Blockchain-based multi-center tumor data collaboration system

Abstract

The present application belongs to the field of medical data privacy computing and blockchain technology, and specifically relates to a tumor multi-center data collaboration system based on a blockchain. First, the privacy sensitivity is calculated according to the genomic mutation spectrum, the clipping threshold and the privacy budget are determined in linkage, and the clipping perturbation data package is generated and chained. Second, the reputation score and the ladder margin are calculated according to the violation record and the sensitivity, and the authorized data carrying the reputation certificate is issued. Finally, the weighted aggregation is executed by combining the multi-modal feature cross-modal alignment with the reputation quantification contribution, the behavior fingerprint is constructed by extracting the gradient sequence, the attack source is located, and the zero-knowledge proof is chained. Through the dynamic noise adding mechanism driven by sensitivity and reputation and the ladder margin mechanism, combined with the zero-knowledge tamper-proof audit technology, the present application realizes the balance between model high availability and high privacy security in tumor multi-center data sharing, effectively preventing illegal and unauthorized attacks and gradient reverse attacks.

Description

Technical Field

[0001] This invention relates to the fields of medical data privacy computing and blockchain technology, specifically a blockchain-based multi-center tumor data collaboration system. Background Technology

[0002] With the rapid development of precision medicine and artificial intelligence technologies, in-depth research on the diagnosis and treatment of tumor diseases increasingly relies on massive, multimodal clinical data. Since the number of tumor samples and subtype distribution held by a single medical institution are limited, cross-institutional multi-center tumor data collaboration has become an inevitable trend for pooling high-quality medical resources and promoting the generalized application of medical models. In multi-center collaborative scenarios in the medical field, the flow and sharing of tumor data not only requires strict adherence to high standards of privacy protection and compliance regulations, but also necessitates the establishment of reliable trust mechanisms, transparent access governance rules, and a fair data contribution quantification system among participating centers. Blockchain technology, with its decentralized, immutable, smart contract-automatic execution, and end-to-end traceability characteristics, naturally aligns with the needs for rights confirmation, authorization, and supervision in multi-party medical collaboration. It provides an ideal technological foundation for building a mutually trusted and recognized data sharing network, fully leveraging the aggregated value of tumor data, and comprehensively meeting the integrated needs of privacy encapsulation, dynamic governance, and traceability auditing in the process of multi-institutional joint modeling. Therefore, there is an urgent need to develop a blockchain-based multi-center tumor data collaboration system.

[0003] However, existing technologies have two main shortcomings in multi-center tumor data collaboration scenarios: First, existing privacy protection strategies often use globally fixed parameters, failing to dynamically adjust for the sensitivity differences of different tumor subtypes. Furthermore, data access mechanisms are mostly static authorizations, lacking a dynamic tiered economic constraint mechanism that combines high data sensitivity with the requester's historical reputation, making it difficult to balance model availability while accurately defending against unauthorized access. Second, existing node contribution quantification dimensions are relatively singular, failing to deeply integrate complex cross-modal semantics and the node's long-term compliance reputation. Moreover, when facing malicious behaviors such as gradient inversion attacks, there is a lack of in-depth mining of cross-round behavioral characteristics. Traditional auditing methods often rely on accessing plaintext gradients, posing a risk of secondary privacy leaks and failing to achieve a trustworthy traceability and accountability system that balances content anonymity with the tamper-proof characteristics of blockchain.

[0004] Therefore, this invention proposes a blockchain-based multi-center tumor data collaboration system. Summary of the Invention

[0005] To address the technical problems mentioned in the background section regarding the rigid privacy protection strategies and singular access mechanisms in existing multi-center collaborative tumor scenarios, which fail to dynamically combine constraints based on data sensitivity and node reputation, and also face the challenge of a single dimension for quantifying node contributions and the reliance on plaintext gradients in traditional auditing methods which can easily lead to secondary privacy leaks, the present invention aims to provide a blockchain-based multi-center tumor data collaboration system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A blockchain-based multi-center tumor data collaboration system includes:

[0008] M1: Privacy encapsulation module, which uses information entropy to calculate the subtype privacy sensitivity score based on the genome mutation spectrum, and adjusts the pruning threshold and privacy budget to perform differential privacy noise addition, generating pruning perturbation data packets containing perturbation feature vectors and writing them into the consortium blockchain network;

[0009] M2: Access governance module, receives the trimming disturbance data packet, calculates the reputation anchor score and tiered deposit by combining the violation record with the sensitivity score, and issues an authorization data packet containing reputation certificate after the pledge is approved. When a violation occurs, a penalty is deducted and the reputation anchor score is updated.

[0010] M3: Contribution aggregation module, receives the authorized data packet, performs cross-modal alignment by extracting the perturbation feature vector, quantifies the contribution score by combining the Shapley value and the reputation anchor score, and generates a global aggregation gradient based on the secret sharing protocol and writes it into the consortium blockchain network.

[0011] M4: The source tracing and auditing module extracts multi-round local gradients to construct cross-round behavioral fingerprint vectors, locates the attack source through Mahalanobis distance deviation, generates a plaintext-free zero-knowledge source tracing proof, and writes it into the consortium blockchain network.

[0012] Furthermore, the privacy encapsulation module is specifically used for:

[0013] The local tumor sample set was divided into multiple subtype subsets according to subtype, and the mutation frequency of genomic mutation sites within each subtype subset was counted to construct a mutation spectrum vector;

[0014] Based on the mutation spectrum vector, a smoothing constant is introduced, and the privacy sensitivity score of each subtype is measured using information entropy. Then, the privacy sensitivity score is used as the driving force, and the gradient pruning threshold and privacy budget corresponding to each subtype are calculated in conjunction with the preset maximum pruning threshold and the maximum privacy budget.

