and threshold encryption technology in balancing user privacy and responsibility tracing is intuitively displayed.

[0090] In the embodiments of the present application, in order to further enhance the security of the system, the following steps can also be implemented:

[0091] Step 4.1, generating a temporary access token and behavior monitoring configuration;

[0092] Based on the zero-knowledge proof verification result of step 1, the anonymous identifier, and the data access policy configuration file, a time-limited access credential is created through a timestamp token generation algorithm, and a normal access behavior model is constructed and the allowed operation range is determined. The specific implementation is as follows:

[0093] Step 4.2, generating a time-limited access token;

[0094] In the JWT (JSON Web Token) format, it contains an anonymous identifier (without any information that can identify the real identity of the user), a permission range (allowing access to data types and operation types), a valid period (expiration time based on the current timestamp), a request source feature (such as IP range, device characteristics, etc.), and a digital signature (to ensure token integrity and non-falsifiability).

[0095] The attribute credential verification process in this step is implemented as follows: the system first verifies the digital signature of the trusted agency by RSA-PSS algorithm (parameters: salt length is 32 bytes, hash function is SHA-256) to confirm the authenticity of the user's submitted attribute credentials. After verification, the system uses the attribute value range mapping table to map the user attributes to the standard format that the system can handle, and the mapping table is stored in the secure key management module and obtained from the key server through the HTTPS protocol (TLS1.3). For numerical attributes, the min-max scaling method is used for normalization to map the values to the [0, 1] interval.

[0096] The specific implementation of the zero-knowledge proof generation algorithm is based on the Bulletproofs framework, using the elliptic curve secp256k1 as the underlying cryptographic primitive. This framework supports efficient zero-knowledge proofs for range proofs and predicate logic, with the following specific steps: convert user attributes to arithmetic circuit representation; use Pedersen commitment to hide attribute values; construct an arithmetic circuit that proves that the attributes satisfy the access conditions; use the Fiat-Shamir transformation to convert the interactive proof to a non-interactive format; generate a compact zero-knowledge proof. The system uses C++ to implement the core cryptographic primitives, interacts with the Python business logic layer through the CFFI interface, and the average time complexity of zero-knowledge proof generation is , where is the number of constraints included in the proof.

[0097] The behavior pattern baseline establishment algorithm in this embodiment is a multi-dimensional behavior feature model constructed based on user historical access data. This behavior pattern baseline establishment algorithm not only considers static access permissions, but also integrates dynamic features such as time, frequency, and operation sequence to construct a more detailed behavior baseline, providing a basis for subsequent anomaly behavior detection. The behavior pattern baseline establishment algorithm receives historical access logs and system parameters as input and outputs a multi-dimensional behavior baseline model. The feature extraction function in the behavior pattern baseline establishment algorithm is defined as from the historical access records in the time window The features include: access frequency , data volume, operation type distribution, etc.

[0098] Time window in behavior pattern baseline establishment algorithm The time window is set in the form of a sliding window, with the basic unit being an hour. In this embodiment, the window size is set to 24 hours (corresponding to short-term, medium-term, and long-term behavior patterns, respectively). The features include: access frequency (the number of accesses in a unit of time, standardized to the interval [0, 1]),data volume (the amount of data accessed in a unit of time, standardized to the interval [0, 1]),operation type distribution (vector representation of the proportion of each type of operation, with each component in the interval [0, 1] and the sum being 1).

[0099] When processing behavior features, the feature extraction function performs standardization processing on each feature. Specifically, the current value of each feature is subtracted from the historical minimum value of the feature, and then divided by the difference between the historical maximum and minimum values of the feature, thereby mapping all feature values to the interval [0, 1]. Here, the historical minimum and maximum values refer to the minimum and maximum values that the feature has taken in historical data.

[0100] The behavior baseline model uses a Gaussian Mixture Model (GMM) to represent it. The parameters of this model include the mean vector, the covariance matrix, and the mixing weights of each Gaussian component. These parameters are obtained by training the historical behavior data using the Expectation-Maximization (EM) algorithm, thereby effectively characterizing the user's multi-dimensional behavior feature distribution.

