A cross-institutional nutrition data joint analysis method based on secure multi-party computation

Abstract

The application discloses a kind of cross-institutional nutrition data joint analysis methods based on secure multi-party computation, it is related to data transmission technical field, the method of the present application is by collecting intelligent terminal diet behavior and eating biological signal, gradient is calculated using sparse random variance reduction gradient method, according to the number of non-zero elements deduction chain privacy budget, the seed generated using eating biological signal is used to add secret sharing and key encapsulation to gradient, after data is received by aggregation node, first using homomorphic addition eliminates mask noise in ciphertext domain, combined with preset medical knowledge graph executes causal consistency check, removes the gradient that violates medical common sense and then updates global model with weighting, finally parse global weight and generate nutrition attribution graph;The present application effectively utilizes biological characteristic entropy to realize physical anti-fake, ensures that model meets medical prior distribution through logic gate control, while strictly protecting user privacy, realizes high precision and clinical interpretability of model.

Description

Technical Field

[0001] This invention relates to the field of data transmission technology, and in particular to a method for joint analysis of cross-institutional nutritional data based on secure multi-party computation. Background Technology

[0002] With the rapid development of IoT technology and smart healthcare, using digital means to monitor and manage dietary behavior has become an important trend in the field of public health. Accurate dietary data collection is the foundation for personalized nutritional intervention and blood pressure and blood sugar control. Traditional dietary recording methods rely on manual input by users, which suffers from problems such as large subjective errors and poor compliance. Therefore, automatic data collection technology based on smart hardware has gradually become a research hotspot, aiming to acquire users' eating data in real time through multimodal sensors to provide objective evidence for health management. In the existing technology, patent CN118538374A discloses a multifunctional smart eating device for monitoring dietary nutrition and realizing health management. This technical solution, through a built-in identification device in the handle of the device, can identify food types and calculate the cumulative nutrient intake for each meal and each day. The solution also outputs a dietary assessment report based on the user's body indicators. It provides personalized food recommendations and health guidance, with the advantage of automated collection and local integration of dietary data. It can provide individual users with intuitive nutritional intake feedback and short-term health warnings. However, the above-mentioned existing technologies still have obvious limitations when facing joint analysis of nutritional data from cross-institutions and large-scale groups. Existing technologies focus on the collection of local data and simple statistical analysis, lacking secure data sharing and model co-construction mechanisms, and cannot make full use of massive distributed data. In the process of data transmission and processing, directly uploading the original dietary characteristics or health indicators poses a significant risk of privacy leakage and lacks privacy protection measures. Existing intelligent collection devices mainly focus on the measurement of physical quantities and lack a mechanism to verify the authenticity of the data source. They cannot effectively identify whether the data is generated by real organisms eating and are easily attacked by simulated data, leading to distorted analysis results. Summary of the Invention

[0003] The technical problem solved by this invention is that existing technologies focus on the collection of local data and simple statistical analysis, lack secure data sharing and model co-construction mechanisms, and cannot make full use of massive distributed data. In the process of data transmission and processing, directly uploading the original dietary characteristics or health indicators poses a significant risk of privacy leakage and lacks privacy protection measures. Existing intelligent collection devices mainly focus on the measurement of physical quantities and lack a mechanism to verify the authenticity of data sources. They cannot effectively identify whether the data is generated by real organisms eating and are easily attacked by simulated data, leading to distorted analysis results.

[0004] To address the aforementioned technical problems, this invention provides the following technical solution: a method for joint analysis of cross-institutional nutritional data based on secure multi-party computation, comprising the following steps: Step S1: Collect raw data of dietary behavior, user health indicators and eating biosignals from the smart terminal, and preprocess the raw data of dietary behavior, user health indicators and eating biosignals. Step S2: Construct a nutrition regression model based on the original dietary behavior data and the user's health indicators, calculate the sparse gradient matrix, deduct the on-chain privacy budget based on the sparseness matrix, generate an encryption seed based on the eating biosignal, encrypt the sparse gradient using the encryption seed, and generate data to be uploaded to the cloud. Step S3: Use secure multi-party computation to perform causal consistency verification on the data to be uploaded to the cloud, and aggregate the data to be uploaded to the cloud that has passed the causal consistency verification to update the global nutrient regression model. Step S4: Construct a nutrient attribution map using the updated global nutrient regression model and generate feedback instructions to adaptively adjust the working mode of the smart terminal.

[0005] This invention generates encrypted seeds using biological signals from eating on a smart terminal, combines sparse gradients and blockchain privacy budgets at the transmission end, and performs causal consistency verification based on a medical knowledge graph in the cloud. This invention differs from traditional federated learning, which focuses only on the statistical characteristics of data while ignoring medical logic and the authenticity of data sources. It can effectively defend against simulated data attacks and model poisoning attacks while protecting user privacy data from leakage, ensuring that the final global nutritional regression model has both high accuracy and clinical interpretability.

[0006] Preferably, the raw data of the eating behavior of the smart terminal includes: food appearance image data, eating speed data, and food nutrient content data; The health indicators include systolic blood pressure, diastolic blood pressure, and blood glucose; The feeding biosignals include hand tremor signals; The preprocessing includes normalization and noise reduction; The raw data of dietary behavior is converted into dietary feature vectors.

[0007] Preferably, in step S2, the process of obtaining the sparse gradient matrix specifically includes: A nutrient regression model is constructed, which is in the form of a linear mapping function. The mathematical expression of the nutrient regression model is as follows: ; in, This is an augmented feature vector of dietary behavior. This is a matrix of dietary weight parameters corresponding to the feature dimensions of the original dietary behavior data and the feature dimensions of the health data. A vector composed of health indicators; The process of calculating the original gradient matrix of the diet weight parameter matrix in the nutrition regression model specifically includes: Based on the sparse random variance reduction gradient method, combined with Regularization constraint terms are applied, and the original gradient matrix is ​​calculated. The mathematical expression of the original gradient matrix is ​​as follows: ; in, Based on raw data of dietary behavior, health indicators, and current parameters The calculated stochastic gradient, For the snapshot parameter matrix, For snapshot parameter matrix The calculated reference gradient, The mean of the full gradients is based on the snapshot parameter matrix. Preset Regularization coefficient; The original gradient matrix is ​​filtered according to a preset retention ratio to obtain a sparse gradient matrix. The filtering operation specifically includes: using the dietary weight parameter matrix of the nutrition regression model to perform linear prediction on the dietary behavior feature vector in the original dietary behavior data, comparing the linear prediction result with health indicators, and calculating the fitting residual vector. When the magnitude of the fitted residual vector is greater than or equal to the preset residual threshold (1.0 in this embodiment), the preset retention ratio is increased; When the magnitude of the fitted residual vector is less than the preset residual threshold, the preset retention ratio is reduced. The original gradient matrix is ​​truncated according to the adjusted retention ratio. The top few gradients with the largest absolute values ​​are retained, and the remaining gradients are set to zero to obtain a sparse gradient matrix.