[0015] Furthermore, the privacy encapsulation module is further used for:

[0016] The original feature vector is pruned using the gradient pruning threshold to obtain a pruned feature vector. The noise standard deviation is determined by the gradient pruning threshold, the privacy budget, and the differential privacy relaxation parameter. Gaussian noise determined by the noise standard deviation is then injected into the pruned feature vector to obtain a perturbed feature vector. Finally, the perturbed feature vector, the privacy sensitivity score, the gradient pruning threshold, and the privacy budget are jointly encapsulated into a pruning perturbation data packet carrying a subtype sensitivity label and uploaded to the consortium blockchain network.

[0017] Furthermore, the access governance module is specifically used for:

[0018] The privacy sensitivity score is extracted from the received cropped and perturbed data packet. Based on the applicant's historical violation records stored in the consortium blockchain network, the applicant's reputation anchoring score is calculated by combining the applicant's historical cumulative violation count and violation decay coefficient. Then, the reputation anchoring score and the privacy sensitivity score are combined, and a sensitivity deposit adjustment coefficient is introduced to adjust the preset benchmark deposit amount. The tiered deposit amount required for the applicant to access the corresponding subtype data is calculated and determined.

[0019] Furthermore, the access governance module is further used for:

[0020] The smart contract issues a staking instruction to the applicant. After the tiered guarantee amount is pledged and the matching verification is passed, the reputation anchoring score is encapsulated into a reputation certificate, and together with the perturbation feature vector and the privacy sensitivity score, it is encapsulated into an authorized cropped perturbation data packet and sent to the applicant.

[0021] If any data usage violations are detected through the smart contract during the applicant's usage period, the tiered guarantee amount will be automatically deducted, and the applicant's cumulative number of violations will be incremented by 1 to calculate the updated reputation anchor score. Finally, the deduction record and the updated reputation anchor score will be written into the consortium blockchain network.

[0022] Furthermore, the contribution aggregation module is specifically used for:

[0023] The received authorized cropping perturbation data packets are modally split and averaged to obtain node-level image, genome, and pathology modal feature vectors. These vectors are then projected onto a unified semantic space using a preset projection matrix to perform cross-modal semantic alignment and are concatenated and fused into a cross-modal fused feature vector. Subsequently, the Shapley value of each node is quantified using the Monte Carlo Shapley approximation algorithm. The Shapley value is multiplied by the reputation anchoring score in the reputation credential and normalized to obtain the cross-modal contribution score. Finally, using the cross-modal contribution score as the weight, the local model gradient uploaded by each node through the secret sharing protocol is demasked and weighted for secure aggregation to generate the global aggregated gradient.

[0024] Furthermore, the source tracing and auditing module is specifically used for:

[0025] The system receives global aggregation results with cross-modal contribution scores, constructs a cross-round gradient upload behavior sequence from the local model gradients uploaded by each node in multiple consecutive aggregation rounds, extracts the mean vector and variance vector of the sequence, concatenates them, and weights them with the cross-modal contribution scores of the corresponding nodes to construct node behavior fingerprint vectors. Then, it calculates the normal contribution baseline and baseline covariance matrix based on the node behavior fingerprint vectors of all historical compliant nodes, and calculates the Mahalanobis distance of each node behavior fingerprint vector relative to the normal contribution baseline. When the Mahalanobis distance exceeds a preset anomaly detection threshold, the corresponding node is determined to be a potential gradient reverse attack source node and is packaged into an abnormal node evidence package.

[0026] Furthermore, the source tracing and auditing module is further used for:

[0027] After obtaining the abnormal node evidence package, a verification logic is constructed through a non-interactive zero-knowledge proof protocol. Without touching the plaintext content of the cross-round gradient upload behavior sequence, a zero-knowledge tracing proof is generated to verify that the Mahalanobis distance calculation process is correct and exceeds the preset anomaly detection threshold. Finally, the zero-knowledge tracing proof, the abnormal node evidence package, and the current aggregation round number are jointly encapsulated into a timestamp tracing record and appended to the consortium blockchain network in chronological order to form a global audit chain.

[0028] Compared with the prior art, the advantages of the present invention are as follows:

[0029] 1. This invention utilizes a privacy encapsulation module to calculate the privacy sensitivity score of each subtype based on the local tumor genome mutation spectrum. This adaptively adjusts the gradient pruning threshold and privacy budget, breaking away from the limitations of traditional one-size-fits-all fixed parameters. It maximizes the usability of the model while strengthening the protection of highly sensitive data. Simultaneously, the access governance module combines the data subtype sensitivity with the applicant's historical violation records, innovatively introducing a tiered margin mechanism. This achieves dual risk premium pricing for data access. Through smart contract execution of dynamic authorization and automated penalties, it constructs a rigorous dynamic game constraint from both economic and credit dimensions, effectively curbing unauthorized access by low-credit nodes.

[0030] 2. In the model aggregation stage, this invention not only performs unified semantic alignment of multimodal features from imaging, genomics, and pathology, but also combines node reputation credentials and Shapley value approximation algorithms to accurately quantify the comprehensive data quality and historical compliance level of each node. This is used as the weight for weighted secure aggregation, preventing low-quality or malicious data from polluting the global model from the source. In the security audit stage, this invention creatively extracts the gradient behavior sequence of nodes across aggregation rounds to construct a behavioral fingerprint vector. The Mahalanobis distance deviation of the normal contribution baseline is used to accurately locate the source of gradient reverse attacks. A non-interactive zero-knowledge proof protocol is used to generate a traceability proof, which is then written into the blockchain. This allows the system to achieve automated, tamper-proof determination of rights and responsibilities without ever touching the plaintext gradients of each node, completely eliminating the risk of secondary privacy leakage in traditional auditing processes.