[0101] Step 5: When a data breach event is detected, extract watermark information based on the leaked data to reconstruct the anonymous identifier;

[0102] In this step, first, obtain the suspected leaked scientific research data sample, the encrypted watermark information distribution mapping table generated in step 3, and the watermark extraction authorization credential. Then, use a multi-modal watermark detection algorithm to extract watermark information from each modality in parallel, recover the possibly damaged identity information fragments using an error correction decoding algorithm, and confirm the integrity of the identity information using a cross-validation algorithm. Finally, output the extracted identity information fragment set, watermark integrity score, and preliminary anonymous identifier reconstruction result.

[0103] ​​It should be noted that the watermark extraction for different modalities adopts the extraction method corresponding to the corresponding embedding algorithm, as follows:

[0104] For the text modality, a semantic watermark extraction algorithm is used. The semantic watermark extraction algorithm receives the text embedded with the watermark as input, extracts the watermark information by analyzing the word replacement pattern, and outputs the extracted watermark bit sequence.

[0105] For the image modality, a DCT domain watermark extraction algorithm is used. The DCT domain watermark extraction algorithm receives the image embedded with the watermark as input, extracts the watermark information by analyzing the DCT coefficients, and outputs the extracted watermark bit sequence.

[0106] For the audio modality, a spread spectrum watermark extraction algorithm is used. The spread spectrum watermark extraction algorithm receives the audio signal embedded with the watermark as input, extracts the watermark information by calculating the cross-correlation with the spread spectrum sequence, and outputs the extracted watermark bit sequence.

[0107] For numerical data, a least significant bit extraction algorithm is used. The least significant bit extraction algorithm receives the numerical data embedded with the watermark as input, extracts the watermark information by analyzing the least significant bits, and outputs the extracted watermark bit sequence.

[0108] The multi-modal cross-validation mechanism in this embodiment is an innovative watermark extraction enhancement technology. The multi-modal cross-validation mechanism does not rely on the integrity of the watermark of a single modality, but by aggregating the watermark information extracted from each modality, comprehensively utilizing the Reed-Solomon error correction capability and cross-modal data consistency verification, it realizes the ability to reliably reconstruct the anonymous identifier even in the case of severe damage or loss of part of the modalities. The multi-modal cross-validation mechanism receives the watermark fragments extracted from each modality as input, and outputs the reconstructed anonymous identifier and the watermark integrity score.

[0109] Watermark integrity score in multi-modal cross-validation mechanism The calculation formula is as follows:

[0110] ;

[0111] Wherein is the multi-modal watermark integrity score; denotes the summation symbol; is the weight factor of modality is the watermark extraction reliability of modality

[0112] Watermark extraction reliability ​​The calculation method is: first, compare the Hamming distance between the original watermark and the extracted watermark, that is, count the number of inconsistent bits at the same position. Then, divide the number of inconsistent bits by the total length of the original watermark to get the proportion of error bits. Finally, subtract the proportion from 1 to get the credibility score of watermark extraction. That is, the closer the original watermark and the extracted watermark, the higher the credibility; if they are completely consistent, the credibility is 1, and if they are completely inconsistent, the credibility is 0.

[0113] Watermark reconstruction threshold Set to When , it is considered that the watermark extraction result is reliable; when , the enhanced recovery mode is enabled; when , it is determined that the watermark extraction fails. The error correction capability parameter of Reed-Solomon error correction coding is set to be able to correct no more than 30% of the total number of symbols.

[0114] Weight factor in multi-modal cross-validation mechanism It will be dynamically adjusted over time. Specifically, each modality (such as text, image, audio, and numerical value) will be assigned an initial weight at the beginning of watermark embedding, which is 0.25 for text, 0.30 for image, 0.25 for audio, and 0.20 for numerical value. As the time (months) after watermark embedding increases, the weight of each modality will gradually decay according to its time sensitivity coefficient. The time sensitivity coefficient reflects the stability of different modalities over time, and the larger the coefficient, the faster the weight decays. In this way, the system can dynamically reduce the dependence on modalities that are easily affected by time, and improve the long-term reliability of overall watermark extraction.

