Scientific research data security protection method based on multi-model cooperation

Abstract

The invention relates to the technical field of scientific research data security protection, and discloses a multi-model cooperation-based scientific research data security protection method, which comprises the following steps of: obtaining a user identity attribute statement and an access request parameter, and generating an anonymous identifier and an attribute proof through a zero-knowledge proof technology; analyzing a multi-mode composition structure of scientific research data; dispersing and embedding the coded anonymous identifiers into different modes of scientific research data; constructing a revocable mapping mechanism between the anonymous identifier and the real identity, and performing fragmentation storage on the mapping relation; when a data leakage event is detected, extracting watermark information based on leaked data to reconstruct an anonymous identifier; a multi-party authorization revocation protocol is executed, and the identity mapping relation is reconstructed after authorization is obtained; according to the method, zero knowledge proves that the multi-modal watermark and the revocable anonymous architecture are innovatively combined, so that the privacy of scientific research data accessors is protected, and meanwhile, effective balance between data tracking and leakage responsibility investigation is realized.

Description

Technical Field

[0001] This invention relates to the field of scientific research data security protection technology, and more specifically, to a scientific research data security protection method based on multi-model collaboration. Background Technology

[0002] Several key technical challenges exist in the process of sharing scientific research data: on the one hand, protecting the privacy of data access users requires allowing anonymity or minimizing the disclosure of identity information; on the other hand, preventing data leaks necessitates identity-based data distribution for traceability and accountability. Simultaneously, watermarking technology must possess both unforgeability and robustness, resisting various attacks, and supporting cross-modal data integrity protection. Existing technologies struggle to balance these seemingly contradictory needs, often sacrificing privacy for accountability or protecting privacy while weakening accountability. Specifically, existing scientific research data sharing technologies suffer from three core contradictions: The contradiction between identity verification and privacy protection: verifying a user's access rights usually requires collecting complete identity information, leading to excessive exposure of privacy; the contradiction between anonymous access and accountability: allowing anonymous access protects privacy but makes it difficult to hold people accountable in case of leakage, while recording identity loses privacy protection; the contradiction between watermark robustness and data usability: single-modal watermarks are easily destroyed, while multi-modal watermarks may affect data usability.

[0003] Furthermore, in existing technologies, identity authentication systems, watermark embedding systems, and accountability mechanisms are independent of each other, lacking a unified architectural design, and thus failing to establish a data sharing mechanism that both respects privacy and effectively enables accountability. Summary of the Invention

[0004] This invention provides a scientific research data security protection method based on multi-model collaboration, which solves the contradiction between identity verification and privacy protection, anonymous access and accountability, watermark robustness and data availability in related technologies, as well as the technical problem that identity authentication system, watermark embedding system and accountability mechanism are independent and lack a unified architecture design.

[0005] This invention provides a scientific data security protection method based on multi-model collaboration, comprising: Obtain user identity attribute declarations and access request parameters, and generate anonymous identifiers and attribute proofs using zero-knowledge proof technology; Based on anonymous identifiers and attribute proofs, we analyze the multimodal composition structure of scientific research data and generate corresponding watermark embedding strategies accordingly. Based on the watermark embedding strategy, the encoded anonymous identifier is embedded into different modalities of scientific research data to achieve multimodal collaborative watermark embedding. After completing the watermark embedding, an anonymous identifier is used to construct a revocable mapping mechanism between anonymous identifiers and real identities, and the mapping relationship is stored in segments. Upon detecting a data breach, the watermark information is extracted from the breached data to reconstruct the anonymous identifier; By combining the reconstructed anonymous identifiers and executing a multi-party authorization revocation protocol, the identity mapping relationship is reconstructed after authorization is obtained, thereby enabling accountability for data breaches.

[0006] Furthermore, the steps of obtaining user identity attribute declarations and access request parameters, and generating anonymous identifiers and attribute proofs using zero-knowledge proof technology, specifically include: preprocessing user identity attribute data, including normalizing numerical attributes, encoding categorical attributes, and converting time-related attributes to relative timestamps; generating verifiable attribute proofs using a zero-knowledge proof generation algorithm without disclosing specific attribute values; and generating unique anonymous identifiers using a hash chain algorithm.

[0007] Furthermore, the step of analyzing the multimodal composition structure of scientific research data and generating corresponding watermark embedding strategies specifically includes: performing modality-specific preprocessing on each type of data to achieve consistency in the dimensions of different modal data; identifying the feature regions of each modality using a multimodal content analysis algorithm; calculating the amount of information that can be embedded in each modality using an adaptive multimodal watermark capacity evaluation algorithm; and encoding and fragmenting the identity information using the ReedSolomon error correction coding algorithm.

[0008] Furthermore, in the multimodal collaborative watermarking embedding step, the following watermarking algorithms are used for different modalities: for text modality, a semantically preserved watermarking algorithm is used; for image modality, a DCT domain watermarking algorithm is used; for audio modality, a spread spectrum watermarking algorithm is used; and for numerical data, a least significant bit replacement algorithm is used.

[0009] Furthermore, the step of constructing the revocable mapping mechanism between anonymous identifiers and real identities specifically includes: The identity mapping relationship is segmented using the Shamir secret sharing algorithm; Threshold encryption algorithms are used to ensure that the complete mapping relationship can only be reconstructed when certain conditions are met. The specific conditions are that the complete identity mapping relationship can only be reconstructed through the threshold decryption algorithm when no less than a preset threshold number of authorized nodes jointly submit the identity mapping relationship fragments they hold. The identity mapping relationships are distributed and stored on different authorized nodes.

[0010] Furthermore, the step of reconstructing the anonymous identifier based on watermark information extracted from leaked data 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 algorithms; and using a multimodal cross-validation mechanism to integrate the watermark information extracted from each modality, so as to reconstruct the anonymous identifier even if some modal watermarks are severely damaged or missing.

[0011] Furthermore, the steps of executing the multi-party authorization revocation protocol specifically include: confirming the legality of the revocation request through a threshold signature verification algorithm; restoring 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.

[0012] Furthermore, in the step of extracting watermark information and reconstructing anonymous identifiers based on leaked data, to further enhance data security monitoring capabilities, it also includes a step of real-time monitoring of data access behavior and updating risk scores. Specifically, this includes: analyzing access patterns in real time using a sliding window algorithm based on the user's real-time access log stream, behavior monitoring rule set, and historical behavior baseline model; using the IsolationForest algorithm to detect abnormal behavior; and dynamically adjusting the user risk score by combining historical scores, the weight of abnormal behavior types, the degree of abnormality, and the time decay factor through a user risk score dynamic update algorithm.

[0013] Furthermore, in the step of executing the multi-party authorization revocation agreement and reconstructing the identity mapping relationship after obtaining authorization, in order to form a complete security closed loop, it also includes the steps of generating an accountability evidence chain and updating the protection strategy. Specifically, this includes: constructing a complete leakage propagation path based on the watermark extraction results, identity mapping reconstruction results, access behavior logs, and leakage event context information through an evidence association analysis algorithm; determining the responsible party based on multi-source evidence using a responsibility determination algorithm; and updating security rules according to the leakage pattern through a protection strategy adaptive optimization algorithm.