[0008] Preferably, in step S2, the operation of deducting the on-chain privacy budget specifically includes: The number of non-zero elements in the sparse gradient matrix is ​​counted, and the privacy budget consumption is calculated based on the preset unit privacy price. The privacy budget consumption is the product of the preset unit privacy price and the number of non-zero elements in the sparse gradient matrix. The smart terminal sends a verification request to the blockchain smart contract, and the smart contract reads the current privacy budget balance of the smart terminal's account on the blockchain; If the privacy budget consumption exceeds the difference between the smart terminal's privacy budget balance on the blockchain and the preset balance threshold, the verification fails, the smart contract refuses to issue the calculation access certificate, and the upload is terminated. If the privacy budget consumption is less than or equal to the difference between the privacy budget balance on the blockchain and the preset balance threshold, the verification is deemed successful. The smart contract then performs a deduction operation on the privacy budget balance according to the privacy budget consumption, updates the privacy budget balance, and generates a timestamped access certificate to be returned to the smart terminal.

[0009] Preferably, in step S2, the process of generating the data to be uploaded to the cloud specifically includes: The Shannon entropy of the hand microtremor signal is calculated, and the Shannon entropy is concatenated with the timestamp. A biomass seed is generated by using a hash function. The biomask seed is used to drive a pseudo-random number generator to generate a random mask matrix with the same dimension as the sparse gradient matrix. Based on a random mask matrix, the sparse gradient matrix is ​​split using an additive secret sharing algorithm to generate gradient secret shares that mask the original gradient values.

[0010] Preferably, in step S2, the process of generating the data to be uploaded to the cloud further includes: Based on the privacy budget consumption, combined with a pre-defined mapping function, the reputation weight is obtained, and the reputation weight is converted into an encrypted reputation weight using a homomorphic encryption public key. The mathematical expression for the preset mapping function is: ; in, Privacy budget spending, This represents the average privacy budget expenditure. This is the upper limit of the preset reputation weight; The gradient secret share and encrypted reputation weight constitute the data to be uploaded to the cloud.

[0011] Preferably, in step S3, the process of updating the global nutrient regression model specifically includes: The aggregation node has a pre-set medical knowledge graph in a homomorphic encryption state, and the medical knowledge graph defines a ciphertext constraint matrix of positive and negative medical correlations between specific dietary characteristics and health indicators. The aggregation node receives data to be uploaded to the cloud from various smart terminals, and constructs a dense-state gating coefficient matrix for each smart terminal that is consistent with the gradient secret share dimension. The process of generating the dense-state gating coefficient matrix specifically includes: The pre-generated multiplication triples are used as auxiliary calculation parameters to perform element-level secure multi-square multiplication on the ciphertext constraint matrix and gradient secret share to obtain the secret state verification matrix. Using the symbol extraction protocol in secure multi-party computation, the symbol features of each corresponding element in the closed-state parity check matrix are extracted, and a closed-state gating coefficient matrix based on the symbol features is constructed. The dimension of the closed-state gating coefficient matrix is ​​consistent with that of the closed-state parity check matrix. When the sign characteristic of an element in the dense state parity check matrix is ​​negative, the value of the corresponding element in the dense state gating coefficient matrix is ​​0; When the sign characteristic of an element is determined to be positive or zero, the value of the corresponding element in the gating coefficient matrix is ​​1; Homomorphic dot multiplication is performed on the pure encrypted gradient matrix and the encrypted gating coefficient matrix of each smart terminal. The results of the homomorphic dot multiplication are then weighted and aggregated with the encrypted reputation weights of each smart terminal. The aggregated results are then decrypted to obtain the plaintext global gradient matrix that conforms to the medical prior distribution. The dietary weight parameter matrix of the local nutrition regression model on the smart terminal is updated using the plaintext global gradient to obtain the global dietary weight matrix. The mathematical expression for the update is: ; in, As a learning factor, This represents the global gradient of the plaintext.

[0012] Preferably, the process of obtaining the multiplication triple specifically includes: During the offline phase, each smart terminal generates a massive number of multiplication triples, randomly selects half of the triples as the cut-out set, publishes their values, and verifies them. If all triples in the cut set satisfy the multiplication relation, the verified hash value is written to the blockchain as a computing access credential, and the remaining undisclosed triples are used as a selection set for the dense multiplication operation in the online phase. If any triple in the cut set does not satisfy the multiplication relationship, the verification is deemed to have failed, and an exception handling operation is executed. The exception handling operation includes a circuit breaker operation and a protocol reset operation in sequence.

[0013] Preferably, the circuit breaker operation includes: immediately clearing all pre-generated triplet data of the current batch from memory and broadcasting a stop instruction to all smart terminals to prevent erroneous triplets from being used in subsequent calculations; The protocol reset operation includes: after confirming that the above steps are completed, distributing a new random number seed, triggering a new round of triplet preprocessing protocol, until a triplet that meets the verification conditions is generated.

[0014] Preferably, step S4 specifically includes: Extract the indices of elements with non-zero weights in the global diet weight matrix and map these indices back to specific diet features. Dietary characteristics are ranked according to the absolute value of non-zero weights to generate a nutritional attribution map that reflects the key factors of current population health risk. If the analysis results show that the weight of dietary features is zero, an instruction is generated to reduce the sampling frequency of the dietary features by the smart terminal. If the analysis results show that a certain type of dietary feature has a high weight, an instruction is generated to increase the sampling frequency of the dietary feature on the smart terminal.