[0115] The time sensitivity coefficient of the text modality watermark is , which reflects that the text watermark decays at a low rate over time (i.e., the text replacement frequency is low, and the watermark is relatively stable); The time sensitivity coefficient of the image modality is , which belongs to medium sensitivity (common image processing operations may damage the watermark); The time sensitivity coefficient of the audio modality is , which has a high sensitivity (audio processing and compression operations are common, and the watermark is easily affected); The time sensitivity coefficient of the numerical value modality is , which belongs to the low-to-moderate sensitivity (numerical data modification is less, but accuracy adjustment occurs occasionally). According to the time sensitivity coefficient of each modality, the system adjusts the weight of each modality with time, reduces the weight of high sensitivity modality, and increases the weight of stable modality, so as to maximize the success rate of long-term watermark extraction. All dynamic weights are normalized.

[0116] As shown in Figure 4 , the change of watermark extraction success rate of different modal data in the face of various intensity attacks is shown, the watermark extraction and multi-modal cross-validation mechanism corresponding to step 5 verifies the anti-attack ability of the multi-modal collaborative watermark technology, and reflects the robustness technical effect in the patent implementation mode.

[0117] In the embodiments of the present application, in order to further enhance the data security monitoring capability, the following steps can also be implemented:

[0118] Step 5.1, real-time monitoring of data access behavior and updating risk score;

[0119] Based on the user's real-time access log stream (including operation type, access frequency, data volume, etc.), the behavior monitoring rule set generated in step 4.1 and the historical behavior baseline model, the access mode is analyzed in real time through a sliding window algorithm, the abnormal detection algorithm is used to identify the behavior deviating from the normal mode, and the user trust degree is dynamically adjusted.

[0120] Before applying the risk score algorithm, the system first performs dimensionless processing on the multi-dimensional abnormal behavior indicators: different types of abnormal indicators (such as access frequency abnormal score, data volume abnormal score, operation type abnormal score, etc.) are converted to a standard distribution with a mean of 0 and a variance of 1 through standardization processing; time decay weight is applied to time-related abnormal signals to ensure that recent abnormal behaviors have higher influence weight; through principal component analysis (PCA, Principal Components Analysis), the multi-dimensional abnormal features are mapped to a unified risk space. These preprocessing steps ensure that in the risk score function , different dimensions and different types of abnormal indicators can be effectively combined to generate a consistent risk score; wherein, represents the multi-dimensional abnormal behavior feature vector at the current time, reflecting the user's performance in various abnormal behaviors; is the historical risk score, representing the user's past risk state; is the weight vector of each abnormal behavior type, used to adjust the influence degree of different abnormal types on the total risk score; is the time decay factor, used to control the speed of the influence of historical abnormal behaviors on the current risk score decreasing over time.

[0121] It should be noted that the sliding window algorithm used in the embodiment receives a real-time access log stream as input and outputs behavior statistical features within a specific time window, providing basic data for anomaly detection.

[0122] In addition, the embodiment uses the IsolationForest algorithm for anomaly behavior detection. The IsolationForest algorithm receives current behavior features and baseline models as input and outputs behavior anomaly scores for determining whether the current operation deviates from the normal mode.

[0123] The user risk score dynamic updating algorithm in the embodiment is an innovative algorithm that fuses multiple source anomaly signals and considers the time decay characteristics. The user risk score dynamic updating algorithm assigns different weights to different types of abnormal behaviors, while introducing an exponential decay function to handle the timeliness of historical abnormal records, achieving accurate quantification and dynamic adjustment of user risk status. The user risk score dynamic updating algorithm receives anomaly detection results and historical risk scores as input and outputs updated user risk scores and corresponding access permission adjustment suggestions. The scoring function in the user risk score dynamic updating algorithm is defined as follows: the current abnormal behavior is combined with the historical risk score, and the updated risk score is calculated based on the weight vector of different types of abnormal behaviors and the time decay factor, returning a real value between 0 and 1, representing the current risk level of the user.

[0124] The specific calculation formula of the risk score function is as follows:

[0125] ;

[0126] wherein represents the risk score function, represents the summation symbol, is the historical risk score, is the historical score weight factor (value is 0.7), is the weight coefficient of the abnormal behavior type, is the abnormal degree of the abnormal behavior type, is the time decay factor (value is 0.1 / hour), is the time elapsed since the last scoring (hours).

[0127] ​​​​​​​​In this embodiment, abnormal behavior types include: abnormal access frequency (weight 0.3), abnormal data volume (weight 0.25), abnormal operation type (weight 0.25), and abnormal access time (weight 0.2). The risk level thresholds are set as: low risk [0, 0.3), medium risk [0.3, 0.7), and high risk [0.7, 1], with different access restriction policies corresponding to different risk levels.