[0014] This invention provides a computer storage medium, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned method for protecting scientific research data based on multi-model collaboration.

[0015] The beneficial effects of this invention are as follows: In terms of the contradiction between identity verification and privacy protection, this invention adopts attribute verification technology based on zero-knowledge proof, which enables users to prove that they meet the access qualification conditions without disclosing specific identity information. This "verification but not disclosure" mechanism effectively avoids the problem of excessive privacy exposure caused by the collection of complete identity information in traditional methods, and achieves the coordination and unity of identity verification and privacy protection. Regarding the conflict between anonymous access and accountability, this invention constructs a revocable anonymity watermarking architecture. By using Shamir secret sharing and threshold encryption technology to protect and separate the mapping relationship between anonymous identifiers and real identities, it ensures that users can access data anonymously under normal circumstances, and that anonymity can only be revoked and accountability pursued only after a leak occurs and multiple parties have authorized it. This mechanism separates identity information management from accountability authority, forming an effective check and balance mechanism and achieving a dynamic balance between anonymity protection and accountability. Regarding the trade-off between watermark robustness and data availability, this invention employs a multimodal collaborative watermarking mechanism. User identity information is encoded using Reed-Solomon and then embedded into different data modalities. The most suitable watermarking algorithm is used for each modality (such as semantically preserving watermarking for text, DCT domain watermarking for images, spread spectrum watermarking for audio, and least significant bit replacement for numerical data). This achieves a balance between high watermark robustness and high data availability. Even if the watermarks in some modalities are destroyed, the system can still recover the identity information from other modalities through cross-validation, thus improving the reliability of evidence for accountability. This invention also effectively solves the problem that identity authentication systems, watermark embedding systems, and accountability mechanisms are independent of each other in the prior art. It constructs a complete security closed loop from access verification, data distribution, behavior monitoring to leakage accountability, so that each link forms an organic and unified protection system. The anonymous identifier generated by zero-knowledge proof is directly used for watermark embedding, while 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 forms a unified architecture design that protects privacy and effectively holds people accountable. Attached Figure Description

[0016] Figure 1 This is a flowchart of a scientific research data security protection method based on multi-model collaboration in this invention; Figure 2 It is a bar chart comparing the watermark capacity and robustness coefficient of different data modalities; Figure 3 This is a diagram showing the mapping relationship between anonymous identifiers and real identities, as well as the relationship between multi-party authorization mechanisms. Figure 4 It is a line graph showing the changes in watermark extraction success rate when different modalities of data face attacks of various intensities; Figure 5 It is a scatter plot showing the relationship between different types of abnormal behavior and risk scores; Figure 6 This is a Sankey diagram illustrating the complete process from data access to accountability. Detailed Implementation

[0017] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0018] At least one embodiment of the present invention discloses a method for protecting scientific research data based on multi-model collaboration, such as... Figure 1 As shown, it includes: Step 1: Obtain the user identity attribute declaration and access request parameters, and generate an anonymous identifier and attribute proof using zero-knowledge proof technology; This step first obtains the user-submitted identity attribute declaration (in JSON format), access request parameters (including the required permission level), and attribute credentials (in digital signature format) issued by a trusted authority. Then, a zero-knowledge proof generation algorithm processes the input data to generate a verifiable attribute proof without revealing the specific attribute values. Simultaneously, a unique anonymous identifier is generated using a 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.

[0019] Before processing the input data, the system preprocesses the user identity attribute data, including: normalizing numerical attributes by mapping different ranges of values ​​to the [0, 1] interval; encoding categorical attributes by using one-hot encoding to convert non-numerical data into numerical vectors; and converting time-related attributes into relative timestamps. These preprocessing steps ensure that attribute data of different types and scales can be processed uniformly within the zero-knowledge proof system.

[0020] It should be noted that the zero-knowledge proof generation algorithm described above is implemented based on the Schnorr protocol. This algorithm takes the user's identity attributes and related credentials as input and outputs a zero-knowledge proof certificate, enabling the verifier to confirm that the user possesses the required attributes without needing to know the specific attribute values.

[0021] In this implementation, the main parameters of the Schnorr protocol are set as follows: a 2048-bit prime number is used as the main parameter, and a 256-bit prime number is used as the sub-parameter. The sub-parameter is divisible by the main parameter minus one, and the order of the generator is equal to that of the sub-parameter. The security parameter (challenge value) in the protocol is limited to a large numerical range, and the random number (commitment value) also has a defined value range. During the proof generation process, the system sets a verification threshold; only when the verification score is higher than 0.95 is the user's attribute considered to meet the access conditions.

[0022] The multi-round hash chaining algorithm in this embodiment is an improvement on the conventional hash chaining algorithm. By introducing a multi-round iteration combined with a timestamp mechanism, it enhances the security of anonymous identifiers, making the identifier both unique and irreversibly deducible from the user's original identity. This multi-round hash chaining algorithm receives user attribute information, a timestamp, and a random salt value as input, and outputs a 256-bit hash value for the anonymous identifier.

[0023] In this implementation, the multi-round hash chain algorithm iterates 10,000 times. Each iteration uses the SHA-256 hash function. The timestamp used to generate the hash chain has millisecond precision, and the random salt value is 128 bits long. Specifically, the hash value of the user attribute is used as the initial input. In each iteration, the previous hash result is concatenated with the current timestamp and random salt value, and then the concatenated result is hashed using SHA-256. After 10,000 iterations, the final hash value is the anonymous identifier. Throughout the process, string concatenation is used to combine the hash value, timestamp, and salt value into a single input, ensuring the uniqueness and unpredictability of the input in each round.

[0024] Step 2: Based on anonymous identifiers and attribute proofs, analyze the multimodal composition structure of scientific research data and generate corresponding watermark embedding strategies accordingly. In this step, the research data to be shared (including multiple modalities such as text, images, audio, and numerical data), data structure metadata (JSON format describing the distribution of each modality), and the anonymous identifier generated in step 1 are first obtained. Then, a multimodal content analysis algorithm is used to identify the feature regions of each modality, a watermark capacity assessment algorithm is used to calculate the amount of information that can be embedded in each modality, and an error correction coding algorithm is used to encode and segment the identity information. Finally, a multimodal feature map (identifying the embeddable regions of each modality), a set of encoded identity information segments (each segment contains partial identity information and a check code), and a watermark embedding strategy table (specifying the segment index for embedding each modality) are output.