[0015] The beneficial effects of this invention are as follows: This invention utilizes the high entropy characteristics of the micro-tremor signals in a user's hands during eating, combined with timestamps, to generate a one-time seed. This seed effectively defends against simulated data attacks and replay attacks by leveraging the non-replicability of biological characteristics, ensuring that the data used in the calculation originates from real live eating behavior. The logical gating mechanism of this invention can automatically identify and eliminate abnormal data that violates medical common sense without decrypting the data, ensuring that the final updated nutrition regression model conforms to the medical prior distribution. This solves the problem of model divergence or distortion caused by uneven data quality in distributed learning. This invention introduces blockchain as a trust anchor, linking the upload volume of non-zero gradients to the on-chain privacy budget, transforming abstract privacy protection into a concrete economic cost constraint, forcing the model to retain only high-value information for interaction. Attached Figure Description

[0016] Figure 1 The present invention provides a flowchart of a method for joint analysis of cross-institutional nutritional data based on secure multi-party computation, as an embodiment of the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example, refer to Figure 1 This paper presents a method for joint analysis of cross-institutional nutrition data based on secure multi-party computation, comprising the following steps: Step S1: Collect raw data of dietary behavior, user health indicators and eating biosignals from the smart terminal, and preprocess the raw data of dietary behavior, user health indicators and eating biosignals. Step S2: Construct a nutrition regression model based on the original data of dietary behavior and the user's health indicators, calculate the sparse gradient matrix, deduct the on-chain privacy budget based on the sparse matrix, generate an encrypted seed based on the eating biological signal, encrypt the sparse gradient using the encrypted seed, and generate data to be uploaded to the cloud. Step S3: Use secure multi-party computation to perform causal consistency verification on the data to be uploaded to the cloud, and aggregate the data to be uploaded to the cloud that has passed the causal consistency verification to update the global nutrient regression model. Step S4: Construct a nutrient attribution map using the updated global nutrient regression model and generate feedback instructions to adaptively adjust the working mode of the smart terminal.

[0019] This invention generates encrypted seeds using biological signals from eating on a smart terminal, combines sparse gradients and blockchain privacy budgets at the transmission end, and performs causal consistency verification based on a medical knowledge graph in the cloud. This invention differs from traditional federated learning, which focuses only on the statistical characteristics of data while ignoring medical logic and the authenticity of data sources. It can effectively defend against simulated data attacks and model poisoning attacks while protecting user privacy data from leakage, ensuring that the final global nutritional regression model has both high accuracy and clinical interpretability.

[0020] The raw data on eating behavior from smart terminals includes: food appearance image data, eating speed data, and food nutrient content data; Health indicators include systolic blood pressure, diastolic blood pressure, and blood sugar; Feeding biosignals include hand tremor signals; Preprocessing includes normalization and noise reduction; The raw data of dietary behavior is converted into dietary feature vectors.

[0021] In one specific embodiment of the present invention, the smart terminal is a smart dining terminal (smart tableware, such as smart forks and spoons) held by the user. In the architecture of this embodiment, in addition to the smart terminal (data sender) and the aggregation node (main computing party), a computing collaboration node (auxiliary computing party) is also introduced. The aggregation node and the computing collaboration node are controlled by different institutions (the aggregation node is a cloud service provider, and the computing collaboration node is a blockchain alliance node). The two do not trust each other and do not collude. Both parties hold a secret share of the pre-generated multiplication triple and jointly complete the logical verification and model update of the encrypted data through a secure multi-party computation protocol. Neither party can independently decrypt and obtain the user's original gradient. Data collection is accomplished through smart tableware deployed at the user's end. The smart tableware collects raw data on eating behavior and converts it into dietary feature vectors. , It is A column vector, where each element is a column vector. Representing a specific feature value, in this embodiment, the dimension of the dietary feature vector is... dimension; in to The method involves capturing food images using the front-end camera of smart tableware, processing them through a lightweight ResNet-18 network, and extracting the output of the penultimate fully connected layer as a 256-dimensional texture feature. Represents the average eating speed (unit: grams per second). to Represents the content of trace elements; For the original feature vector Perform augmentation, that is, in Append a dimension whose value is always 1 to the end. .

[0022] Biodynamic characteristics include spectral data of hand micro-tremors during feeding (hand micro-tremor signal) collected by a high-frequency inertial measurement unit. The hand micro-tremor signal refers to six-axis motion data (including triaxial acceleration and triaxial angular velocity) collected by a high-sensitivity inertial measurement unit at a sampling rate of 100 Hz. The active motion components of 0.1 to 3 Hz are filtered out by a bandpass filter, and the physiological tremor components in the frequency band of 4 to 12 Hz are extracted. In the time domain, this component is a high-frequency shaking sequence superimposed on the feeding trajectory, and in the frequency domain, it is a power spectral density distribution with individual-specific peaks. After a preset time period (2 hours in this example) following the end of the meal, the user's smart bracelet collects health indicators at this time. The augmented dietary feature vector is then paired with the health indicators, which are normalized according to the average level of healthy people to eliminate the influence of dimensions.

[0023] Step S2, the process of obtaining the sparse gradient matrix specifically includes: A nutrient regression model is constructed, which is in the form of a linear mapping function. The mathematical expression of the nutrient regression model is: ; in, This is an augmented feature vector of dietary behavior. This is a matrix of dietary weight parameters corresponding to the feature dimensions of the original dietary behavior data and the feature dimensions of the health data. A vector composed of health indicators; used to characterize the degree of influence of each dietary characteristic on the health of the population. Before the first round of calculation begins, all elements of the diet weight parameter matrix are initialized to preset initial values. In subsequent rounds of calculation, the parameter matrix of the global nutrition regression model updated in the previous round is used as the current diet weight parameter matrix. The process of calculating the original gradient matrix of the diet weight parameter matrix in the nutrition regression model includes: Based on the sparse random variance reduction gradient method, combined with Regularization constraint terms are applied, and the original gradient matrix is ​​calculated. The mathematical expression for the original gradient matrix is: ; in, Based on raw data of dietary behavior, health indicators, and current parameters The calculated stochastic gradient, For the snapshot parameter matrix, For snapshot parameter matrix The calculated reference gradient, The mean of the full gradients is based on the snapshot parameter matrix. Preset Regularization coefficient; The original gradient matrix is ​​filtered according to a preset retention ratio to obtain a sparse gradient matrix. The filtering operation specifically includes: using the dietary weight parameter matrix of the nutrition regression model to perform linear prediction on the dietary behavior feature vector in the original dietary behavior data; comparing the results of the linear prediction with health indicators; and calculating the fitting residual vector; the absolute value of the fitting residual is used as a control signal to measure the value of the data. When the magnitude of the fitted residual vector is greater than or equal to the preset residual threshold (1.0 in this embodiment), the preset retention ratio is increased; this indicates that the original dietary behavior data contains important information that the nutritional regression model has not learned. When the magnitude of the fitted residual vector is less than the preset residual threshold, the preset retention ratio is reduced; it is determined that the current data has been fully fitted by the model. The original gradient matrix is ​​truncated according to the adjusted retention ratio. The top few gradients with the largest absolute values ​​are retained, and the remaining gradients are set to zero to obtain a sparse gradient matrix.

