Trusted AI inquiry model generation method and device, electronic equipment and storage medium

Abstract

The invention discloses a trusted AI inquiry model generation method and device, electronic equipment and a storage medium, which are used for solving the technical problem of how to realize effective sharing and analysis of data while ensuring the privacy of a patient. The method comprises the steps that model fragments of a previous round of AI inquiry global model are received through polynomial interpolation, when medical nodes collect medical data, a local AI inquiry model is trained through the medical data and the model fragments, model parameters are obtained, and zero-knowledge proof is generated; when the zero-knowledge proof verification is passed, reliability verification is carried out on model parameters of the medical nodes in sequence through a three-dimensional reputation evaluation model, and the medical nodes passing the verification are obtained; the model parameters of the medical nodes passing verification are stored in a block, an AI inquiry global model is obtained, and the block is generated through the verification nodes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of diagnosis models, and particularly relates to a trusted AI diagnosis model generation method and device, electronic equipment and a storage medium. BACKGROUND

[0002] Blockchain is a distributed database system that links data blocks through cryptography. It is a "digital trust machine" maintained by multiple parties. Blockchain is composed of blocks and chain links. The block encapsulates the information to be saved and is equipped with a unique hash value. When new data is generated, the new block will be chained with the previous block, and any tampering will destroy the integrity of the entire chain. This design allows data to be synchronized across multiple parties without relying on a central authority, making it impossible to modify and traceable throughout. In the medical industry, applications include but are not limited to: patient medical records can be recorded in a decentralized distributed ledger, ensuring data cannot be tampered with and is traceable throughout; distributed network structure and asymmetric encryption algorithm ensure the security and reliability of drug data, enabling transparency and traceability of the drug supply chain, etc.

[0003] Federated learning is a distributed machine learning technique that allows multiple participants, such as different devices, institutions or servers, to collaboratively train a global model without sharing local data. Data is always kept locally, and only model parameters or intermediate calculation results are transmitted through encryption or secure protocols, thus achieving privacy protection and data security. This method not only protects patient privacy, but also enables cross-institutional data collaboration.

[0004] With the growth of medical data and the increasing demand for privacy protection, how to ensure patient privacy while effectively sharing and analyzing data has become an important research direction. Blockchain technology has great potential in building a trusted data exchange platform due to its decentralization and tamper-proof features. However, traditional blockchain systems face performance bottlenecks when dealing with large-scale data, especially when data privacy needs to be guaranteed.

[0005] The prior art uses consortium chain technology for medical data management. This type of system usually uses smart contracts to control data access permissions and ensures all participants' consistent approval of data through consensus mechanisms. However, most existing blockchain solutions rely on Proof of Work (PoW) or Proof of Stake (PoS) as consensus mechanisms, which can lead to slow transaction processing speed and poor scalability. SUMMARY

[0006] The application provides a trusted AI diagnosis model generation method and device, electronic equipment and a storage medium, and aims to solve the technical problem of how to ensure patient privacy while effectively sharing and analyzing data.

[0007] The application provides a trusted AI diagnosis model generation method, which comprises the following steps:

[0008] The model slices of the previous round of AI diagnosis global model are received through polynomial interpolation, and when medical nodes collect medical data, the local AI diagnosis model is trained through the medical data and the model slices to obtain model parameters and generate zero-knowledge proof.

[0009] When the zero-knowledge proof is verified, the reliability of the model parameters of each medical node is verified in sequence through a three-dimensional reputation evaluation model to obtain a medical node that passes the verification.

[0010] The model parameters of the medical node that passes the verification are stored in a block to obtain an AI diagnosis global model, and the block is generated by a verification node.

[0011] Optionally, when the medical nodes collect medical data, the local AI diagnosis model is trained through the medical data and the model slices to obtain model parameters.

[0012] When the medical nodes collect medical data, the local AI diagnosis model is trained through the medical data and the model slices to obtain initial model parameters.

[0013] The gradient of the loss function of the local AI diagnosis model is calculated.

[0014] The update gradient is calculated using the gradient and a preset injection noise.

[0015] The update model parameters are calculated using the update gradient and the initial model parameters.

[0016] It is determined whether the loss function meets the convergence condition, and if not, the step of calculating the gradient of the loss function of the local AI diagnosis model is returned.

[0017] If yes, the update model parameters are taken as the model parameters of the trained local AI model.

[0018] Optionally, when the zero-knowledge proof is verified, the reliability of the model parameters of each medical node is verified in sequence through a three-dimensional reputation evaluation model to obtain a medical node that passes the verification.