[0031] 3. This invention deeply integrates the immutability of consortium blockchains with the automated execution capabilities of smart contracts, enabling the full-chain temporal on-chaining of authorization credentials, penalty records, global gradients, and traceability proofs, thus forming a legally binding distributed trust system. Furthermore, at the underlying algorithmic level, this invention introduces a smoothing constant in information entropy calculation to prevent numerical divergence, utilizes cross-modal contrastive learning to pre-train projection matrices to overcome the semantic gap in heterogeneous data, and introduces relaxation parameters in differential privacy noise addition to balance the probability of privacy loss. These advancements significantly improve the numerical stability and feature representation capabilities of complex calculations in multi-center medical federated learning scenarios, providing a solid technical foundation for the secure sharing and efficient circulation of medical data. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of the system workflow of the present invention;

[0034] Figure 2 This is a schematic diagram of the privacy encapsulation module workflow of the present invention;

[0035] Figure 3 This is a schematic diagram of the workflow of the traceability audit module of the present invention;

[0036] Figure 4 This is a schematic diagram of multi-node collaboration and data flow timing provided in an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] To achieve the above objectives, the present invention provides a blockchain-based multi-center tumor data collaboration system, such as... Figures 1-3 As shown, the system includes:

[0039] M1: Privacy encapsulation module, which uses information entropy to calculate the subtype privacy sensitivity score based on the genome mutation spectrum, and adjusts the pruning threshold and privacy budget to perform differential privacy noise addition, generating pruning perturbation data packets containing perturbation feature vectors and writing them into the consortium blockchain network.

[0040] In this embodiment, the original tumor samples collected locally by each participating central node are multimodal integrated data, specifically image data. genomic data and pathological data The system is composed of three modalities. Considering that the privacy risks contained in each modality are significantly different, especially since the mutation distribution of genomic data can specifically expose patient characteristics, the system first conducts an independent privacy sensitivity assessment of the genomic modality through the privacy encapsulation module before carrying out multimodal joint modeling and extraction. The more uniform the genomic mutation distribution, the stronger the privacy sensitivity.

[0041] M101: Divide the local tumor sample set into subtypes. To quantify the differentiated privacy protection needs of each tumor subtype subset, targeting the [number] subtype subset, [further details are needed]. For each subtype subset, the system extracts genomic modality information separately at this stage and calculates the preset... The mutation frequency at each site in the genome is used to construct a mutation spectrum vector. :

[0042]

[0043] in, For subtype number, and , The total number of subtypes; Number the mutation site, and , This represents the total number of mutation sites. For the first The first of the subtypes The mutation frequency of each mutation site;

[0044] The mutation frequency The corresponding calculation formula is: ;in, For the first The first subtype subset The number of samples with mutations at each mutation site; For the first The total number of samples in each subtype subset;

[0045] Based on the mutation spectrum vector Information entropy is used to measure the privacy sensitivity score of each subtype. The calculation formula is:

[0046]

[0047] in, A smoothing constant is used to prevent logarithmic operations from diverging, and ;

[0048] In this embodiment, based on the principle of numerical stability for zero-probability terms in information theory, a small smoothing constant is added to the input term of the logarithmic operation when the sample frequency approaches zero to prevent the operation from diverging. The preferred value range for is defined between and , and those skilled in the art can determine a specific value within this range based on the accuracy and numerical stability requirements of tumor mutation frequency data; the smoothing constant described in this embodiment The preferred value is At this point, the privacy sensitivity score is... The disturbance magnitude is controlled within Within this range, it will not affect the privacy sensitivity scores of each subtype. The relative size relationship;

[0049] The privacy sensitivity score The larger the value, the more likely it is to be the first. The more uniform the distribution of genomic mutations among subtypes, the stronger the privacy sensitivity; the privacy sensitivity score... The subtype sensitivity label will be written into the trimmed perturbation data packet. The access control module uses the privacy sensitivity score as a basis. Determine the tiered guarantee amount;

[0050] M102: After obtaining the privacy sensitivity score Then, based on the aforementioned privacy sensitivity score For the calculation of the driving quantity linkage, the first Gradient clipping thresholds for each subtype With privacy budget The calculation formula is:

[0051]

[0052] in, Set the maximum cropping threshold; This is the adjustment coefficient for the cropping threshold sensitivity, and ;

[0053]

[0054] in, To preset the maximum privacy budget; This is a privacy budget sensitivity adjustment factor, and ;

[0055] In this embodiment, based on the gradient pruning threshold design principle in differential privacy federated learning, it is a recognized trade-off in the field that when the gradient pruning threshold is too large, the noise amplitude increases, leading to a decrease in model accuracy, while when it is too small, the gradient direction distortion rate increases. The preset maximum pruning threshold is used to determine this trade-off. The preferred value range is defined as between 0.5 and 2.0. Those skilled in the art can determine the value within this range based on the gradient amplitude distribution of multicenter tumor data and the model convergence speed requirements. The preset maximum pruning threshold mentioned in this embodiment... The preferred value is 1.0, at which point a balance is achieved between gradient direction distortion and noise amplification, and the number of convergence epochs increases by no more than 8% compared to the unpruned baseline.

[0056] In this embodiment, based on the linear sensitivity adjustment principle in differential privacy multi-category data differentiation protection, the pruning threshold sensitivity adjustment coefficient is... The preferred value range is defined as between 0.5 and 2.0. Those skilled in the art can determine the appropriate value based on the privacy sensitivity score for each subtype. The normalized magnitude and the gradient clipping threshold The dynamic range requirement is to specify a value within this range; the clipping threshold sensitivity adjustment coefficient mentioned in this embodiment The preferred value is 1.0, at which point the gradient clipping threshold is... The dynamic range covers the preset maximum clipping threshold. The range of 50% to 100% meets the strong protection requirements of highly sensitive subtypes;