[0024] In one specific embodiment of the present invention, the constructed nutrient regression model is in the form of a multiple linear mapping function, and its mathematical expression is: ; in, This is an augmented feature vector of dietary behavior. The diet weight parameter matrix is ​​a matrix corresponding to the feature dimensions and health indicator dimensions of the original diet behavior data. Before the first iteration begins, the diet weight parameter matrix is ​​initialized to a preset initial value (in this embodiment, all elements in the matrix are initialized to all zero values). Dietary weight parameter matrix for 3D matrix ( To augment the feature dimensions of the original data, (For the dimensions of health indicators), the rows of the diet weight parameter matrix correspond to dietary behavior characteristics, and the columns correspond to health indicators. The elements of the diet weight parameter matrix... Characterizing the first Dietary characteristics of vitamins on the first Linear regression weighting coefficients of health indicators; Use the current dietary feature vector and health indicators Substitute into the nutrient regression model and perform matrix multiplication. Calculate the current predicted health indicators ,Will and The difference is used as the fitting residual vector Each component of the fitted residual vector represents the prediction bias of the nutrition regression model on the corresponding health indicator. To establish an accurate and sparse mapping between health indicators and dietary behavior characteristics, a composite objective function is constructed. Its mathematical expression is as follows: ; in, Let be the objective function. For the number of dietary samples, For the first A vector of health indicators for a dietary sample This is the transpose of the weight matrix. For the first Augmented feature vectors of a diet sample for Regularization coefficient, Let L1 be the norm of the weight matrix; The least squares loss function is used to characterize the effect of the nutrient regression model on the th... A real-time sensing diet sample The fitting residuals; For the diet weight parameter matrix The norm, by minimizing this term, utilizes the geometric properties of Lasso regression to compress the weights of non-critical features to zero, thereby achieving sparse feature selection in the nutrient regression model; To address the issues of large variance and unstable convergence in traditional stochastic gradient descent, this embodiment introduces a two-layer loop update mechanism with variance reduction techniques. The outer loop is triggered after each certain period of calculation, updating the outer loop with the latest dietary weight parameter matrix. Perform a full backup, denoted as a snapshot parameter matrix. And calculate the arithmetic mean of the gradients of the objective function for all historical diet samples currently cached locally, denoted as . ; When new real-time sensing data dietary samples Upon input, the inner loop is entered. The inner loop calculates the original gradient matrix for the diet weight parameter matrix. The mathematical expression for the original gradient matrix is: ; in, Based on the current diet sample and the current diet weight parameter matrix The calculated stochastic gradient, For snapshot parameter matrix The reference gradient calculated for the current diet sample. The mean of the full gradient based on the snapshot parameters. Preset Regularization coefficient, The sign function is used in this formula to obtain the unbiased estimated gradient through the variance reduction principle and to promote sparsity through the L1 regularization term. The current dietary sample includes raw data on dietary behavior from smart terminals and users' health indicators; The method for calculating the stochastic gradient is to take the current diet sample's... and the current diet weight parameter matrix Multiply to obtain the current predicted health indicator vector, and The residual vector of the current predicted health indicator vector and Multiplication yields stochastic gradients; The reference gradient is calculated by taking the current diet sample's gradient as an example. and snapshot parameter matrix Multiply to obtain the current predicted health indicator vector, and The residual vector of the current predicted health indicator vector and Multiplication yields the reference gradient; This invention employs a sparse stochastic variance reduction gradient algorithm combined with L1 regularization constraints. By introducing a snapshot parameter matrix and the mean of the full gradient, variance reduction is achieved, effectively suppressing noise fluctuations in stochastic gradient descent and ensuring stable convergence of the model under a non-convex objective function. Simultaneously, the gradient retention ratio is dynamically adjusted using the fitting residuals, and the gradient matrix is ​​truncated, forcing the model to retain only the key features that have the greatest impact on health indicators. This sparsity processing not only significantly reduces the network transmission bandwidth requirements but also reduces the rate at which the privacy budget is consumed by reducing the number of uploaded non-zero elements, achieving a balance between computational efficiency and economic cost.

[0025] In step S2, the operation of deducting the on-chain privacy budget specifically includes: The number of non-zero elements in the sparse gradient matrix is ​​counted. Based on the preset privacy unit price, the privacy budget consumption is calculated. The privacy budget consumption is the product of the preset privacy unit price and the number of non-zero elements in the sparse gradient matrix. The smart terminal sends a verification request to the blockchain smart contract, and the smart contract reads the current privacy budget balance of the smart terminal's account on the blockchain; If the privacy budget consumption exceeds the difference between the smart terminal's privacy budget balance on the blockchain and the preset balance threshold, the verification fails, the smart contract refuses to issue the calculation access certificate, and the upload is terminated. If the privacy budget consumption is less than or equal to the difference between the privacy budget balance on the blockchain and the preset balance threshold, the verification is deemed successful. The smart contract then performs a deduction operation on the privacy budget balance according to the privacy budget consumption, updates the privacy budget balance, and generates a timestamped access certificate to be returned to the smart terminal.