[0019] When the zero-knowledge proof is verified, the reliability of the model parameters of each medical node is verified in sequence through a three-dimensional reputation evaluation model to obtain a suspected abnormal node.

[0020] detecting a Byzantine node in the suspected abnormal node;

[0021] removing the Byzantine node from the medical node to obtain a medical node passed verification.

[0022] Optionally, the step of detecting the Byzantine node in the suspected abnormal node comprises:

[0023] analyzing gradient distribution difference of the suspected abnormal node through KL divergence, and determining a suspicious node according to the gradient distribution difference;

[0024] performing interpretability analysis on the suspicious node to determine a Byzantine node.

[0025] Optionally, the three-dimensional reputation evaluation model is:

[0026]

[0027] wherein, is the credibility of the model parameter, represents the accuracy rate of historical updates of the medical node, is the time sequence mean of the gradient verification pass rate, quantifies the storage and computing power contribution of the node, is a model parameter, and the parameter ensures weight normalization.

[0028] Optionally, it further comprises:

[0029] calculating the contribution degree of each medical node to the AI diagnosis global model through the contribution quantification model, and assigning rewards and penalties to the medical nodes according to the contribution degree.

[0030] The application also provides a trusted AI diagnosis model generation device, comprising:

[0031] a model parameter generation module, configured to receive a model segment of a previous round of AI diagnosis global model through polynomial interpolation, train a local AI diagnosis model through medical data collected by a medical node and the model segment when the medical data is collected, obtain model parameters, and generate zero-knowledge proof;

[0032] a verification module, configured to perform reliability verification on model parameters of each medical node in sequence through a three-dimensional reputation evaluation model when the zero-knowledge proof is passed, and obtain a medical node passed verification;

[0033] an AI diagnosis global model generation module, configured to store the model parameters of the medical node passed verification in a block to obtain an AI diagnosis global model, and the block is generated by a verification node.

[0034] Optionally, the model parameter generation module comprises:

[0035] An initial model parameter generation submodule is configured to, when medical data is collected by the medical node, train a local AI diagnosis model by using the medical data and the model shard, and obtain initial model parameters;

[0036] A gradient calculation submodule is configured to calculate a gradient of a loss function of the local AI diagnosis model;

[0037] An update gradient calculation submodule is configured to calculate an update gradient by using the gradient and a preset injection noise;

[0038] An update model parameter calculation submodule is configured to calculate update model parameters by using the update gradient and the initial model parameters;

[0039] A return submodule is configured to determine whether the loss function meets a convergence condition, and if not, return to the step of calculating the gradient of the loss function of the local AI diagnosis model;

[0040] A model parameter determination submodule is configured to, if yes, take the update model parameters as model parameters of a trained local AI model.

[0041] The application further provides an electronic device, which comprises a processor and a memory:

[0042] The memory is configured to store program code and transmit the program code to the processor;

[0043] The processor is configured to execute the trusted AI diagnosis model generation method according to instructions in the program code.

[0044] The application further provides a computer readable storage medium, which is configured to store program code, and the program code is configured to execute the trusted AI diagnosis model generation method.

[0045] As can be seen from the above technical solutions, the application has the following advantages: the application receives model shards of an AI diagnosis global model of the last round by polynomial interpolation, trains a local AI diagnosis model by using medical data collected by the medical node and the model shards, obtains model parameters, and generates zero-knowledge proof; when the zero-knowledge proof is verified, the model parameters of each medical node are verified for reliability in sequence by a three-dimensional reputation evaluation model, and a medical node that passes the verification is obtained; the model parameters of the medical node that passes the verification are stored in a block, an AI diagnosis global model is obtained, and the block is generated by a verification node. Thus, the effective sharing and analysis of data are realized while the privacy of patients is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0047] Figure 1 A step flow chart of a trusted AI diagnosis model generation method provided by the embodiment of the present application is provided.

[0048] Figure 2 An architecture schematic diagram of a trusted AI diagnosis model generation method provided by the embodiment of the present application is provided.

[0049] Figure 3 A structure block diagram of a trusted AI diagnosis model generation device provided by the embodiment of the present application is provided. DETAILED DESCRIPTION

[0050] The embodiment of the present application provides a trusted AI diagnosis model generation method, device, electronic equipment and storage medium, which is used to solve the technical problem of how to guarantee patient privacy while realizing effective sharing and analysis of data.

[0051] In order to make the purposes, features and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the following described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0052] Please refer to Figure 1 , Figure 1 A step flow chart of a trusted AI diagnosis model generation method provided by the embodiment of the present application is provided.