[0128] The time dimension processing in the risk scoring function employs a multi-timescale ensemble approach, integrating short-term, medium-term, and long-term risk assessments. The complete time-sensitive risk scoring function is defined as follows:

[0129] ;

[0130] in, , , These represent short-term (within 1 hour), medium-term (within 24 hours), and long-term (within 7 days) risk scores, respectively. , , These represent the weighting coefficients for short-term, medium-term, and long-term risk scores, respectively. The overall risk score after integration.

[0131] Risk scores at all time scales are calculated using the same basic formula, but with different parameter settings: short-term assessments use high time precision ( The time accuracy is 0.5 / hour, used to capture sudden abnormal behavior; the mid-term evaluation uses medium time accuracy ( The time precision is 0.1 / hour, used to identify persistent anomalous patterns; long-term assessment uses low time precision ( The score is 0.01 / hour, used to monitor slowly evolving abnormal trends. The system updates the short-term score every 10 minutes, the medium-term score every hour, and the long-term score every day, thus ensuring that the risk assessment can reflect changes in user behavior in a timely manner at different time scales.

[0132] like Figure 5 As shown, the relationship between different types of abnormal behavior and risk scores is illustrated. Corresponding to the risk score dynamic update algorithm in step 5.1, the influence weight of different abnormal behavior types on user risk scores is intuitively demonstrated, verifying the actual effect of the risk scoring function in the patent.

[0133] Step 6: Combine the reconstructed anonymous identifier with the multi-party authorization revocation protocol, and reconstruct the identity mapping relationship after obtaining authorization to achieve accountability for data leakage.

[0134] In this step, first, the reconstructed anonymous identifier in step 5, the mapping relationship fragments held by each authorized party in step 4, and the revocation key share, as well as the revocation request and legal authorization documents, are obtained. Then, the legality of the revocation request is confirmed by the threshold signature verification algorithm, the complete mapping relationship is recovered from the fragments using the secret reconstruction algorithm, and the entire revocation process is recorded by the audit log generation algorithm. Finally, the decrypted real identity information, the revocation operation audit report, and the evidence chain integrity proof document are output.

[0135] It should be noted that the secret reconstruction algorithm in the present embodiment is based on the principle of Lagrange interpolation. The secret reconstruction algorithm receives a sufficient number of secret shares as input, reconstructs the original polynomial through interpolation calculation, and outputs the recovered anonymous identifier and real identity mapping relationship.

[0136] The threshold signature verification and audit log generation mechanism in the present embodiment is an innovative combination that ensures the legality and traceability of the revocation process. This threshold signature verification and audit log generation mechanism not only requires the signature approval of a predetermined threshold of authorized parties, but also links each step of the entire revocation process into an unalterable audit trail through hash chain technology, ensuring the legality of the operation and the transparency of the process. The threshold signature verification and audit log generation mechanism receives the revocation request and authorized signature as input, and outputs the verification result and complete operation audit chain.

[0137] Threshold value in threshold signature verification mechanism Set to two-thirds of the total number of authorized parties and round up the result. That is, the system will calculate how many valid signatures of authorized parties (i.e., threshold value ) are needed to perform the revocation operation according to the actual number of authorized parties . For example, in the present embodiment, the total number of authorized parties is , and the threshold value is , that is, at least valid signatures of authorized parties are needed. The signature verification function is defined as: verifying the validity of the signature set for the request . Specifically, first, count the number of valid signatures in the signature set that pass mathematical verification, denoted as . When and only when and all valid signatures pass the verification, the function returns True, otherwise False.

[0138] The audit log hash chain uses the SHA-256 algorithm, and the hash value of each log entry is calculated as:

[0139] ;

[0140] wherein is the hash value of the th entry, is the current timestamp, is the operation description, is the digital signature of the operation executor, denotes the string concatenation operation; denotes the secure hash algorithm SHA-256.

[0141] In the embodiments of the present application, in order to form a complete security closed loop, the following steps can also be implemented:

[0142] Step 6.1, generating accountability evidence chain and updating protection strategy;

[0143] Based on the watermark extraction result, the identity mapping reconstruction result, the access behavior log and the leakage event context information, a complete leakage propagation path is constructed through an evidence correlation analysis algorithm, a responsible party is determined based on multi-source evidence using a responsibility identification algorithm, and the security rules are updated according to the leakage mode.