[0057] In this embodiment, based on the privacy budget range design principle in differential privacy engineering deployment, the preset maximum privacy budget... The preferred value range is defined as between 1.0 and 10.0. Those skilled in the art can determine the value within this range based on the regulatory compliance requirements of multi-center tumor data and the tolerance for accuracy loss in downstream classification tasks; the preset maximum privacy budget mentioned in this embodiment The preferred value is 8.0, at which point the accuracy loss of downstream tumor classification tasks is less than 3% and meets the regulatory requirements for safe sharing of medical data;

[0058] In this embodiment, based on the principle of differentiated allocation of privacy budget in differential privacy multi-subtype scenarios, the privacy budget sensitivity adjustment coefficient is... The preferred value range is defined as between 0.5 and 3.0. Those skilled in the art can determine the specific value within this range based on the privacy protection strength requirements of rare mutation subtypes and the budget loss tolerance of low-sensitivity subtypes. The privacy budget sensitivity adjustment coefficient described in this embodiment... The preferred value is 1.5, at which point the actual privacy budget for the highly sensitive subtype is compressed to the preset maximum privacy budget. The privacy budget loss for low-sensitivity subtypes should be less than 25%, while the privacy budget loss for low-sensitivity subtypes should not exceed 15%.

[0059] The privacy sensitivity score The larger the gradient clipping threshold, the better. With the privacy budget Simultaneous reduction achieves a linked enhancement of both cutting intensity and privacy protection intensity;

[0060] M103: The first The first subtype in the subset Each original tumor sample is input into a pre-trained multimodal feature extraction network deployed locally at each participating central node for forward mapping calculation to obtain the original feature vector representing the multimodal information of the original tumor sample. At this time, the original tumor sample input into the network for joint modeling is restored to a complete modal overview, specifically represented as image data. genomic data and pathological data ;

[0061] The specific feature extraction process is as follows: First, the image feature extraction branch network with ResNet-50 as the backbone, the genome coding branch network with multi-hidden layer perceptron, and the pathological feature extraction branch network based on visual Transformer, all included in the pre-trained multimodal feature extraction network, are used to perform independent forward mapping calculations on the image data, the genome data, and the pathological data, respectively, to extract the single-modal feature vectors corresponding to each modality. The calculation formula is as follows:

[0062]

[0063]

[0064]

[0065] in, , , These are the extracted image, genomic, and pathological single-modal feature vectors, respectively. , , These represent the forward mapping functions of the corresponding branch networks; These are the weight parameters for the image feature network; For genome network weight parameters; These are the network weight parameters for pathological features;

[0066] In this embodiment, based on the prior requirement of performing transfer learning on a large-scale medical image dataset to extract the underlying visual features of lesions, the image feature network weight parameters are... The values ​​are taken from a pre-trained high-dimensional tensor containing approximately 25.5 million floating-point elements; based on the nonlinear mapping requirement of autoencoding reconstruction on massive gene sequencing data to learn the co-expression features of gene loci, the genomic network weight parameters are... The values ​​are taken from a pre-trained weight matrix containing hundreds of thousands to millions of floating-point elements; based on the need for high-dimensional feature extraction to capture cell nuclear polymorphism and complex spatial dependencies in publicly available digital pathological slides, the weight parameters of the pathological feature network are... The value is a pre-trained high-dimensional tensor containing approximately 86 million floating-point elements;

[0067] Subsequently, each of the single-modal feature vectors is projected onto a unified semantic space using a preset projection matrix to perform cross-modal semantic alignment, and then concatenated and fused to obtain the original feature vectors representing the multimodal information of the original tumor sample. The calculation formula is:

[0068]

[0069] in, , , These are the preset projection matrices corresponding to imaging, genomics, and pathological modalities, respectively. , , This is the bias term for the corresponding modal projection; This represents a vector concatenation operation;

[0070] Then, based on the gradient clipping threshold For the original feature vector The L2 norm clipping yields the clipped feature vector, calculated as follows:

[0071]

[0072] in, For the first Sample IDs within each subtype subset; For the first The first subtype subset The original feature vectors corresponding to each sample; Original feature vector The L2 norm; The cropped feature vector;

[0073] After obtaining the cropped feature vector Then, the cropped feature vector is... Injection at the gradient clipping threshold With the privacy budget The Gaussian noise of the joint calibration is calculated as follows:

[0074]

[0075] in, It follows a Gaussian distribution; The standard deviation of noise; To the clipped feature vector Identity matrices of the same dimension; The perturbation feature vector;

[0076] The noise standard deviation The gradient clipping threshold With the privacy budget Substituting all the values ​​into the following formula, we get:

[0077]

[0078] in, For privacy relaxation parameters, and difference ; The natural logarithm operator;

[0079] In this embodiment, the differential privacy relaxation parameter The system defines the upper limit for tolerance of privacy guarantee failure, and limits the differential privacy relaxation parameters to establish a substantially effective protection boundary. The value of is no greater than the reciprocal of the total number of samples used in training. To accurately balance privacy strength and model usability in medical data collaboration scenarios and avoid excessive noise or failure of privacy defenses, the differential privacy relaxation parameter... The preferred value range is defined within to Between these parameters, each central node can dynamically adapt; the differential privacy relaxation parameters described in this embodiment... The preferred value is At this point, the probability of privacy protection failure is negligible, and the noise standard deviation... The calculation results are within a reasonable range, meeting the dual requirements of privacy protection and model availability in the multi-center tumor data federated learning scenario;

[0080] The noise standard deviation With the gradient clipping threshold Reduce and decrease, along with the privacy budget Decrease and increase; the disturbance feature vector Satisfying the privacy budget With the differential privacy relaxation parameters For differential privacy guarantees of parameters, the perturbation feature vector Will participate in cross-modal semantic alignment and weighted secure aggregation computation in the contribution aggregation module;

[0081] M104: Finally, the perturbation feature vector The privacy sensitivity score mentioned above The gradient clipping threshold and the privacy budget Encapsulated as trimmed and perturbed data packets carrying subtype sensitivity tags ,and The data is uploaded to the consortium blockchain network and received and processed by the access governance module.