[0025] Before performing multimodal content analysis, the system first performs modality-specific preprocessing on each type of data: text data undergoes inverse document frequency (IVF) processing to convert the text into numerical feature vectors; image data is normalized to scale pixel values ​​to the [0, 1] interval; audio signals undergo frequency domain transformation and energy normalization; and numerical data is standardized to convert it into a distribution with a mean of 0 and a variance of 1. These preprocessing operations ensure the consistency of dimensions between different modal data, enabling the multimodal capacity assessment algorithm to accurately calculate and compare the watermark carrying capacity of different modalities.

[0026] It should be noted that the error correction coding algorithm described above is implemented using Reed-Solomon coding. This Reed-Solomon coding algorithm takes an anonymous identifier (256-bit hash value) as input and outputs a fault-tolerant sequence of coded symbols and its fragments, enabling the system to recover the original information even if some data is lost or tampered with.

[0027] Feature Region Identification Function in Multimodal Content Analysis Algorithms Defined as a data mode In feature space The system extracts regions suitable for watermarking. The feature region recognition function, based on information entropy and perceived importance score, returns a set of coordinates of regions in each modality that are suitable for embedding watermarks without affecting data availability.

[0028] Feature region recognition function The information entropy calculation uses the Shannon entropy method. Specifically, it involves calculating the probability of each feature value appearing in the data, multiplying all probabilities by their base-2 logarithm, and finally taking the negative sum of these products to obtain the entropy value of the region. Information entropy reflects the degree of disorder in the feature distribution; the more uniform the probability distribution, the higher the entropy value, indicating that the region is more suitable for watermark embedding. Feature region selection threshold. Set as ,when At that time, the corresponding region is identified as a feature region suitable for watermark embedding. Perceived importance scoring. Defined as normalized The value within the interval is calculated using the following formula: ; in Indicates perceived importance rating, For information entropy weighting coefficients, The marginal complexity weight coefficient, This is the content significance weighting coefficient. For feature regions Information entropy For feature regions edge complexity, For feature regions Content salience; in this embodiment , , .

[0029] The adaptive multimodal watermarking capacity assessment algorithm in this embodiment is an innovative algorithm designed for the characteristics of different data modalities. This algorithm dynamically calculates the amount of watermark information that each modality can carry based on the inherent characteristics of each modality (such as semantic redundancy of text, complexity of images, spectral characteristics of audio, and precision of numerical data), and optimizes the watermark fragment allocation strategy based on the calculation results, thereby maximizing the robustness of the watermark while maintaining data availability. This adaptive multimodal watermarking capacity assessment algorithm receives data characteristic parameters of various modalities as input and outputs the watermark capacity value for each modality and the optimized watermark fragment allocation scheme.

[0030] The formulas for calculating the watermark capacity of each modality in the adaptive multimodal watermark capacity evaluation algorithm are as follows: Text modality: ; in For the watermark capacity of the text modality, Text modal capacity coefficient ( ), For text length, This refers to semantic redundancy.

[0031] Image modality: ; in For the watermark capacity of the image modality, This is the image modal capacity coefficient (set to 0.01). Image width, Image height, Let be the image complexity.

[0032] Audio modality: ; in For the watermark capacity of the audio modality, This is the audio modal capacity factor (set to 0.001). Audio duration (in seconds) Sampling frequency, The masking threshold.

[0033] Numerical data: ; in For the watermark capacity of the numerical modality, This is the numerical modal capacity coefficient (set to 0.1). The number of values. The number of bits of available precision.

[0034] The watermark segmentation allocation optimization adopts a linear programming method, with the constraint that the watermark information of each modality does not exceed its capacity limit, and the objective function is to maximize the robustness score of the overall watermark.

[0035] The complete mathematical expression for watermark fragment allocation optimization is to maximize the objective function: ; in To maximize, The summation symbol is used to represent the summation symbol. For modality robustness coefficient, To assign to the mode Number of watermark bits For modality The time decay function, This is the expected storage time.

[0036] 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 that modality; the number of watermark bits allocated to each modality must not be less than the minimum necessary number of bits for that modality, 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.

[0037] , , , These represent the time decay rates for text, audio, image, and numerical modalities, respectively, reflecting the resistance of each modality to common attacks. In this embodiment, , , , This optimization problem is solved using the branch and bound method, with the maximum number of iterations set to... The convergence accuracy is .

[0038] Step 3: Based on the watermark embedding strategy, the encoded anonymous identifier is embedded into different modalities of the scientific research data to achieve multimodal collaborative watermark embedding. In this step, we first acquire the modal content of the research data, the identity information encoding fragments generated in step 2, and the watermark embedding strategy table. Then, we use appropriate watermark embedding algorithms for different modalities to embed the identity information fragments into each modality, and use homomorphic encryption algorithms to protect the watermark mapping information. Finally, we output multimodal research data with embedded watermarks, an encrypted watermark information distribution mapping table (recording the watermark position and extraction method for each modality), and a watermark embedding strength parameter set.

[0039] It should be noted that this implementation method employs the following specific algorithms for watermark embedding in different modalities: For text-based modalities, a semantically preserving watermarking algorithm is used. This algorithm takes the original text content and a watermark information bit sequence as input, embeds the watermark information through synonym substitution, syntactic structure adjustment, etc., and outputs text content that retains the semantics but includes the watermark.

[0040] For image modalities, a DCT domain watermarking algorithm is used. This DCT domain watermarking algorithm takes the original image data and the watermark information bit sequence as input, and embeds the watermark by modifying the mid-frequency components of the DCT coefficients, outputting image data with a basically unchanged visual effect but containing the watermark.

[0041] For audio modalities, a spread spectrum watermarking algorithm is used. This algorithm takes the original audio signal and a watermark information bit sequence as input, calculates the watermark information by using the spread spectrum sequence and a masking threshold, and outputs audio data that retains the watermark without affecting the auditory effect.

[0042] For numerical data, a least significant bit replacement algorithm is used. This algorithm takes the original numerical dataset and the watermark information bit sequence as input, and embeds the watermark by replacing the least significant bit of the numerical data, outputting numerical data that contains the watermark but with minimal impact on precision.

[0043] Furthermore, this implementation uses the Paillier homomorphic encryption algorithm to protect the watermark information distribution mapping table. This Paillier homomorphic encryption algorithm receives watermark location information as input and outputs an encrypted mapping table, ensuring that even if the mapping table is obtained, the specific location of the watermark cannot be known without the key. The encrypted mapping function in the Paillier homomorphic encryption algorithm... The message is encrypted using the Paillier homomorphic encryption algorithm, whereby... This is a plaintext message. The encrypted mapping function ensures that the encrypted data can be homomorphically operated on without revealing the original location information, using random numbers.

[0044] In the Paillier homomorphic encryption algorithm, the public key parameter (That is, the modulus of the Paillier encryption algorithm, which determines the security strength of the encryption system) The length is set to 2048 bits. Two large prime numbers with the same number of digits (The first large prime number in Paillier's algorithm) and The result is obtained by multiplying (the second largest prime number in Paillier's algorithm). Parameters (The generators in the Paillier algorithm) must guarantee that they satisfy the correctness and security requirements of the Paillier encryption algorithm. Parameters for and The least common multiple of. Parameters (The non-zero integers used in the Paillier algorithm for key generation) are non-zero integers. Random numbers. (The random number used in Paillier encryption) needs to be... Select from the range, and require... and Mutual elements. Plain text message. (The plaintext data to be encrypted) is limited to Within a certain range, to ensure the security and correctness of encryption operations.