[0026] In one specific embodiment of the present invention, blockchain is introduced as a trust anchor and accounting tool for third parties; When a smart terminal (i.e., smart tableware) is initialized, it registers an account on the blockchain. This account is initially allocated a certain amount of privacy budget balance. In this embodiment, the initial privacy budget balance is 1000 privacy points. The privacy budget amount represents the total amount of information entropy that the smart terminal is allowed to release to the outside world. The preset unit privacy price is set to 1 integral per non-zero gradient element; The balance threshold is set to 50 privacy points. Since blockchain transactions (uploading hashes and verifying credentials) consume a small amount of on-chain resources, a balance threshold needs to be maintained to prevent the account from running out of resources and thus being unable to send control signals to request replenishment. After obtaining the sparse gradient matrix, instead of immediately transmitting it encrypted, the privacy budget is deducted first to calculate the amount of privacy budget consumed. The smart terminal, as a light node in the blockchain network, initiates a query request to the blockchain network through the light node protocol, calls the smart contract interface deployed on the chain, and reads the current privacy budget balance of the account. The existence of privacy deduction mechanisms transforms abstract privacy protection into concrete economic costs. The sparsity of nutrient regression models is no longer just for fast computation, but also for saving budget. If the model cannot effectively reset the weights of a large number of irrelevant features to 0, too many non-zero gradients will be generated, and users will quickly exhaust their budget and be unable to continue participating in the service. This requires nutrient regression models to maintain high sparsity and attribution accuracy at all times.

[0027] This invention introduces blockchain as a trust anchor and privacy measurement tool, transforming the abstract differential privacy protection into a concrete on-chain economic cost. By linking the privacy budget consumption to the number of non-zero elements in the sparse gradient matrix and using smart contracts for balance verification and deduction, this mechanism forces terminal devices to consume budget for uploading only when the gradient contains high-value information (i.e., a large fitting residual). This effectively prevents malicious nodes from probing global model privacy by sending invalid gradients at high frequencies, and also provides an immutable on-chain certificate for measuring cross-institutional data contributions.

[0028] Step S2, the process of generating data to be uploaded to the cloud, specifically includes: Calculate the Shannon entropy of the hand microtremor signal, concatenate the Shannon entropy with the timestamp, and generate a biomask seed using a hash function. A pseudo-random number generator driven by a biological mask seed is used to generate a random mask matrix with the same dimension as the sparse gradient matrix. Based on a random mask matrix, an additive secret sharing algorithm is used to split the sparse gradient matrix and generate gradient secret shares that mask the original gradient values.

[0029] In one specific embodiment of the present invention, the hand tremor signal collected by the smart terminal is intercepted, and a high-pass filter with a cutoff frequency of 3Hz is applied to remove the gravity component and the active motion component. Subsequently, a fast Fourier transform is performed on the remaining high-frequency tremor signal to calculate its power spectral density in the physiological tremor frequency band of 4Hz to 12Hz, denoted as . ; To transform physical characteristics into information-theoretic characteristics, the power spectral density is normalized to construct a probability distribution of tremor energy. Based on this probability distribution, using Shannon's formula The biodynamic entropy value was calculated. This entropy value precisely quantifies the microscopic instability of the user's neuromuscular control system and has extremely high individual specificity. The Shannon entropy in floating-point form is concatenated with the current timestamp, and a long integer value is generated using a hash function (SHA-256). This long integer is the biomass seed. Using this calculated biomass seed as the initialization seed, the pseudo-random number generator is initialized. The dimension of the current sparse gradient matrix is ​​read and set to 1. (This embodiment is) A mask matrix is ​​constructed by generating a sequence of random numbers with the same dimension as the sparse gradient matrix using a pseudo-random number generator. Based on the random mask matrix, the sparse gradient matrix is ​​split using the additive secret sharing algorithm to generate gradient secret shares that mask the original gradient values. The gradient secret shares contain share 1 and share 2. Share 1 is the random mask matrix, and share 2 is the result of subtracting the random mask matrix from the sparse gradient.

[0030] Using the homomorphic encryption public key generated locally on the smart terminal, the random mask matrix is ​​homomorphically encrypted to generate a secret mask matrix. The secret mask matrix and the gradient secret share are then sent to the aggregation node.

[0031] This invention utilizes Shannon entropy from hand micro-tremors with time-varying characteristics to generate a one-time session seed and decomposes the original gradient into a random mask matrix and a masked share, ensuring the mathematical randomness of network transmission data. In particular, the seed is independently protected by the public key of the aggregation node through a key encapsulation mechanism, so that even if the transmission link is compromised, the attacker cannot recover the gradient because they cannot obtain the seed. At the same time, the mechanism that the aggregation node must decrypt the seed to reconstruct the mask enforces implicit verification of data timeliness and biological activity, greatly enhancing the system's anti-attack capability.

[0032] Step S2, the process of generating data to be uploaded to the cloud, also includes: Based on the privacy budget consumption, combined with a pre-defined mapping function, the reputation weight is obtained, and the reputation weight is converted into an encrypted reputation weight using a homomorphic encryption public key. The mathematical expression for the predefined mapping function is: ; in, Privacy budget spending, This represents the average privacy budget expenditure. This is the upper limit of the preset reputation weight. For mapping functions; The gradient secret share and encrypted reputation weight constitute the data to be uploaded to the cloud.

[0033] In one specific embodiment of the present invention, the upper limit of the preset reputation weight is set to 2.0. This mechanism ensures that the magnitude of model updates is positively correlated with the information density (budget cost) of the data, while also possessing robustness against malicious high-budget attacks.

[0034] This invention transforms the user's privacy investment cost (budget consumption) into reputation weights during model aggregation and sets a weight cap. This incentivizes participating institutions (smart terminals) to contribute data with rich information content while preventing high-budget nodes from maliciously dominating the global model. By converting reputation weights into encrypted form for aggregation, the importance distribution of each participating institution is further hidden, preventing targeted attacks on high-weight nodes and improving the robustness of the federated network.