[0082] M2: Access governance module, receives the trimmed disturbance data packet, calculates the reputation anchor score and tiered deposit by combining the violation record with the sensitivity score, issues an authorization data packet containing reputation credentials after the pledge is approved, deducts penalties and updates the reputation anchor score when a violation occurs.

[0083] M201: Receives the trimming perturbation data packet And extract the cropped perturbation data packets. The subtype sensitivity label, which is the privacy sensitivity score, is mentioned in the text. Based on the applicant's historical violation records stored in the consortium blockchain network, the applicant's... Reputation anchoring score The calculation formula is:

[0084]

[0085] in, Set a maximum score for reputation anchoring; The attenuation coefficient is the violation factor, and ; For the applicant node The cumulative number of violations in history;

[0086] The reputation anchoring score The smaller the value, the more likely it is to be the requesting node. The lower the historical compliance level, the lower the reputation anchoring score. It will participate in the tiered guarantee amount calculation in step M202 and be encapsulated into the authorized trimmed disturbance data packet in step M203. Then it is passed to the contribution aggregation module;

[0087] In this embodiment, the preset reputation anchoring score upper limit is... The preferred value range is defined as between 0.8 and 1.0. Those skilled in the art can set a specific value within this range based on the initial trust level of the system. The upper limit of the preset reputation anchoring score mentioned in this embodiment is... The preferred value is 1.0;

[0088] In this embodiment, if the attenuation coefficient is too large, a single violation will cause a sharp drop in the reputation anchoring score, resulting in excessive punishment for occasional errors; if it is too small, the cumulative punishment for violations will be insufficient to restrain malicious behavior. The violation attenuation coefficient... The preferred value range is defined as between 0.1 and 0.5. Those skilled in the art can determine the specific value within this range based on the alliance network's tolerance and penalty requirements for historical violations; the violation attenuation coefficient described in this embodiment... The preferred value is 0.3, which is the reputation anchor score after a node has committed 3 historical violations. Reduced to the preset reputation anchoring score upper limit Less than 40%;

[0089] M202: After obtaining the reputation anchoring score Then, the reputation anchor score is used. With the privacy sensitivity score Jointly determine the tiered guarantee amount The calculation formula is:

[0090]

[0091] in, The preset benchmark guarantee amount; This is the sensitivity margin adjustment factor, and ; For the applicant node Request to access No. The tiered guarantee amount required to pledge when trimming perturbation data packets of each subtype;

[0092] The tiered guarantee amount According to the privacy sensitivity score As the reputation anchor score increases, so does the reputation anchor score. By reducing the size of the data, a dual risk premium can be achieved for both highly sensitive data and low-compliance nodes.

[0093] In this embodiment, based on the design principle of the benchmark margin amount in blockchain smart contracts, the preset benchmark margin amount is... The preferred value range is defined as between 100 and 500 units of measurement. Those skilled in the art can determine a specific value within this range based on the incentive mechanism design of the alliance network and the magnitude of data value. The preset benchmark guarantee amount mentioned in this embodiment... The preferred value is 200 units of measurement;

[0094] In this embodiment, based on the design principle that data sensitivity and access cost are positively correlated in the differential privacy data market, the sensitivity deposit adjustment coefficient is... The preferred value range is defined as between 0.5 and 2.0. Those skilled in the art can determine the appropriate value based on the subtype sensitivity score. The actual distribution range and the required margin difference are clearly defined within this range; the sensitivity margin adjustment coefficient mentioned in this embodiment The preferred value is 1.0;

[0095] M203: Upon obtaining the aforementioned tiered guarantee amount Then, a smart contract is used to notify the applicant. Issue a pledge instruction, requiring the applicant to The tiered guarantee amount is pledged to the smart contract. After the pledge is confirmed, the applicant is encrypted according to the attribute-based encryption strategy. The set of institutional attributes is matched and verified against a preset data access strategy tree; after successful matching and verification, the reputation anchor score is... Encapsulate and generate reputation credentials; combine the reputation credentials with the trimmed and perturbed data packets. The included perturbation feature vector and the privacy sensitivity score mentioned above They are jointly encapsulated into authorized trimmed perturbation packets carrying reputation credentials. ,and Issued to the applicant It is received and processed by the contribution aggregation module;

[0096] M204: At the applicant's Use the authorized trimmed perturbation data packets During this period, smart contracts were used to contact the applicant. The data usage behavior of the applicant will be continuously monitored; the monitoring content includes the applicant's... For the perturbation feature vector Whether the access scope, access frequency, and data usage purpose are consistent with the authorization terms in the preset data access policy tree;

[0097] If no violations are detected, the tiered guarantee amount... The authorized cropping of the perturbation data packet The money will be automatically refunded to the applicant by the smart contract after the usage period ends. The record of this compliant use will be written into the consortium blockchain network;

[0098] If a violation is detected, the smart contract will automatically deduct the tiered guarantee amount. The penalty records and evidence of violations will be written into the consortium blockchain network; the applicant will... The cumulative number of violations Updated to And calculate the updated reputation anchoring score. The calculation formula is: ;

[0099] The updated reputation anchoring score Write it into the consortium blockchain network, replacing the original reputation anchor score. For the applicant The updated reputation anchor score will be used the next time an access request is initiated. Recalculate the tiered guarantee amount use.

[0100] M3: Contribution aggregation module, receives the authorized data packet, performs cross-modal alignment by extracting the perturbation feature vector, quantifies the contribution score by combining the Shapley value and the reputation anchor score, and securely aggregates and generates a global aggregation gradient based on the secret sharing protocol, which is then written into the consortium blockchain network.