[0144] Before performing the evidence correlation analysis, the system uniformly preprocesses the multi-source heterogeneous evidence data: the confidence degree of the watermark matching result is standardized and converted into a matching confidence score in the interval [0, 1]; the time and space characteristics are uniformly coded, the time information is converted into a relative timestamp, and the space information is converted into a standardized position code; the behavior characteristics are vectorized, and different types of behavior modes are mapped to a feature vector space. These preprocessing steps ensure that evidence data of different sources and different dimensions can be correlated in a unified mathematical framework.

[0145] In addition, the output of the evidence correlation analysis algorithm is decoded and transformed to generate an actionable accountability result: the abstract evidence chain graph structure is converted into a time-sequential visualized leakage propagation path; a confidence score and an explainability analysis report are generated for the responsibility identification result; and the protection strategy parameters are converted into specific security configuration instructions, including watermark parameter adjustment, access control rule modification and monitoring threshold update, etc. executable operation instructions.

[0146] The evidence correlation analysis algorithm in the present embodiment is an innovative algorithm for multi-source heterogeneous data fusion, which identifies potential leakage paths and key nodes by establishing a spatio-temporal correlation model of leakage data characteristics and access behavior. The evidence correlation analysis algorithm receives leakage sample characteristics and historical access logs as input, and outputs a visualized leakage propagation chain and a key node list. Among them, the correlation function in the evidence correlation analysis algorithm is defined as: the watermark evidence with access logs in a time window , by calculating the spatiotemporal consistency score and the behavior feature similarity, constructing the evidence chain graph structure, and outputting the possible path set of the leakage propagation.

[0147] Time window in evidence correlation analysis algorithm is set as an adjustable parameter, with a basic unit of day, and in this embodiment, the value range is days, and the default value is days. The calculation formula of the spatiotemporal consistency score is:

[0148] ;

[0149] Among them, represents the spatiotemporal consistency score, is the time similarity, is the location similarity, is the time weight factor (the value is 0.6).

[0150] The calculation of the behavior feature similarity adopts the cosine similarity method. Specifically, the system will represent the watermark feature and the access behavior feature as vectors, and then calculate the cosine value between the two vectors to measure their similarity. The closer the cosine value is to 1, the more similar the behavior features of the two are; on the contrary, the smaller the value is, the lower the similarity is.

[0151] Correlation threshold is set as . Among them, represents the behavior feature similarity score, which specifically refers to the similarity between the watermark feature vector and the access behavior feature vector calculated by the cosine similarity method, and the value range is , and the closer the value is to 1, the more similar the behavior features are. Finally, when , it is determined that there is a correlation.

[0152] The evidence chain construction algorithm adopts the maximum likelihood estimation method to calculate the probability score of different propagation paths and outputs the path set whose probability exceeds the threshold (set as ).

[0153] Time similarity in evidence correlation analysis algorithm The complete calculation formula is:

[0154] ;

[0155] Among them, is the time similarity,​ for the time of the leak event, for the time of the suspicious behavior, for the reference time point (usually the time of the leak discovery), for the time importance function;

[0156] The time importance function is used to measure the importance of a certain event time point relative to the reference time point . Specifically, as the distance increases, the importance decreases exponentially. The speed of the time importance decay is controlled by the coefficient (2.0 in this embodiment), and the maximum time range under consideration is 365 days. That is, the closer the event is to the reference time point, the higher its weight in evidence analysis, and the farther the event is from the reference time point, the lower its weight.

[0157] The time importance function ensures that the time similarity evaluation takes into account both the temporal proximity between events and the importance of events relative to the reference time point, so that recent events obtain a higher weight in evidence correlation. The calculation of the location similarity is based on a weighted combination of network topology and geographic location, and the network topology similarity and geographic location similarity are combined.

[0158] The responsibility determination algorithm in this embodiment adopts an evidence weight fusion model to comprehensively evaluate multi-dimensional factors such as watermark matching degree, behavior abnormality degree, and spatio-temporal consistency, and generate a credibility score of responsibility determination. The responsibility determination algorithm receives multiple evidence source data as input and outputs the responsibility determination result and the contribution weight of each evidence.