[0035] Step S3, the process of updating the global nutrient regression model specifically includes: The aggregation node is pre-configured with a medical knowledge graph in a homomorphic encryption state. The medical knowledge graph defines a ciphertext constraint matrix of positive and negative medical correlations between specific dietary characteristics and health indicators. The aggregation node receives data to be uploaded to the cloud from various smart terminals, and constructs a dense-state gating coefficient matrix for each smart terminal that is consistent with the gradient secret share dimension. The process of constructing the dense-state gating coefficient matrix specifically includes: The pre-generated multiplication triples are used as auxiliary calculation parameters. The aggregation node and the computational collaboration node jointly perform element-level secure multi-party multiplication on the ciphertext constraint matrix and the gradient secret share to obtain the secret state verification matrix. Specifically, the aggregation node and the computational collaboration node exchange intermediate calculation results based on the Beaver multiplication triple protocol without exchanging the original ciphertext share, thereby calculating the share of the secret state verification matrix. By utilizing the symbol extraction protocol in secure multi-party computation, the symbol features of each corresponding element in the closed-state parity check matrix are extracted, and a closed-state gating coefficient matrix based on the symbol features is constructed. The dimension of the closed-state gating coefficient matrix is ​​consistent with that of the closed-state parity check matrix. The sign characteristic is the positive or negative mathematical attribute of each element in the dense state parity check matrix; When the sign characteristic of an element in the dense state parity check matrix is ​​negative, the value of the corresponding element in the dense state gating coefficient matrix is ​​0; When the sign characteristic of an element is determined to be positive or zero, the value of the corresponding element in the gating coefficient matrix is ​​1; In step S3, before calling the pre-generated multiplication triples as auxiliary calculation parameters, in this embodiment, after receiving the data, the aggregation node does not perform any decryption operation. It utilizes the homomorphic addition property of the homomorphic encryption algorithm to directly perform homomorphic addition on the received encrypted mask matrix and the plaintext form of the gradient secret share. The aggregation node directly restores the pure encrypted gradient matrix in the ciphertext domain. During this process, the aggregation node cannot touch the plaintext of the mask matrix, thus it cannot reverse-engineer the original gradient.

[0036] Homomorphic dot multiplication is performed on the gradient secret share of each smart terminal and the encrypted gating coefficient matrix, and the homomorphic dot multiplication result is weighted and aggregated with the encrypted reputation weight corresponding to each smart terminal. The aggregation result is then decrypted to obtain the plaintext global gradient matrix that conforms to the medical prior distribution. The dietary weight parameter matrix of the local nutrition regression model on the smart terminal is updated using the plaintext global gradient to obtain the global dietary weight matrix. The mathematical expression for the update is: ; in, This is the weight parameter matrix of the local nutrition regression model on the smart terminal before the update. As a learning factor, For plaintext global gradient; In one specific embodiment of the present invention, a preset significance level threshold is used. The preset correlation threshold is 0.05. It is 0.2; By collecting medical statistical data from the Cochrane Evidence-Based Medicine Database using big data, and for the 1024 dietary characteristics and health indicators involved in this embodiment, statistical indicators derived from large-scale meta-analysis were automatically retrieved and extracted. The extracted statistical indicators include the Pearson correlation coefficient (denoted as Pearson correlation coefficient), which measures the strength of linear correlation between variables. ) and the probability value that measures statistical significance (denoted as If the meta-analysis shows a correlation coefficient between dietary characteristics (such as sodium) and health indicators... and Then set the element at the position corresponding to that dietary characteristic. ; Dietary characteristics are derived from the dietary features represented in the dietary feature vector, including eating speed data and the content of food nutrients; If the correlation coefficient and Then set the element at the position corresponding to that dietary characteristic. ; like or Then set the element at the position corresponding to that dietary characteristic. This indicates that the region is undefined, allowing the nutrient regression model to learn freely. During the initialization phase, the prior constraint matrix is ​​fully encrypted using the homomorphic encryption public key of the aggregation node, and transformed into a ciphertext constraint matrix. In terms of form, the prior constraint matrix is ​​represented as a high-dimensional hexadecimal string array. Construct a dense-state gating coefficient matrix consistent with the gradient secret share dimension; Element-wise homomorphic Hadamard product operations are performed on the ciphertext constraint matrix and the gradient secret share to obtain the secret state verification matrix; Let the elements in the dense state parity check matrix be To determine whether an element violates medical common sense, the aggregation node runs the MPC comparison protocol to homomorphically extract the most significant bit of the plaintext value corresponding to the element's ciphertext, denoted as . ; like The corresponding mathematical meaning is negative (which violates medical common sense), so the protocol outputs... ,like The corresponding mathematical meaning is positive or zero (consistent with medical common sense), then the protocol output is... ; The most significant bit of the ciphertext needs to be checked. Perform a logical reversal. Since homomorphic encryption supports ciphertext subtraction, perform a subtraction operation on the elements in the cryptographic verification matrix. ), and fill the calculation results into the corresponding positions in the dense-state gating coefficient matrix; When elements in the dense-state parity-check matrix represent elements that violate medical common sense, the elements in the dense-state gating coefficient matrix... ,coefficient The corresponding gradient data will be annihilated in subsequent dot products; When consistent with medical common sense, the elements in the dense-state gating coefficient matrix ,coefficient The original gradient data will be fully preserved in subsequent dot products; Repeat the above steps for each element in the dense state check matrix to finally assemble the dense state gate coefficient matrix. This invention first uses homomorphic addition to eliminate biological mask noise in the ciphertext domain to obtain a pure encrypted gradient share. Then, it uses multiplication triples and a symbol extraction protocol to calculate the Hadamard product of the gradient direction and medical prior constraints. The system can accurately identify and remove abnormal gradients that violate medical common sense (such as the negative correlation between sodium intake and blood pressure) without decrypting user data.

[0037] The process of obtaining a multiplication triple includes the following: During the offline phase, each smart terminal generates a massive number of multiplication triples, randomly selects half of the triples as the cut-out set, publishes their values, and verifies them. If all triples in the cut set satisfy the multiplication relation, the verified hash value is written to the blockchain as a computing access credential, and the remaining undisclosed triples are used as a selection set for the dense multiplication operation in the online phase. If any triple in the cut set does not satisfy the multiplication relationship, the verification is deemed to have failed, and an exception handling operation is executed. The exception handling operation includes a circuit breaker operation and a protocol reset operation in sequence.

[0038] This invention randomly samples half of the triples and publicly verifies their multiplication relationships, and stores the verified hash values ​​on the blockchain for evidence. This can detect and eliminate erroneous triples generated by malicious nodes with a very high probability. This mechanism prevents attackers from destroying the homomorphic multiplication results in step S3 by providing tampered auxiliary calculation parameters, and ensures the mathematical correctness of the encrypted verification logic and the security of the execution environment.