[0101] M301: Receives authorized pruning perturbation data packets from all participating central nodes. ,in , The total number of central nodes participating in the federated aggregation; pruning perturbation data packets from each of the authorized data packets. Extract the perturbation feature vectors respectively. The privacy sensitivity score mentioned above and the reputation anchoring score ;

[0102] Based on the pre-defined modal segmentation index agreed upon during the alliance initialization phase, for each node The perturbation feature vector Perform modal decomposition and extract image modal perturbation subvectors. Genome mode perturbation subvector and pathological modal perturbation subvector ;

[0103] For nodes The node-level image feature vector is obtained by averaging the modal perturbation subvectors of all samples. Node-level genome feature vectors and node-level pathological feature vectors The calculation formula is:

[0104]

[0105] in, For nodes The number of samples involved in the calculation; They are nodes Inner Image, genomic and pathological modal perturbation vectors for each sample;

[0106] M302: After obtaining the node-level image feature vector The node-level genome feature vector and the node-level pathological feature vector Then, the feature vectors of the three modalities are projected onto a unified semantic space using a preset projection matrix, and cross-modal semantic alignment is performed. The calculation formula is as follows:

[0107]

[0108] in, , , These are preset projection matrices for imaging, genomics, and pathology modalities, respectively. The preset projection matrices are pre-trained through cross-modal contrastive learning during the consortium initialization phase and shared among the participating central nodes. , , They are nodes Aligning feature vectors for image, genome, and pathology modalities in a unified semantic space;

[0109] The three modality alignment feature vectors are concatenated and fused to obtain the node. Cross-modal fusion feature vector The expression is: ;in, This is a vector concatenation operation;

[0110] The preset projection matrix , , The alliance was pre-trained through cross-modal contrastive learning during the initialization phase. The specific training process is as follows:

[0111] Using a small number of labeled paired samples provided by each participating central node as training data, each record in the paired samples contains the original features of the same patient across three modalities: imaging, genome, and pathology, a paired sample set is constructed. ,in Number the paired samples. The total number of paired samples is given. Each paired sample is mapped to a unified semantic space using the preset projection matrix to obtain alignment feature vectors for three modalities. The positive sample objective is to maximize the cosine similarity between the three modal alignment feature vectors of the same patient, and the negative sample objective is to minimize the cosine similarity between the cross-modal alignment feature vectors of different patients. A cross-modal contrastive loss is then constructed. The calculation formula is:

[0112]

[0113] in, Number the negative samples during the traversal; The cosine similarity function; This is a temperature coefficient used to control the degree of concentration in the similarity distribution;

[0114] By minimizing the cross-modal contrast loss For the preset projection matrix , , Perform iterative updates until the cross-modal contrast loss is reached. Convergence, the converged preset projection matrix , , It is broadcast and shared among all participating central nodes and remains fixed during subsequent federated aggregation;

[0115] In this embodiment, based on the principle of temperature coefficient controlling the concentration of feature distribution in contrastive learning, the temperature coefficient... The preferred value range is defined as between 0.05 and 0.5. Those skilled in the art can determine the specific value within this range based on the size of the paired sample set and the modal feature dimension. The temperature coefficient mentioned in this embodiment... The preferred value is 0.1, at which point the intra-class similarity of cross-modal aligned feature vectors is significantly higher than the inter-class similarity, satisfying the feature discrimination requirements for subsequent Shapley value approximation calculation;

[0116] M303: Combining the aforementioned reputation anchoring score The Monte Carlo Shapley approximation algorithm is used to quantify the cross-modal contribution scores of each node; for nodes Through random sampling A number that does not contain nodes Random Alliance Approximate calculation nodes Shapley value The calculation formula is:

[0117]

[0118] in, Number of Monte Carlo samples; For the sampling number, and ; For the first The second sampling does not include nodes. Randomized alliance; The utility function is the cross-modal fusion feature vector of all nodes within the alliance. AUC index of tumor classification on the reserved validation set after aggregation; For nodes The Shapley value is used to quantify the nodes. Marginal contribution to the performance of the global model;

[0119] The utility function The specific function expression is defined as follows:

[0120]

[0121] in, Represents any given node alliance, as described above. Alliance or join a node The new alliance after; and These represent the set of positive and negative tumor samples in the reserved validation set, respectively. and These represent the total number of positive and negative samples in the reserved validation set, respectively. and These represent specific positive and negative samples, respectively. Indicates based on alliance The classification model constructed by cross-modal fusion feature aggregation of internal nodes provides a probability scoring function for predicting a positive result for an input sample; indicator function. The value is 1 when the predicted score of a positive sample is greater than the predicted score of a negative sample; otherwise, the value is 0.

[0122] In this embodiment, based on the trade-off principle between the number of Monte Carlo samplings and the estimated variance in the Shapley value approximation algorithm, the number of Monte Carlo samplings... The preferred value range is defined as between 50 and 200, and those skilled in the art can determine the optimal value based on the total number of participating central nodes. The computational resource constraints are explicitly set within this range; the number of Monte Carlo sampling times mentioned in this embodiment... The preferred value is 100, at which point the Shapley value is... The estimation error is within an acceptable range and the computational cost is reasonable;

[0123] In obtaining the Shapley value Then, the Shapley value is... With the reputation anchoring score After multiplication and normalization, the nodes are obtained. The cross-modal contribution score is calculated as follows:

[0124]

[0125] in, Let the summation traversal variable be the node number, and ; For nodes The Shapley value; For nodes Reputation anchoring score; For nodes The cross-modal contribution score, and ;

[0126] The cross-modal contribution score It also reflects the node's data quality contribution and historical compliance level; and is used as the weight in the weighted security aggregation in the M304 step, and is written into the consortium blockchain network after the M304 step is completed for the traceability and audit module to receive and process;

[0127] M304: After obtaining the cross-modal contribution score After that, each node Based on the cross-modal fusion feature vector Calculate the gradient of the local model locally. and the local model gradient is shared via a secret sharing protocol. After masking, the data is uploaded to the aggregation server; the aggregation server, unable to obtain the plaintext gradient of any single node, uses the cross-modal contribution score. Perform demasking aggregation on the mask gradients of each node for the weights to obtain the global aggregated gradient. The calculation formula is:

[0128]

[0129] in, For nodes The local model gradient;

[0130] The global aggregated gradient It integrates a weighted average of each node's cross-modal data quality contribution and historical compliance level; in obtaining the global aggregate gradient... Then, the global aggregated gradient Cross-modal contribution scores of each node Encapsulated as a global aggregation result with cross-modal contribution scores ,and .