[0045] like Figure 2 As shown, this diagram compares the watermark capacity and robustness coefficients of different data modalities. It reflects the core parameters of the multimodal watermark allocation strategy in step 200 of the patent, demonstrates the differences in watermark carrying capacity among different data modalities, and intuitively illustrates the role of each modality in security protection.

[0046] Step 4: After completing the watermark embedding, use anonymous identifiers to construct a revocable mapping mechanism between anonymous identifiers and real identities, and store the mapping relationship in segments. In this step, 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 first obtained. Then, the identity mapping relationship is divided into multiple fragments using a secret sharing algorithm, and a threshold encryption algorithm is used to ensure that the complete mapping relationship can only be reconstructed when specific conditions are met. Finally, the identity mapping relationship fragment set (distributed and stored on different authorized nodes), the revocation authorization key share (distributed to each authorized party), and mapping relationship metadata (including creation time and expiration conditions) are output. The specific condition is that the complete identity mapping relationship can only be reconstructed through the threshold decryption algorithm when no less than a preset threshold number of authorized nodes jointly submit the identity mapping relationship fragments they hold.

[0047] It should be noted that the above secret-sharing algorithm is implemented using the Shamir secret-sharing scheme. This algorithm takes as input the mapping relationship between anonymous identifiers and users' real identities, and outputs multiple shares, ensuring that the original secret can only be reconstructed by collecting a sufficient number of shares.

[0048] In the Shamir secret sharing scheme, the threshold parameter Set as ,in For the total number of shares, This represents the floor function. In this embodiment, Values ,therefore That is, at least Only one share is needed to rebuild the secret. The polynomial degree is set to... , ensure any Each share cannot obtain any information about the secret. Polynomial coefficients ( In a finite field Randomly selected from among the prime numbers The bit length is twice the secret length, and is at least 256 bits.

[0049] The calculation method for each share is as follows: First, construct a polynomial with the secret to be shared as the constant term and the remaining coefficients as random numbers. Then, assign a unique identifier to each share (this identifier is within a specified range). Substitute this identifier into the polynomial and take the modulus over a finite field to obtain the corresponding share value. In this way, only by collecting a sufficient number of different shares can the original secret be reconstructed through polynomial interpolation.

[0050] This implementation combines threshold encryption with time-lock encryption to form an enhanced revocable mapping protection mechanism. This enhanced revocable mapping protection mechanism not only requires the cooperation of a preset threshold number of authorized parties to decrypt the mapping relationship, but also sets a time constraint, ensuring that even with sufficient authorization, decryption cannot be completed before the predetermined time, thus providing dual security guarantees. This enhanced revocable mapping protection mechanism receives the secret mapping relationship and security parameters as input, and outputs the encrypted mapping relationship and distributed decryption key shares.

[0051] In the enhanced revocable mapping protection mechanism, the threshold value The setting method is as follows: the total number of authorized parties Multiply by two-thirds and round up to get the minimum number of authorized parties required for collaboration. (Time lock constraint) The difficulty parameter represents the minimum computation time required for the unlocking operation. In this embodiment, the computational complexity of the time lock is set to one million consecutive square operations. The time lock function encrypts the plaintext mapping relationship, and unlocking is only possible after sufficient computation. Specifically, it is implemented based on the RSA algorithm using a 2048-bit modulus. The logic for the decryption threshold verification is as follows: only when the number of collected valid decryption shares is not less than the threshold value... The system will only allow decryption operations under certain conditions; otherwise, the decryption request will be rejected.

[0052] like Figure 3 As shown, the mapping relationship between anonymous identifiers and real identities, as well as the multi-party authorization mechanism, are illustrated. Corresponding to the revocable mapping mechanism in step 4, this intuitively demonstrates the role of secret sharing and threshold encryption technology in balancing the protection of user privacy and accountability.

[0053] In this embodiment of the application, to further enhance system security, the following steps may also be implemented: Step 4.1: Generate temporary access tokens and configure behavior monitoring; Based on the zero-knowledge proof verification result from step 1, the anonymous identifier, and the data access policy configuration file, a time-sensitive access credential is created using a timestamped token generation algorithm. Simultaneously, a normal access behavior model is constructed, and the allowed scope of operations is determined. The specific implementation is as follows: Step 4.2: Generate a time-limited access token; It adopts the JWT (JSON Web Token) format and contains an anonymous identifier (without any information that can identify the user's real identity), permission scope (the types of data and operations that are allowed to be accessed), validity period (the expiration time set based on the current timestamp), request source characteristics (such as IP range, device characteristics, etc.), and digital signature (to ensure the integrity and non-forgeability of the token).

[0054] The attribute credential verification process in this step is implemented as follows: First, the system verifies the digital signature of the trusted institution using the RSA-PSS algorithm (parameters: salt length 32 bytes, hash function SHA-256) to confirm the authenticity of the attribute credential submitted by the user. After successful verification, the system uses an attribute value range mapping table to map the user's attributes to a standard format that the system can process. This mapping table is stored in the security key management module and obtained from the key server via the HTTPS protocol (TLS 1.3). For numeric attributes, the normalization process uses a min-max scaling method to map the values ​​to the [0, 1] range.

[0055] The specific implementation of the zero-knowledge proof generation algorithm is based on the Bulletproofs framework, using elliptic curve secp256k1 as the underlying cryptographic primitive. This framework supports efficient zero-knowledge proofs for range proofs and predicate logic. The specific steps are as follows: converting user attributes into arithmetic circuit representations; hiding attribute values ​​using Pedersen commitments; constructing arithmetic circuits proving that attributes satisfy access conditions; converting interactive proofs into non-interactive formats using the Fiat-Shamir transform; and generating compact zero-knowledge proofs. The system uses C++ to implement the core cryptographic primitives and interacts with the Python business logic layer through the CFFI interface. The average time complexity of zero-knowledge proof generation is O(log n). ,in The number of constraints included in the proof.

[0056] The behavior pattern baseline establishment algorithm in this embodiment is a multi-dimensional behavior feature model constructed based on user historical access data. This algorithm not only considers static access permissions but also integrates dynamic features such as time, frequency, and operation sequences to construct a more refined behavior baseline, providing a basis for subsequent abnormal behavior detection. The 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... Defined from historical access records In the middle, within the time window Internally extract features such as access frequency, data volume, and operation type distribution. The process outputs a standardized multidimensional behavioral feature vector.