[0039] The circuit breaker operation includes: immediately clearing all pre-generated triplet data of the current batch from memory and broadcasting a stop instruction to all smart terminals to stop uploading cloud data to prevent erroneous triplets from being used in subsequent calculations; The protocol reset operation includes: after confirming that the above steps are completed, distributing a new random number seed, triggering a new round of triplet preprocessing protocol, until triplets that meet the verification conditions are generated.

[0040] The immediate circuit breaker strategy can block the path of error propagation. Combined with the redistribution of random number seeds and protocol reset, it ensures that the system can quickly recover to a safe and trustworthy initial state after being attacked or experiencing a failure, thus guaranteeing the continuity and reliability of the joint analysis task.

[0041] Step S4 specifically includes: Extract the indices of elements with non-zero weights in the global diet weight matrix and map these indices back to specific diet features. Dietary characteristics are ranked according to the absolute value of non-zero weights to generate a nutritional attribution map that reflects the key factors of current population health risk. If the analysis results show that the weight of dietary features is zero, then an instruction is generated to reduce the sampling frequency of dietary features by the smart terminal. If the analysis results show that a certain type of dietary feature has a high weight, then an instruction is generated to increase the sampling frequency of dietary features on the smart terminal.

[0042] This invention generates a nutritional attribution map by analyzing a global dietary weight matrix, which can accurately identify key influencing factors of current population health risks and automatically adjust the sampling frequency of smart terminals accordingly. It performs high-frequency and precise monitoring of high-risk factors and reduces sampling of low-weight factors to save on privacy budgets, significantly improving the energy efficiency of hardware and the targeting of health monitoring.

[0043] This invention achieves dual privacy protection and source authenticity verification during data transmission by constructing a dynamic mask and key encapsulation mechanism based on feeding biological signals. It utilizes the high entropy characteristics of the micro-tremor signals of a user's hand during eating, combined with a timestamp, to generate a one-time seed, which drives a pseudo-random number generator to construct a random mask matrix to secretly split the sparse gradient. This mechanism effectively defends against simulated data attacks and replay attacks by utilizing the non-replicability of biological features, ensuring that the data involved in the calculation originates from real live feeding behavior. Unlike traditional methods that rely solely on statistical aggregation, this invention utilizes a secure multi-party computation protocol before aggregation. It performs Hadamard product operations and sign determination on the encrypted gradient to be uploaded and the pre-set encrypted pathophysiological constraint matrix. This logical gating mechanism can automatically identify and eliminate abnormal data that violates medical common sense without decrypting the data, ensuring that the final updated nutritional regression model conforms to the medical prior distribution. This solves the problem of model divergence or distortion caused by uneven data quality in distributed learning.

[0044] This invention also utilizes L1 regularization constraints to compress non-critical feature weights to zero, significantly reducing the data dimension of encrypted transmission. By introducing blockchain as a trust anchor, the upload volume of non-zero gradients is linked to the on-chain privacy budget, transforming abstract privacy protection into a concrete economic cost constraint. This forces the model to retain only high-value information for interaction. Combined with an adaptive feedback mechanism based on nutrient attribution graphs, the system can dynamically adjust the sampling frequency of terminal sensors according to feature weights, achieving accurate monitoring of key health risk factors while reducing energy consumption.

[0045] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0046] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for joint analysis of cross-institutional nutritional data based on secure multi-party computation, characterized in that, Includes the following steps: Step S1: Collect raw data of dietary behavior, user health indicators and eating biosignals from the smart terminal, and preprocess the raw data of dietary behavior, user health indicators and eating biosignals. Step S2: Construct a nutrition regression model based on the original dietary behavior data and health indicators, calculate the sparse gradient matrix, deduct the on-chain privacy budget based on the sparseness matrix, generate an encryption seed based on the eating biosignal, encrypt the sparse gradient using the encryption seed, and generate data to be uploaded to the cloud. Step S3: Use secure multi-party computation to perform causal consistency verification on the data to be uploaded to the cloud, and aggregate the data to be uploaded to the cloud that has passed the causal consistency verification to update the global nutrient regression model. Step S4: Construct a nutrient attribution map using the updated global nutrient regression model and generate feedback instructions to adaptively adjust the working mode of the smart terminal.

2. The method for joint analysis of cross-institutional nutritional data based on secure multi-party computation as described in claim 1, characterized in that, The raw data of the smart terminal's eating behavior includes: food appearance image data, eating speed data, and food nutrient content data; The health indicators include systolic blood pressure, diastolic blood pressure, and blood glucose; The feeding biosignals include hand tremor signals; Preprocessing includes normalization and noise reduction; The raw data of dietary behavior is converted into dietary feature vectors.

3. The method for joint analysis of cross-institutional nutritional data based on secure multi-party computation as described in claim 2, characterized in that, In step S2, the process of obtaining the sparse gradient matrix specifically includes: A nutrient regression model is constructed, which is in the form of a linear mapping function. The mathematical expression of the nutrient regression model is as follows: ； in, This is an augmented feature vector of dietary behavior. This is a matrix of dietary weight parameters corresponding to the feature dimensions of the original dietary behavior data and the feature dimensions of the health data. A vector composed of health indicators; The process of calculating the original gradient matrix of the diet weight parameter matrix in the nutrition regression model specifically includes: Based on the sparse random variance reduction gradient method, combined with Regularization constraint terms are applied, and the original gradient matrix is ​​calculated. The mathematical expression of the original gradient matrix is ​​as follows: ； in, Based on raw data of dietary behavior, health indicators, and current parameters The calculated stochastic gradient, For the snapshot parameter matrix, For snapshot parameter matrix The calculated reference gradient, The mean of the full gradients is based on the snapshot parameter matrix. Preset Regularization coefficient; The original gradient matrix is ​​filtered according to a preset retention ratio to obtain a sparse gradient matrix. The filtering operation specifically includes: using the dietary weight parameter matrix of the nutrition regression model to perform linear prediction on the dietary behavior feature vector in the original dietary behavior data, comparing the linear prediction result with health indicators, and calculating the fitting residual vector. When the magnitude of the fitted residual vector is greater than or equal to the preset residual threshold (1.0 in this embodiment), the preset retention ratio is increased; When the magnitude of the fitted residual vector is less than the preset residual threshold, the preset retention ratio is reduced. The original gradient matrix is ​​truncated according to the adjusted retention ratio. The top few gradients with the largest absolute values ​​are retained, and the remaining gradients are set to zero to obtain a sparse gradient matrix.