[0131] M4: The source tracing and auditing module extracts multi-round local gradients to construct cross-round behavioral fingerprint vectors, locates the attack source through Mahalanobis distance deviation, generates a plaintext-free zero-knowledge source tracing proof, and writes it into the consortium blockchain network.

[0132] M401: Receives the global aggregation result To extract the global aggregated gradient and the cross-modal contribution scores of each node ; Regarding the first In each round of federated aggregation, nodes are extracted. The uploaded local model gradient , will node In continuous The local model gradient sequences uploaded within each aggregation round are used to construct nodes. Cross-round gradient upload behavior sequence The calculation formula is:

[0133]

[0134] in, For the aggregation round number, and ; This represents the total number of aggregate rounds within the audit window; For nodes In the The local model gradient uploaded during the round aggregation; For nodes The sequence of gradient upload behavior across rounds;

[0135] In this embodiment, based on the principle of balancing audit round coverage and computational overhead in the design of the federated learning audit window, the total number of aggregated rounds within the audit window is... The preferred value range is defined as between 10 and 50, and those skilled in the art can determine a specific value within this range based on the constraints of alliance network communication frequency and storage capacity; the total number of aggregation rounds within the audit window mentioned in this embodiment. The preferred value is 20, at which point the sample size of the behavioral sequence is sufficient to support stable statistical anomaly detection.

[0136] M402: After obtaining the cross-round gradient upload behavior sequence Then, based on the cross-modal contribution score... Build nodes for weights Node behavior fingerprint vector ;

[0137] First, the sequence of gradient upload behaviors across rounds is analyzed. The statistical features are extracted from the gradient at each round, and the calculation formula is as follows:

[0138]

[0139] in, For nodes The mean vector of gradients across rounds; For nodes The variance vector of the gradient across rounds;

[0140] The mean vector With the variance vector After splicing, the cross-modal contribution score is used. Weighted summation yields nodes Node behavior fingerprint vector The expression is:

[0141]

[0142] in, This is a vector concatenation operation; For nodes The node behavior fingerprint vector;

[0143] The cross-modal contribution score The larger the node The node behavior fingerprint vector In subsequent anomaly detection, the higher the weight, the more significant the impact of deviation from the normal behavior baseline on the overall audit conclusion;

[0144] M403: Obtain the node behavior fingerprint vector of each node. Then, the normal contribution baseline is calculated using the node behavior fingerprint vectors of all historical compliant nodes. With baseline covariance matrix The calculation formula is:

[0145]

[0146] in, To contribute baseline normally; The baseline covariance matrix; This is a vector transpose operation;

[0147] After obtaining the normal contribution baseline With the baseline covariance matrix Then, calculate each node. Node behavior fingerprint vector Relative to the normal contribution baseline Mahalanobis distance The calculation formula is: ;in, The baseline covariance matrix The inverse matrix;

[0148] Mahalanobis distance Larger values ​​indicate a larger node The greater the deviation of the gradient upload behavior from the normal contribution baseline;

[0149] After obtaining the Mahalanobis distance Then, using a preset anomaly detection threshold Decision Node Is it a potential source node for gradient inversion attacks? When the Mahalanobis distance... Exceeding the preset anomaly detection threshold At that time, determine the node As a potential source node for gradient inversion attacks, the node will be... number The Mahalanobis distance and the cross-round gradient upload behavior sequence Encapsulated as an abnormal node evidence package ,and ;

[0150] In this embodiment, based on the Mahalanobis distance threshold setting principle in statistical anomaly detection, the preset anomaly detection threshold is... The preferred value range is defined as between 2.0 and 4.0. Those skilled in the art can determine the value within this range based on the tolerance ratio of false positives and false negatives of the consortium network. The preset anomaly detection threshold mentioned in this embodiment... The preferred value is 3.0, at which point the false alarm rate is approximately 0.27% under the Gaussian distribution assumption, which meets the audit accuracy requirements of multi-center medical data collaboration scenarios;

[0151] M404: Upon obtaining the evidence package of the abnormal node Then, based on the abnormal node evidence package Generate zero-knowledge traceability proof Zero-knowledge traceability proof Constructed using a non-interactive zero-knowledge proof protocol, this allows the auditor to upload the sequence of actions across different rounds of gradient uploads without accessing the protocol. Given the plaintext content, the following proposition can be verified: Node The Mahalanobis distance Exceeding the preset anomaly detection threshold And the Mahalanobis distance Fingerprint vector of the node behavior The normal contribution baseline and the baseline covariance matrix The distance was correctly calculated based on the Mahalanobis distance calculation formula in M403.

[0152] The zero-knowledge traceability proof The abnormal node evidence package and the current aggregation round number Encapsulated as a timestamp traceability record ,and ;

[0153] The timestamp trace record Numbered by aggregation round The data are added to the consortium blockchain network in the order they appear, forming a global audit chain that is immutable and time-traceable.