[0057] Time window in the behavior pattern baseline establishment algorithm It is set as a sliding window, with the basic unit being hours. In this embodiment, the window size is set to [size missing]. Hours (corresponding to short-term, medium-term, and long-term behavioral patterns respectively). Characteristics Includes: access frequency characteristics (Number of visits per unit time, standardized to) (interval), data volume characteristics (Data access volume per unit time, standardized to) (Interval), Operation Type Distribution Characteristics (A vector representation of the proportion of various operations, with each component in...) The interval and the sum is ).

[0058] When processing behavioral features, feature extraction functions standardize each feature. Specifically, they subtract the historical minimum value of each feature from its current value, and then divide by the difference between its historical maximum and minimum values, thus mapping all feature values ​​to the range of 0 to 1. Here, the historical minimum and maximum values ​​refer to the minimum and maximum values ​​that the feature has ever taken in historical data, respectively.

[0059] The behavioral baseline model is represented using a Gaussian Mixture Model (GMM). The model's parameters include the mean vector, covariance matrix, and the mixing weights of each Gaussian component. These parameters are obtained by training historical behavioral data using the Expectation-Maximization (EM) algorithm, thus effectively characterizing the multidimensional distribution of user behavioral features.

[0060] Step 5: Upon detecting a data breach, extract watermark information based on the breached data to reconstruct the anonymous identifier; In this step, we first obtain samples of suspected leaked scientific research data, the encrypted watermark information distribution mapping table generated in step 3, and the watermark extraction authorization certificate. Then, we use a multimodal watermark detection algorithm to extract watermark information from each modality in parallel, recover potentially damaged identity information fragments through an error correction decoding algorithm, and confirm the integrity of the identity information using a cross-validation algorithm. Finally, we output the extracted identity information fragment set, the watermark integrity score, and the preliminary anonymized identifier reconstruction results.

[0061] It should be noted that the extraction method corresponding to the embedding algorithm is used for watermark extraction for different modalities, as detailed below: For text-based modalities, a semantic watermark extraction algorithm is used. This algorithm takes the text with the embedded watermark as input, extracts the watermark information by analyzing word substitution patterns, and outputs the extracted watermark bit sequence.

[0062] For image modalities, a DCT domain watermark extraction algorithm is used. This DCT domain watermark extraction algorithm takes an image with an embedded watermark as input, extracts watermark information by analyzing the DCT coefficients, and outputs the extracted watermark bit sequence.

[0063] For audio modes, a spread spectrum watermark extraction algorithm is used. This algorithm receives the audio signal with the embedded watermark as input, extracts the watermark information by calculating the cross-correlation with the spread spectrum sequence, and outputs the extracted watermark bit sequence.

[0064] For numerical data, a least significant bit extraction algorithm is used. This algorithm takes the numerical data with the embedded watermark as input, analyzes the least significant bits to extract the watermark information, and outputs the extracted watermark bit sequence.

[0065] The multimodal cross-validation mechanism in this embodiment is an innovative watermark extraction and enhancement technique. This mechanism does not rely on the integrity of a single modality's watermark; instead, it aggregates watermark information extracted from each modality and comprehensively utilizes Reed-Solomon error correction capabilities and cross-modal data consistency checks. This enables reliable reconstruction of the anonymous identifier even when some modal watermarks are severely damaged or missing. The multimodal cross-validation mechanism receives watermark fragments extracted from each modality as input and outputs the reconstructed anonymous identifier and a watermark integrity score.

[0066] Watermark Integrity Scoring in Multimodal Cross-Validation Mechanism The calculation formula is: ; in Scoring for multimodal watermark integrity; Indicates the summation symbol; For modality Weighting factors; For modality The credibility of watermark extraction.

[0067] Credibility of watermark extraction The calculation method is as follows: First, compare the Hamming distance between the original watermark and the extracted watermark, that is, count the number of bits that are inconsistent at the same position. Then, divide this number of inconsistent bits by the total length of the original watermark to obtain the proportion of erroneous bits. Finally, subtract this proportion from 1 to obtain the watermark extraction confidence score. In other words, the closer the original watermark and the extracted watermark are, the higher the confidence score; if they are completely consistent, the confidence score is 1; if they are completely inconsistent, the confidence score is 0.

[0068] Watermark Reconstruction Threshold Set as ,when When the watermark extraction results are considered reliable, then... When, enable enhanced recovery mode; when If this occurs, watermark extraction is deemed unsuccessful. The error correction capability parameter of Reed-Solomon error correction coding is set to correct errors not exceeding 30% of the total number of symbols.

[0069] Weighting factors in multimodal cross-validation mechanisms The weights are dynamically adjusted over time. Specifically, each modality (such as text, image, audio, and numerical value) is assigned an initial weight during the initial watermark embedding process: 0.25 for text, 0.30 for image, 0.25 for audio, and 0.20 for numerical value. As the time (in months) after watermark embedding increases, the weight of each modality gradually decreases according to its time sensitivity coefficient. The time sensitivity coefficient reflects the stability of different modal watermarks over time; the larger the coefficient, the faster the weight decays. In this way, the system can dynamically reduce its dependence on time-sensitive modalities, improving the long-term reliability of overall watermark extraction.

[0070] The time sensitivity coefficient of the text modal watermark is represented by the value of . This reflects that the rate at which the text watermark decays over time is relatively low (i.e., the text replacement frequency is low, and the watermark is relatively stable). The time sensitivity coefficient representing the image modality has a value of [value missing]. This is considered medium sensitivity (common image processing operations may damage the watermark). The time sensitivity coefficient for the audio modality, with values ​​ranging from [value missing]. It is highly sensitive (audio processing and compression operations are common, and watermarks are easily affected). The time sensitivity coefficient for numerical modes, with values ​​ranging from [value missing]. The sensitivity level is considered low to medium (numerical data modifications are minimal, but precision adjustments occur frequently). Based on the time sensitivity coefficient of each modality, the system dynamically adjusts the weights of each modality over time, reducing the weights of high-sensitivity modalities and increasing the weights of stable modalities to maximize the success rate of long-term watermark extraction. All dynamic weights are normalized.

[0071] like Figure 4 As shown, the success rate of watermark extraction varies when different modal data are subjected to attacks of varying strengths. This corresponds to the watermark extraction and multimodal cross-validation mechanism in step 5, verifying the anti-attack capability of the multimodal collaborative watermarking technology and reflecting the robust technical effect of the patent implementation.

[0072] In this embodiment of the application, to further enhance data security monitoring capabilities, the following steps may also be implemented: Step 5.1: Monitor data access behavior in real time and update the risk score; 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 pattern is analyzed in real time using the sliding window algorithm, the anomaly detection algorithm is used to identify behaviors that deviate from the normal pattern, and the user's trust level is dynamically adjusted.