[0159] The credibility score in the responsibility determination algorithm is calculated by the formula:

[0160] ;

[0161] wherein represents the credibility score, is the weight factor of the evidence type , is the strength score of the evidence type , and represents the summation symbol.

[0162] In this embodiment, the evidence types include: watermark matching evidence (weight ), behavior abnormality evidence (weight ), and spatio-temporal consistency evidence (weight ). The responsibility determination threshold is set to​​ When , it is identified as a high-confidence liability determination; when , it is identified as a medium-confidence determination; when , it is identified as a low-confidence determination, requiring further investigation.

[0163] The confidence score calculation of the liability determination algorithm introduces an evidence strength decay factor with time dimension. The complete scoring formula is:

[0164] ;

[0165] Where: is the liability determination confidence score at time ; is the weight factor of evidence type ; is the strength score of evidence type ; is the time decay function of evidence type , reflecting the effectiveness change of evidence after days since collection; denotes the summation symbol.

[0166] The meaning of the time decay function is: as time goes by since the evidence was collected, the effectiveness of the evidence will gradually decrease. Specifically, each type of evidence has a corresponding time decay rate , while represents the number of days that have passed since the evidence was collected. The time decay function dynamically adjusts the weight of the evidence according to the values of and , so that the more distant in time the evidence is, the less influence it has on the liability determination, reflecting the actual situation that evidence gradually loses reliability over time.

[0167] The time decay rates of different types of evidence reflect the speed at which evidence loses reliability over time: watermark matching evidence (low decay rate, watermark evidence is relatively stable), behavioral anomaly evidence (high decay rate, behavior patterns may change rapidly over time), and spatiotemporal consistency evidence (medium decay rate). This formula ensures that in the process of liability determination, newly collected evidence has a higher weight than old evidence, and the effective period of different types of evidence matches their inherent reliability characteristics. The liability determination result is automatically updated every 24 hours, reflecting the impact of new evidence and time factors.

[0168] The protection strategy adaptive optimization algorithm in this embodiment can extract feature patterns from the occurred leakage events, identify system vulnerabilities and protection blind spots, and adjust the watermark parameters, access control rules and monitoring thresholds accordingly. The protection strategy adaptive optimization algorithm receives leakage case data and current protection configuration as input, and outputs an optimized set of protection strategy parameters.

[0169] The protection strategy adaptive optimization algorithm fully considers the time factor in design, and the optimization goal is to enable the system to select the optimal protection action according to the current state at different time points, so as to obtain the maximum comprehensive benefit in the long term. The algorithm maps the state of the system to the corresponding protection measures through a strategy function, and introduces a discount factor in the decision-making process, so that the importance of future benefits gradually decreases. The evaluation of comprehensive benefits not only considers the improvement of security, but also takes into account the influence of system performance and the change of user experience. Specifically, the algorithm calculates the security improvement, performance change and user experience change respectively, and assigns different weights to the three, among which the security weight is the highest, followed by performance and user experience, and the weight ratio is 0.6, 0.2 and 0.2 respectively.

[0170] The probability of change of system state is modeled based on historical operation data and expert experience, and time series analysis method is used to capture the time dependence of state transition. In order to ensure the timeliness of the model, the state transition probability matrix will be updated once a quarter.

[0171] In addition, the algorithm also sets time-related constraints to ensure the stability and security of the system. For example, the update frequency of the protection strategy is limited to at most once a week to avoid system instability caused by frequent adjustments; the amplitude of each strategy adjustment is also limited within plus or minus 30% to prevent excessive changes affecting system operation. In high-risk periods (such as night and weekend), the system will automatically adopt more stringent security strategies, and the corresponding parameter adjustment amplitude will be increased to 1.5 times of that in normal period to enhance the protection capability.

[0172] As shown in Figure 6 The complete process from data access to responsibility tracing is shown, the evidence association analysis and responsibility identification algorithm corresponding to step 6.1 is shown, the complete responsibility tracing path after scientific research data leakage is shown, and the closed-loop accountability effect of the multi-model collaborative protection mechanism in actual application is verified.

[0173] A computer storage medium, comprising a memory and one or more processors, the memory storing executable code, the one or more processors executing the executable code to implement the above-mentioned scientific research data security protection method based on multi-model collaboration.