4. The method for joint analysis of cross-institutional nutritional data based on secure multi-party computation as described in claim 3, characterized in that, In step S2, the operation of deducting the on-chain privacy budget specifically includes: The number of non-zero elements in the sparse gradient matrix is ​​counted, and the privacy budget consumption is calculated based on the preset unit privacy price. The privacy budget consumption is the product of the preset unit privacy price and the number of non-zero elements in the sparse gradient matrix. The smart terminal sends a verification request to the blockchain smart contract, and the smart contract reads the current privacy budget balance of the smart terminal's account on the blockchain; If the privacy budget consumption exceeds the difference between the smart terminal's privacy budget balance on the blockchain and the preset balance threshold, the verification fails, the smart contract refuses to issue the calculation access certificate, and the upload is terminated. If the privacy budget consumption is less than or equal to the difference between the privacy budget balance on the blockchain and the preset balance threshold, the verification is deemed successful. The smart contract then performs a deduction operation on the privacy budget balance according to the privacy budget consumption, updates the privacy budget balance, and generates a timestamped access certificate to be returned to the smart terminal.

5. The method for joint analysis of cross-institutional nutritional data based on secure multi-party computation as described in claim 4, characterized in that, In step S2, the process of generating the data to be uploaded to the cloud specifically includes: The Shannon entropy of the hand microtremor signal is calculated, and the Shannon entropy is concatenated with the timestamp. A biomass seed is generated by using a hash function. The biomask seed is used to drive a pseudo-random number generator to generate a random mask matrix with the same dimension as the sparse gradient matrix. Based on a random mask matrix, the sparse gradient matrix is ​​split using an additive secret sharing algorithm to generate gradient secret shares that mask the original gradient values.

6. The method for joint analysis of cross-institutional nutritional data based on secure multi-party computation as described in claim 5, characterized in that, In step S2, the process of generating the data to be uploaded to the cloud further includes: Based on the privacy budget consumption, combined with a pre-defined mapping function, the reputation weight is obtained, and the reputation weight is converted into an encrypted reputation weight using a homomorphic encryption public key. The mathematical expression for the preset mapping function is: ； in, Privacy budget spending, This represents the average privacy budget expenditure. This is the upper limit of the preset reputation weight; The gradient secret share and encrypted reputation weight constitute the data to be uploaded to the cloud.

7. The method for joint analysis of cross-institutional nutritional data based on secure multi-party computation as described in claim 6, characterized in that, In step S3, the process of updating the global nutrient regression model specifically includes: The aggregation node has a pre-set medical knowledge graph in a homomorphic encryption state, and the medical knowledge graph defines a ciphertext constraint matrix of positive and negative medical correlations between specific dietary characteristics and health indicators. The aggregation node receives data to be uploaded to the cloud from various smart terminals, and constructs a dense-state gating coefficient matrix for each smart terminal that is consistent with the gradient secret share dimension. The process of generating the dense-state gating coefficient matrix specifically includes: The pre-generated multiplication triples are used as auxiliary calculation parameters to perform element-level secure multi-square multiplication on the ciphertext constraint matrix and gradient secret share to obtain the secret state verification matrix. Using the symbol extraction protocol in secure multi-party computation, the symbol features of each corresponding element in the closed-state parity check matrix are extracted, and a closed-state gating coefficient matrix based on the symbol features is constructed. The dimension of the closed-state gating coefficient matrix is ​​consistent with that of the closed-state parity check matrix. When the sign characteristic of an element in the dense state parity check matrix is ​​negative, the value of the corresponding element in the dense state gating coefficient matrix is ​​0; When the sign characteristic of an element is determined to be positive or zero, the value of the corresponding element in the gating coefficient matrix is ​​1; Homomorphic dot multiplication is performed on the pure encrypted gradient matrix and the encrypted gating coefficient matrix of each smart terminal. The results of the homomorphic dot multiplication are then weighted and aggregated with the encrypted reputation weights of each smart terminal. The aggregated results are then decrypted to obtain the plaintext global gradient matrix that conforms to the medical prior distribution. The dietary weight parameter matrix of the local nutrition regression model on the smart terminal is updated using the plaintext global gradient to obtain the global dietary weight matrix. The mathematical expression for the update is: ； in, This is a nutrient weight parameter matrix. As a learning factor, This represents the global gradient of the plaintext.

8. The method for joint analysis of cross-institutional nutritional data based on secure multi-party computation as described in claim 7, characterized in that, The process of obtaining the multiplication triple specifically includes: During the offline phase, each smart terminal generates a massive number of multiplication triples, randomly selects half of the triples as the cut-out set, publishes their values, and verifies them. If all triples in the cut set satisfy the multiplication relation, the verified hash value is written to the blockchain as a computing access credential, and the remaining undisclosed triples are used as a selection set for the dense multiplication operation in the online phase. If any triple in the cut set does not satisfy the multiplication relationship, the verification is deemed to have failed, and an exception handling operation is executed. The exception handling operation includes a circuit breaker operation and a protocol reset operation in sequence.

9. The method for joint analysis of cross-institutional nutritional data based on secure multi-party computation as described in claim 8, characterized in that, The circuit breaker operation includes: immediately clearing all pre-generated triplet data of the current batch from memory and broadcasting an abort instruction to all smart terminals; The protocol reset operation includes: after confirming that the above steps are completed, distributing a new random number seed, triggering a new round of triplet preprocessing protocol, until a triplet that meets the verification conditions is generated.

10. The method for joint analysis of cross-institutional nutritional data based on secure multi-party computation as described in claim 9, characterized in that, Step S4 specifically includes: Extract the indices of elements with non-zero weights in the global diet weight matrix and map these indices back to specific diet features. Dietary characteristics are ranked according to the absolute value of non-zero weights to generate a nutritional attribution map that reflects the key factors of current population health risk. If the analysis results show that the weight of dietary features is zero, an instruction is generated to reduce the sampling frequency of the dietary features by the smart terminal. If the analysis results show that a certain type of dietary feature has a high weight, an instruction is generated to increase the sampling frequency of the dietary feature on the smart terminal.

Images