[0073] Before applying the risk scoring algorithm, the system first performs dimensional standardization on the multidimensional abnormal behavior indicators: different types of abnormal indicators (such as access frequency abnormal scores, data volume abnormal scores, operation type abnormal scores, etc.) are standardized and converted into a standard distribution with a mean of 0 and a variance of 1; time-decrease weights are applied to time-related abnormal signals to ensure that recent abnormal behaviors have higher impact weights; and principal component analysis (PCA) is used to map the multidimensional abnormal features to a unified risk space. These preprocessing steps ensure that the risk scoring function... In this process, different units and types of anomaly indicators can be effectively weighted and combined to generate consistent risk scores; among them, This represents a multidimensional abnormal behavior feature vector at the current moment, reflecting the user's performance on various abnormal behaviors; Historical risk scores represent a user's past risk status; This is a weight vector for each abnormal behavior type, used to adjust the degree of influence of different abnormal types on the overall risk score; This is a time decay factor used to control the rate at which the impact of historical anomalous behavior on the current risk score decreases over time.

[0074] It should be noted that the sliding window algorithm used in this embodiment receives real-time access log streams as input and outputs behavioral statistical features within a specific time window, providing basic data for anomaly detection.

[0075] Furthermore, this implementation uses the IsolationForest algorithm for abnormal behavior detection. The IsolationForest algorithm takes the current behavioral features and a baseline model as input and outputs an abnormal behavior score to determine whether the current operation deviates from the normal pattern.

[0076] The user risk score dynamic update algorithm in this embodiment is an innovative algorithm that integrates multi-source anomaly signals and considers time-series decay characteristics. This algorithm achieves accurate quantification and dynamic adjustment of user risk status by assigning different weights to different types of abnormal behavior and introducing an exponential decay function to handle the timeliness of historical anomaly records. The algorithm receives anomaly detection results and historical risk scores as input, and outputs the updated user risk score and corresponding access permission adjustment suggestions. The scoring function in the user risk score dynamic update algorithm... Defined as: the current abnormal behavior Compared with historical risk scores Combined, weight vectors based on different types of abnormal behavior and time decay factor Calculate the updated risk score and return a real value between 0 and 1, representing the user's current risk level.

[0077] Risk scoring function The specific calculation formula is as follows: ; in This represents the risk scoring function. The summation symbol is used to represent the summation symbol. Score historical risk. This is the historical score weighting factor (with a value of 0.7). Abnormal behavior type The weighting coefficients, Abnormal behavior type The degree of abnormality, This is the time decay factor (with a value of 0.1 / hour). This represents the time (in hours) that has elapsed since the last rating.

[0078] 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.

[0079] 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: ; 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.

[0080] 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.

[0081] 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.

[0082] 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. This step first obtains the anonymous identifier reconstructed in step 5, the mapping relationship fragments and revocation key shares held by each authorized party in step 4, as well as the revocation request and legal authorization documents. Then, the legality of the revocation request is confirmed using a threshold signature verification algorithm, the complete mapping relationship is restored from the fragments using a secret reconstruction algorithm, and the entire revocation process is recorded using an audit log generation algorithm. Finally, the decrypted true identity information, the revocation operation audit report, and the evidence chain integrity certificate are output.

[0083] It should be noted that the secret reconstruction algorithm in this embodiment is based on the principle of Lagrange interpolation. This algorithm receives a sufficient number of secret shares as input, reconstructs the original polynomial through interpolation, and outputs the restored anonymous identifier and its mapping to the real identity.

[0084] The threshold signature verification and audit log generation mechanism in this embodiment is an innovative combination to ensure the legality and traceability of the revocation process. This mechanism not only requires the authorized party to sign and approve the revocation request upon reaching a preset threshold, but also uses hash chain technology to link each step of the revocation process into an immutable audit trail, ensuring the legality of the operation and the transparency of the process. This threshold signature verification and audit log generation mechanism receives the revocation request and authorized signature as input, and outputs the verification result and a complete operation audit chain.

[0085] Threshold value in threshold signature verification mechanism Set as the total number of authorized parties Two-thirds, and rounded up. In other words, the system will calculate based on the actual number of authorized parties. Calculate the minimum number of valid signatures from authorized parties required (i.e., the threshold value). Only after certain conditions are met can the revocation operation be performed. For example, in this embodiment, the total number of licensors... for Then the threshold value for That is, at least A valid signature from the authorizing party. Signature verification function. Defined as: a set of verified signatures For the request The validity of the signatures. Specifically, first, the signature set is statistically analyzed. The number of valid signatures that pass mathematical verification in the middle is denoted as . If and only if The function returns True if all valid signatures pass verification, otherwise it returns False.

[0086] The audit log hash chain uses the SHA-256 algorithm, with each log entry... The hash value is calculated as follows: ; in For the first The hash value of the entry. This is the current timestamp. For operation description, For the digital signature of the operator, This represents a string concatenation operation; This indicates the Secure Hash Algorithm SHA-256.

[0087] In this embodiment of the application, in order to form a complete security closed loop, the following steps may also be implemented: Step 6.1: Generate a chain of evidence for accountability and update the protection strategy; Based on the watermark extraction results, identity mapping reconstruction results, access behavior logs, and leakage event context information, a complete leakage propagation path is constructed through evidence correlation analysis algorithm. The responsibility determination algorithm is used to determine the responsible party based on multi-source evidence, and the security rules are updated according to the leakage pattern.

[0088] Before conducting evidence association analysis, the system performs unified preprocessing on multi-source heterogeneous evidence data: watermark matching results are standardized to convert confidence scores in the [0, 1] interval; spatiotemporal features are uniformly encoded, converting time information into relative timestamps and spatial information into standardized location codes; and behavioral features are vectorized, mapping different types of behavioral patterns to feature vector space. These preprocessing steps ensure that evidence data from different sources and with different scales can be analyzed for association within a unified mathematical framework.

[0089] Furthermore, the output of the evidence association analysis algorithm is decoded and transformed to generate actionable accountability results: the abstract evidence chain diagram structure is transformed into a time-series visualized leakage propagation path; confidence scores and interpretability analysis reports are generated for the accountability determination results; and protection strategy parameters are transformed into specific security configuration instructions, including directly executable operation instructions such as watermark parameter adjustment, access control rule modification, and monitoring threshold update.

[0090] The evidence correlation analysis algorithm in this embodiment is an innovative algorithm for multi-source heterogeneous data fusion. It identifies potential leakage paths and key nodes by establishing a spatiotemporal correlation model between leaked data characteristics and access behavior. This evidence correlation analysis algorithm receives leaked sample characteristics and historical access logs as input and outputs a visualized leakage propagation chain and a list of key nodes. The correlation function in the evidence correlation analysis algorithm... Defined as: watermark evidence Access logs In the time window The system performs matching within the system, constructs an evidence chain graph structure by calculating spatiotemporal consistency scores and behavioral feature similarities, and outputs a set of possible paths for leakage propagation.

[0091] Time window in evidence association analysis algorithm This parameter is set to be adjustable, with the basic unit being days. In this embodiment, the value range is [value range missing]. Days, default value is Heaven. Spatiotemporal consistency score. The calculation formula is: ; in Indicates the spatiotemporal consistency score. For time similarity, For positional similarity, This is the time weighting factor (with a value of 0.6).