[0174] Herein, the present application provides an implementation example:

[0175] A national medical imaging research center has established a multi-modal medical database containing brain MRI, CT, and pathological images that need to be shared among medical research institutions nationwide for the development of artificial intelligence-assisted diagnosis systems. This database contains sensitive medical data for 500,000 patients, and both patient privacy and data security and traceability must be ensured. Each data set consists of clinical description text, multi-modal medical images, patient physiological indicator values, and some cases of doctor voice recordings, with a total data volume exceeding 200 TB.

[0176] Researcher A of a medical research institution applies to access the brain glioma data set, submitting an application containing his institution's qualifications, research project number, data security level permission, and other identity attributes. After the system collects these attributes, it does not store the original information but verifies the researcher's access qualifications through zero-knowledge proof technology and generates an anonymous identifier. An example of user identity attributes and anonymous identifier generation is shown in Table 1:

[0177] Table 1: User identity attributes and anonymous identifier generation example

[0178]

[0179] The system analyzes the structure of the brain glioma data set that researcher A requests to access, which includes clinical description text, MRI / CT images, patient physiological indicator values, and doctor voice recordings. The system calculates the watermark capacity based on the characteristics of each modality and develops an optimal watermark allocation strategy. The results of multi-modal medical data feature analysis are shown in Table 2:

[0180] Table 2: Results of multi-modal medical data feature analysis

[0181]

[0182] The results of the watermark fragment allocation strategy are shown in Table 3:

[0183] Table 3: Results of watermark fragment allocation strategy

[0184]

[0185] The system extends researcher A's anonymous identifier (256-bit hash value) to 320 bits using Reed-Solomon encoding and embeds it into each modality according to the above strategy. Different watermark algorithms are used for each modality to ensure data usability.

[0186] Six months later, a suspected brain glioma data set from the national medical imaging research center appeared on a medical forum. After obtaining the sample, the security department initiated the accountability process and used multi-modal watermark extraction technology to recover the anonymous identifier. The results of the leaked data watermark extraction are shown in Table 4:

[0187] Table 4: Leakage data watermark extraction results

[0188]

[0189] After the revocation authorization application is initiated, the system notifies 5 authorized parties to execute the identity mapping revocation agreement after obtaining permission through legal procedures. Finally, 4 authorized parties provide valid key shares, reaching the threshold value requirement (minimum 3 shares), and successfully reconstructing the identity mapping relationship. The multi-party authorized revocation process record is shown in Table 5:

[0190] Table 5: Multi-party authorized revocation process record

[0191]

[0192] Through multi-party authorized revocation, the system finally confirms that researcher A is responsible for the leaked data. According to the investigation results, the researcher violated the data use agreement and provided data to unauthorized third parties. The system updates the security policy accordingly, enhances the behavior monitoring rules for this type of research project, and optimizes the watermark allocation strategy to improve tracking capabilities. This case demonstrates that the multi-model collaborative scientific data security protection method successfully protects medical data privacy while achieving data leakage accountability.

[0193] The above describes embodiments of the present application, but the embodiments are not limited to the specific implementation described above, which is only illustrative and not limiting. Those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection of the embodiments.

Claims

1. A scientific research data security protection method based on multi-model cooperation, characterized in that, The method comprises the following steps: Obtain user identity attribute declaration and access request parameters, generate anonymous identifier and attribute proof through zero-knowledge proof technology; Based on the anonymous identifier and the attribute proof, analyze the multi-modal composition structure of the scientific research data, and generate the corresponding watermark embedding strategy accordingly; According to the watermark embedding strategy, the coded anonymous identifier is scattered and embedded into different modalities of the scientific research data, realizing multi-modal collaborative watermark embedding; After completing the watermark embedding, use the anonymous identifier to build a revocable mapping mechanism between the anonymous identifier and the real identity, and store the mapping relationship in fragments; When a data leakage event is detected, extract watermark information from the leaked data to reconstruct the anonymous identifier; Combine the reconstructed anonymous identifier to execute a multi-party authorized revocation protocol, and after obtaining authorization, reconstruct the identity mapping relationship to realize the accountability of the data leakage behavior.