[0154] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0155] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A blockchain-based multi-center tumor data collaboration system, characterized in that, include: M1: Privacy encapsulation module, which uses information entropy to calculate the subtype privacy sensitivity score based on the genome mutation spectrum, and adjusts the pruning threshold and privacy budget to perform differential privacy noise addition, generating pruning perturbation data packets containing perturbation feature vectors and writing them into the consortium blockchain network; M2: Access governance module, receives the trimming disturbance data packet, calculates the reputation anchor score and tiered deposit by combining the violation record with the sensitivity score, and issues an authorization data packet containing reputation certificate after the pledge is approved. When a violation occurs, a penalty is deducted and the reputation anchor score is updated. M3: Contribution aggregation module, receives the authorized data packet, performs cross-modal alignment by extracting the perturbation feature vector, quantifies the contribution score by combining the Shapley value and the reputation anchor score, and generates a global aggregation gradient based on the secret sharing protocol and writes it into the consortium blockchain network. M4: The source tracing and auditing module extracts multi-round local gradients to construct cross-round behavioral fingerprint vectors, locates the attack source through Mahalanobis distance deviation, generates a plaintext-free zero-knowledge source tracing proof, and writes it into the consortium blockchain network.

2. The blockchain-based multi-center tumor data collaboration system according to claim 1, characterized in that, The privacy encapsulation module is specifically used for: The local tumor sample set was divided into multiple subtype subsets according to subtype, and the mutation frequency of genomic mutation sites within each subtype subset was counted to construct a mutation spectrum vector; Based on the mutation spectrum vector, a smoothing constant is introduced, and the privacy sensitivity score of each subtype is measured using information entropy. Then, the privacy sensitivity score is used as the driving force, and the gradient pruning threshold and privacy budget corresponding to each subtype are calculated in conjunction with the preset maximum pruning threshold and the maximum privacy budget.

3. The blockchain-based multi-center tumor data collaboration system according to claim 2, characterized in that, The privacy encapsulation module is further used for: The original feature vector is pruned using the gradient pruning threshold to obtain a pruned feature vector. The noise standard deviation is determined by the gradient pruning threshold, the privacy budget, and the differential privacy relaxation parameter. Gaussian noise determined by the noise standard deviation is then injected into the pruned feature vector to obtain a perturbed feature vector. Finally, the perturbed feature vector, the privacy sensitivity score, the gradient pruning threshold, and the privacy budget are jointly encapsulated into a pruning perturbation data packet carrying a subtype sensitivity label and uploaded to the consortium blockchain network.

4. The blockchain-based multi-center tumor data collaboration system according to claim 1, characterized in that, The access governance module is specifically used for: The privacy sensitivity score is extracted from the received cropped and perturbed data packet. Based on the applicant's historical violation records stored in the consortium blockchain network, the applicant's reputation anchoring score is calculated by combining the applicant's historical cumulative violation count and violation decay coefficient. Then, the reputation anchoring score and the privacy sensitivity score are combined, and a sensitivity deposit adjustment coefficient is introduced to adjust the preset benchmark deposit amount. The tiered deposit amount required for the applicant to access the corresponding subtype data is calculated and determined.

5. The blockchain-based multi-center tumor data collaboration system according to claim 4, characterized in that, The access governance module is further used for: The smart contract issues a staking instruction to the applicant. After the tiered guarantee amount is pledged and the matching verification is passed, the reputation anchoring score is encapsulated into a reputation certificate, and together with the perturbation feature vector and the privacy sensitivity score, it is encapsulated into an authorized cropped perturbation data packet and sent to the applicant. If any data usage violations are detected through the smart contract during the applicant's usage period, the tiered guarantee amount will be automatically deducted, and the applicant's cumulative number of violations will be incremented by 1 to calculate the updated reputation anchor score. Finally, the deduction record and the updated reputation anchor score will be written into the consortium blockchain network.

6. The blockchain-based multi-center tumor data collaboration system according to claim 1, characterized in that, The contribution aggregation module is specifically used for: The received authorized cropping perturbation data packets are modally split and averaged to obtain node-level image, genome, and pathology modal feature vectors. These vectors are then projected onto a unified semantic space using a preset projection matrix to perform cross-modal semantic alignment and are concatenated and fused into a cross-modal fused feature vector. Subsequently, the Shapley value of each node is quantified using the Monte Carlo Shapley approximation algorithm. The Shapley value is multiplied by the reputation anchoring score in the reputation credential and normalized to obtain the cross-modal contribution score. Finally, using the cross-modal contribution score as the weight, the local model gradient uploaded by each node through the secret sharing protocol is demasked and weighted for secure aggregation to generate the global aggregated gradient.

7. The blockchain-based multi-center tumor data collaboration system according to claim 1, characterized in that, The source tracing and auditing module is specifically used for: The system receives global aggregation results with cross-modal contribution scores, constructs a cross-round gradient upload behavior sequence from the local model gradients uploaded by each node in multiple consecutive aggregation rounds, extracts the mean vector and variance vector of the sequence, concatenates them, and weights them with the cross-modal contribution scores of the corresponding nodes to construct node behavior fingerprint vectors. Then, it calculates the normal contribution baseline and baseline covariance matrix based on the node behavior fingerprint vectors of all historical compliant nodes, and calculates the Mahalanobis distance of each node behavior fingerprint vector relative to the normal contribution baseline. When the Mahalanobis distance exceeds a preset anomaly detection threshold, the corresponding node is determined to be a potential gradient reverse attack source node and is packaged into an abnormal node evidence package.

8. The blockchain-based multi-center tumor data collaboration system according to claim 7, characterized in that, The source tracing and auditing module is further used for: After obtaining the abnormal node evidence package, a verification logic is constructed through a non-interactive zero-knowledge proof protocol. Without touching the plaintext content of the cross-round gradient upload behavior sequence, a zero-knowledge tracing proof is generated to verify that the Mahalanobis distance calculation process is correct and exceeds the preset anomaly detection threshold. Finally, the zero-knowledge tracing proof, the abnormal node evidence package, and the current aggregation round number are jointly encapsulated into a timestamp tracing record and appended to the consortium blockchain network in chronological order to form a global audit chain.

Images