[0092] The similarity of behavioral features is calculated using the cosine similarity method. Specifically, the system represents the watermark features and access behavior features as vectors, and then measures their similarity by calculating the cosine of the angle between these two vectors. The closer the cosine value is to 1, the more similar the behavioral features are; conversely, the smaller the value, the lower the similarity.

[0093] Association threshold Set as .in, The behavioral feature similarity score represents the similarity between the watermark feature vector calculated using the cosine similarity method and the access behavior feature vector, with a value range of [value missing]. The closer the value is to 1, the more similar the behavioral characteristics. Ultimately, when At that time, it was determined that a relationship existed.

[0094] The evidence chain construction algorithm uses the maximum likelihood estimation method to calculate the probability scores of different propagation paths, and outputs probabilities exceeding a threshold. (Set as) The set of paths.

[0095] Temporal similarity in evidence association analysis algorithms The complete formula for calculating the time-importance decay function is as follows: ; in For time similarity, To leak the time of the incident, For the time of suspicious behavior, This is a reference time point (usually the time when the leak was discovered). It is a function of time importance; The time importance function is used to measure the time point of an event. Relative to reference time point The importance of this. Specifically, with... distance The greater the distance, the more exponentially the importance decreases. The rate at which the importance of time decays is determined by a coefficient. The control factor is set to 2.0 in this embodiment, and the maximum time range for consideration is 365 days. In other words, events closer to the reference time point have a higher weight in the evidence analysis, while events farther away from the reference time point have a lower weight.

[0096] The time importance function ensures that time similarity assessment considers both the temporal proximity of events and their importance relative to a reference time point, giving more recent events higher weight in evidence association. Location similarity. The calculation is based on a weighted combination of network topology and geographic location, which weights the similarity between network topology and geographic location. The liability determination algorithm in this embodiment employs an evidence weight fusion model, comprehensively evaluating multiple factors such as watermark matching degree, behavioral anomaly degree, and spatiotemporal consistency to generate a credibility score for liability determination. This algorithm receives data from multiple evidence sources as input and outputs the liability determination result and the contribution weight of each piece of evidence.

[0097] Credibility scoring in liability determination algorithms The calculation formula is: ; in Indicates credibility score. For evidence type Weighting factors For evidence type Intensity rating The summation symbol is used.

[0098] In this embodiment, the evidence types include: watermark matching evidence (weighted) ), evidence of abnormal behavior (weight) ) and spatiotemporal consistency evidence (weight) Threshold for determining liability. Set as ,when When, it is considered a high-credibility liability determination; when When, it is considered a moderately reliable determination; when At that time, it was deemed a low-credibility judgment, requiring further investigation and evidence collection.

[0099] The credibility score calculation for liability determination incorporates a time-dimensional evidence strength attenuation factor. The complete scoring formula is as follows: ; in: For a moment Credibility score for liability determination; For evidence type Weighting factors; For evidence type Intensity rating; For evidence type The time decay function reflects the evidence Since the collection began Changes in the effectiveness of the Empress Dowager; The summation symbol is used.

[0100] The time decay function means that the validity of evidence gradually decreases as time goes by after it has been collected. Specifically, each type of evidence has a corresponding time decay rate. ,and This represents the number of days that have elapsed since the evidence was collected. The time decay function is based on... and The numerical value is used to dynamically adjust the weight of evidence, so that the older the evidence is, the less influence it has in determining liability, reflecting the reality that evidence gradually loses its reliability over time.

[0101] The time decay rate of different types of evidence reflects the rate at which the reliability of evidence decreases over time: watermark matching evidence. (Low attenuation rate, watermark evidence is relatively stable), evidence of abnormal behavior (High decay rate, behavioral patterns may change rapidly over time), evidence of spatiotemporal consistency (Medium attenuation rate). This formula ensures that newly collected evidence carries higher weight than older evidence during the liability determination process, and that the validity period of different types of evidence matches their inherent reliability characteristics. The liability determination results are automatically updated every 24 hours to reflect the impact of new evidence and time factors.

[0102] The adaptive optimization algorithm for protection strategies in this embodiment can extract feature patterns from past leakage events, identify system vulnerabilities and protection blind spots, and adjust watermark parameters, access control rules, and monitoring thresholds accordingly. This algorithm receives leakage case data and the current protection configuration as input and outputs an optimized set of protection strategy parameters.

[0103] The adaptive optimization algorithm for protection strategies fully considers the time factor in its design. Its optimization objective is to enable the system to select the optimal protection action based on its current state at different time points, thereby maximizing overall benefits in the long run. The algorithm maps the system's state to corresponding protection measures through a policy function and introduces a discount factor in the decision-making process, gradually reducing the emphasis on future benefits. The evaluation of overall benefits considers not only the improvement in security but also the impact on system performance and changes in user experience. Specifically, the algorithm calculates the improvement in security, changes in performance overhead, and changes in user experience separately, assigning different weights to each. Security has the highest weight, followed by performance and user experience, with weight ratios of 0.6, 0.2, and 0.2, respectively.

[0104] The probability of system state changes is modeled based on historical operating data and expert experience, and time series analysis is used to capture the time dependence of state transitions. To ensure the timeliness of the model, the state transition probability matrix is ​​updated quarterly.

[0105] Furthermore, the algorithm incorporates time-related constraints to ensure system stability and security. For example, the update frequency of the protection strategy is limited to a maximum of once per week to prevent system instability caused by frequent adjustments; the magnitude of each strategy adjustment is also limited to within ±30% to prevent excessive changes from affecting system operation. During high-risk periods (such as nighttime and weekends), the system automatically adopts a more stringent security strategy, and the corresponding parameter adjustment magnitude is increased to 1.5 times that during normal periods to enhance protection capabilities.

[0106] like Figure 6 As shown, the complete process from data access to accountability is demonstrated, corresponding to the evidence correlation analysis and accountability determination algorithm in step 6.1. It fully demonstrates the accountability tracing path after scientific research data leakage and verifies the closed-loop accountability effect of the multi-model collaborative protection mechanism in practical applications.

[0107] A computer storage medium includes a memory and one or more processors, wherein executable code is stored in the memory, and when the one or more processors execute the executable code, it is used to implement the above-described method for protecting scientific research data based on multi-model collaboration.

[0108] Here, the present invention provides an implementation example: A national-level medical imaging research center has established a multimodal medical database containing brain MRI, CT, and pathology images. This database needs to be shared among medical research institutions nationwide for the development of AI-assisted diagnostic systems. It contains sensitive medical data from 500,000 patients, requiring both patient privacy protection and data security and traceability. Each dataset consists of clinical descriptive text, multimodal medical images, patient physiological indicators, and doctor's voice recordings for some cases, totaling over 200TB.