2. The scientific research data security protection method based on multi-model cooperation according to claim 1, characterized in that, The step of obtaining user identity attribute declaration and access request parameters, generating anonymous identifier and attribute proof through zero-knowledge proof technology, specifically includes: preprocessing user identity attribute data, including normalizing numerical attributes, encoding categorical attributes, and converting time attributes to relative time stamps; generating verifiable attribute proof without revealing specific attribute values through zero-knowledge proof generation algorithm; generating a unique anonymous identifier through hash chain algorithm.

3. The scientific research data security protection method based on multi-model cooperation according to claim 1, characterized in that, The step of analyzing the multi-modal composition structure of the scientific research data and generating the corresponding watermark embedding strategy accordingly, specifically includes: preprocessing each type of data to achieve consistency in dimensionality for different modalities of data; identifying feature regions of each modality using multi-modal content analysis algorithm; calculating the amount of information that can be embedded in each modality using adaptive multi-modal watermark capacity evaluation algorithm; and encoding and fragmenting identity information using Reed Solomon error correction encoding algorithm.

4. The scientific research data security protection method based on multi-model cooperation of claim 1, wherein, In the step of multi-modal collaborative watermark embedding, the following watermark algorithms are used for different modalities: semantic preservation watermark algorithm for text modality; DCT domain watermark algorithm for image modality; spread spectrum watermark algorithm for audio modality; and least significant bit replacement algorithm for numerical data.

5. The scientific research data security protection method based on multi-model cooperation according to claim 1, characterized in that, The step of building a revocable mapping mechanism between the anonymous identifier and the real identity, specifically includes: Split the identity mapping relationship through Shamir secret sharing algorithm; Use threshold encryption algorithm to ensure that the complete mapping relationship can only be reconstructed when the condition is met, where the condition means that only when no less than a preset threshold number of authorized nodes jointly submit their identity mapping relationship fragments, the complete identity mapping relationship can be reconstructed through threshold decryption algorithm; Distribute and store the identity mapping relationship fragments in different authorized nodes.

6. The scientific research data security protection method based on multi-model cooperation according to claim 1, characterized in that, The step of reconstructing the anonymous identifier based on the leaked data to extract watermark information, specifically includes: using corresponding watermark extraction algorithms to extract watermark information from each modality in parallel; restoring damaged identity information fragments through error correction decoding algorithm; and using multi-modal cross-validation mechanism to integrate watermark information extracted from each modality, so that even if some modalities have severely damaged and missing watermarks, the anonymous identifier can still be reconstructed.

7. The scientific research data security protection method based on multi-model cooperation according to claim 1, characterized in that, The step of executing the multi-party authorized revocation protocol specifically comprises: confirming the legality of the revocation request through a threshold signature verification algorithm; recovering the complete mapping relationship from the fragments using a secret reconstruction algorithm based on the Lagrange interpolation principle; and recording the entire revocation process through an audit log generation algorithm.

8. The scientific research data security protection method based on multi-model cooperation of claim 1, wherein, The step of extracting watermark information from the leaked data to reconstruct the anonymous identifier further comprises the step of monitoring data access behavior in real time and updating the risk score, specifically comprising: analyzing access patterns in real time based on real-time access log streams, behavior monitoring rules, and historical behavior baseline models through a sliding window algorithm; performing abnormal behavior detection using an IsolationForest algorithm; and dynamically updating the user risk score through a user risk score dynamic updating algorithm combined with historical scores, the weight of abnormal behavior types, the degree of abnormality, and a time decay factor to dynamically adjust the user risk score.

9. The scientific research data security protection method based on multi-model cooperation of claim 1, wherein, The step of executing the multi-party authorized revocation protocol and reconstructing the identity mapping relationship after obtaining authorization further comprises the step of generating an accountability evidence chain and updating the protection strategy to form a complete security closed loop, specifically comprising: constructing a complete leak propagation path through an evidence correlation analysis algorithm based on watermark extraction results, identity mapping reconstruction results, access behavior logs, and leak event context information; determining the responsible party based on multi-source evidence using a responsibility determination algorithm; and updating security rules according to the leak mode through a protection strategy adaptive optimization algorithm.

10. A computer storage medium, characterized in that, The device comprises a memory and one or more processors, the memory stores executable code, and the one or more processors execute the executable code to implement the multi-model cooperation-based scientific research data security protection method of any one of claims 1-9.

Images