[0109] Researcher A from a medical research institution applied for access to a glioma dataset, submitting an application that included identity attributes such as the researcher's institutional qualifications, research project number, and data security level permit. After collecting these attributes, the system does not store the original information but instead verifies the researcher's access eligibility using zero-knowledge proof technology and generates an anonymous identifier. Examples of user identity attributes and anonymous identifier generation are shown in Table 1. Table 1: Examples of User Identity Attributes and Anonymous Identifier Generation

[0110] The system analyzed the structure of the glioma dataset requested by researcher A, which includes four modalities: clinical description text, MRI / CT images, patient physiological indicators, and doctor's voice recordings. The system calculated the watermark capacity based on the characteristics of each modality and formulated the optimal watermark allocation strategy. The results of the multimodal medical data feature analysis are shown in Table 2. Table 2: Results of Multimodal Medical Data Feature Analysis

[0111] The results of the watermark fragment allocation strategy are shown in Table 3: Table 3: Results of Watermark Fragment Allocation Strategy

[0112] The system extends researcher A's anonymous identifier (256-bit hash value) to 320 bits using Reed-Solomon encoding and embeds it into various modalities according to the strategy described above. Different watermarking algorithms are used for each modality during the embedding process to ensure data availability.

[0113] Six months later, a dataset of gliomas, suspected to originate from a national-level medical imaging research center, appeared on a medical forum. After obtaining the samples, the security department initiated an accountability process, using multimodal watermark extraction technology to recover the anonymized identifiers. The watermark extraction results of the leaked data are shown in Table 4. Table 4: Watermark Extraction Results of Leaked Data

[0114] After initiating a revocation request and obtaining permission through legal procedures, the system notified the five authorizing parties to execute the identity mapping revocation agreement. Ultimately, four authorizing parties provided valid key shares, meeting the threshold requirement (minimum three shares), and successfully reconstructed the identity mapping relationship. The multi-party authorization revocation process record is shown in Table 5: Table 5: Record of the Revocation Process of Multiple Authorizations

[0115] Through multiple authorization revocations, the system ultimately confirmed that Researcher A was responsible for the leaked data. According to the investigation, the researcher violated the data usage agreement by providing the data to an unauthorized third party. The system accordingly updated its security policy, enhanced behavioral monitoring rules for this type of research project, and optimized the watermark allocation strategy to improve tracking capabilities. This case demonstrates how a multi-model collaborative approach to scientific data security can successfully protect medical data privacy while simultaneously achieving accountability for data breaches.

[0116] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for protecting scientific research data security based on multi-model collaboration, characterized in that, include: Obtain user identity attribute declarations and access request parameters, and generate anonymous identifiers and attribute proofs using zero-knowledge proof technology; Based on anonymous identifiers and attribute proofs, we analyze the multimodal composition structure of scientific research data and generate corresponding watermark embedding strategies accordingly. Based on the watermark embedding strategy, the encoded anonymous identifier is embedded into different modalities of scientific research data to achieve multimodal collaborative watermark embedding. After completing the watermark embedding, an anonymous identifier is used to construct a revocable mapping mechanism between anonymous identifiers and real identities, and the mapping relationship is stored in segments. Upon detecting a data breach, the watermark information is extracted from the breached data to reconstruct the anonymous identifier; By combining the reconstructed anonymous identifiers and executing a multi-party authorization revocation protocol, the identity mapping relationship is reconstructed after authorization is obtained, thereby enabling accountability for data breaches.

2. The scientific research data security protection method based on multi-model collaboration according to claim 1, characterized in that, The steps of obtaining user identity attribute declarations and access request parameters, and generating anonymous identifiers and attribute proofs using zero-knowledge proof technology, specifically include: preprocessing user identity attribute data, including normalizing numerical attributes, encoding categorical attributes, and converting time-related attributes to relative timestamps; generating verifiable attribute proofs using a zero-knowledge proof generation algorithm without revealing specific attribute values; and generating unique anonymous identifiers using a hash chain algorithm.

3. The scientific research data security protection method based on multi-model collaboration according to claim 1, characterized in that, The steps of analyzing the multimodal composition structure of scientific research data and generating corresponding watermark embedding strategies include: preprocessing each type of data to achieve consistency in the dimensions of different modal data; using a multimodal content analysis algorithm to identify the feature regions of each modality; calculating the amount of information that can be embedded in each modality using an adaptive multimodal watermark capacity evaluation algorithm; and using the ReedSolomon error correction coding algorithm to encode and segment identity information.

4. The scientific research data security protection method based on multi-model collaboration according to claim 1, characterized in that, In the multimodal collaborative watermarking embedding step, the following watermarking algorithms are used for different modalities: for text modality, a semantically preserved watermarking algorithm is used; for image modality, a DCT domain watermarking algorithm is used; for audio modality, a spread spectrum watermarking algorithm is used; and for numerical data, a least significant bit replacement algorithm is used.

5. The scientific research data security protection method based on multi-model collaboration according to claim 1, characterized in that, The steps for constructing the revocable mapping mechanism between anonymous identifiers and real identities specifically include: The identity mapping relationship is segmented using the Shamir secret sharing algorithm; Threshold encryption algorithms are used to ensure that the complete mapping relationship can only be reconstructed when certain conditions are met. These conditions mean that the complete identity mapping relationship can only be reconstructed through threshold decryption when no fewer than a preset threshold number of authorized nodes jointly submit the identity mapping relationship fragments they hold. The identity mapping relationships are distributed and stored on different authorized nodes.

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

7. A scientific research data security protection method based on multi-model collaboration according to claim 1, characterized in that, The steps for executing the multi-party authorization revocation protocol specifically include: confirming the legality of the revocation request through a threshold signature verification algorithm; restoring 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. A scientific research data security protection method based on multi-model collaboration according to claim 1, characterized in that, The step of extracting watermark information and reconstructing anonymous identifiers based on leaked data also includes a step of real-time monitoring of data access behavior and updating risk scores. Specifically, this includes: analyzing access patterns in real time using a sliding window algorithm based on the user's real-time access log stream, behavior monitoring rule set, and historical behavior baseline model; using the IsolationForest algorithm to detect abnormal behavior; and dynamically adjusting the user risk score by combining historical scores, the weight of abnormal behavior types, the degree of abnormality, and the time decay factor through a user risk score dynamic update algorithm.

9. A scientific research data security protection method based on multi-model collaboration according to claim 1, characterized in that, In the step of executing the multi-party authorization revocation agreement and reconstructing the identity mapping relationship after obtaining authorization, in order to form a complete security closed loop, it also includes the steps of generating an accountability evidence chain and updating the protection strategy. Specifically, this includes: constructing a complete leakage propagation path based on watermark extraction results, identity mapping reconstruction results, access behavior logs, and leakage event context information through evidence association analysis algorithms; determining the responsible party based on multi-source evidence using a responsibility determination algorithm; and updating security rules according to the leakage pattern through a protection strategy adaptive optimization algorithm.

10. A computer storage medium, characterized in that, The device includes 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 scientific data security protection method based on multi-model collaboration as described in any one of claims 1-9.

Images