[0053] The trusted AI diagnosis model generation method provided by the present application comprises:

[0054] Step 101, receiving the model segment of the last round AI diagnosis global model through polynomial interpolation, when the medical node collects medical data, training the local AI diagnosis model through the medical data and the model segment, obtaining the model parameter, and generating the zero-knowledge proof;

[0055] In the embodiment of the present application, the architecture of the whole trusted AI diagnosis model generation method is as follows Figure 2As shown, the downstream medical nodes are composed of medical units, including but not limited to large hospitals, small clinics, and mobile patrol facilities. All nodes share AI models on a decentralized blockchain platform. New nodes can also be added after passing management verification.

[0056] First, at the network topology layer, medical nodes participate in dynamic networking as blockchain full nodes, and 21 verification nodes (including 5 backup nodes) are elected by using a DPoS (Delegated Proof of Stake) consensus mechanism. Any medical institution that wants to become a verification node needs to submit a candidate application and be voted on by all nodes. The nodes are ranked from high to low according to the number of votes, and the top 16 nodes are the formal verification nodes, followed by the next 5 as backup nodes. Through the power of staking and rotation mechanism to prevent power concentration, that is, the verification nodes are supervised by all nodes, and when the verification nodes cannot maintain online and on-time block, the verification nodes can be replaced. The system constructs a double-chain structure of "parameter chain + audit chain", the main chain stores the hash value of the aggregated model, and the side chain records the encryption proof of the node contribution (each node performs training and calculation tasks based on local data to obtain "node contribution", and through homomorphic encryption and zero-knowledge proof executed locally, the training contribution of itself is encrypted and the corresponding proof is generated), and the two are realized through atomic interaction based on ICS standard (Inter-Chain Standard) cross-chain protocol. This design enables the trusted transfer of medical data hash (in the format of HIPAA standard) and model parameters between different chains, eliminating the risk of single point failure at the infrastructure level. For example, when a node cluster in a certain region fails, the cross-chain protocol can automatically migrate the computing task to other geographic shard node groups.

[0057] The core innovation of the secure computing layer is to introduce KZG (Kate-Zaverucha-Goldberg) polynomial commitment technology into federated learning. In federated learning, multiple groups of nodes participate in training together, and the central party aggregates the updated gradients / parameters of each node to obtain the global model after this training. The global model is mathematically expressed as a polynomial function where, represents the model parameters, and the model is divided into fragments and transmitted to different medical nodes through polynomial interpolation. Each node can only obtain a part of the original AI consultation global model, thereby ensuring that the entire content cannot be inferred. When medical data is collected, each medical node can train a local AI consultation model through medical data and model fragments, obtain model parameters, and generate zero-knowledge proof (Zero-Knowledge Scalable Transparent Arguments of Knowledge, zk-STARK) proof , This makes it possible to prove to the audit chain that the node did indeed perform the local training computation without revealing the internal data, where is the cryptographic summary of the computation trace, is the time-bound signature of the polynomial coefficients, is the timestamp to prevent replay attacks. The proof size is compressed to logarithmic level by using the FRI (Fast Reed-Solomon Interactive) protocol , greatly reducing the overhead of on-chain verification.

[0058] zk-STARK is a zero-knowledge proof protocol with transparency (no trusted initialization) and scalability, which can prove the correctness of large-scale computation without revealing privacy.

[0059] In one example, when the medical node collects medical data, the step of training a local AI consultation model by medical data and model shards to obtain model parameters can include the following sub-steps:

[0060] S11, when the medical node collects medical data, training a local AI consultation model by medical data and model shards to obtain initial model parameters;

[0061] S12, calculating the gradient of the loss function of the local AI consultation model;

[0062] S13, calculating the update gradient using the gradient and the preset injected noise;

[0063] S14, calculating the update model parameter using the update gradient and the initial model parameter;

[0064] S15, determining whether the loss function meets the convergence condition, if not, returning to the step of calculating the gradient of the loss function of the local AI consultation model;

[0065] S16, if yes, taking the update model parameter as the model parameter of the trained local AI model.

[0066] When training the model, the medical data is input into the model, the output of the model is calculated through forward propagation, and compared with the true label to calculate the loss function. Then use the optimization algorithm to update the parameters of the model according to the gradient of the loss function, the purpose is to make the loss function gradually decrease. Repeat this process until the model converges.

[0067] In order to ensure the privacy in the gradient update process, noise that satisfies - differential privacy is injected in the gradient update process , where is the privacy loss parameter, which determines the strength of privacy protection; is a slack parameter that allows the algorithm to violate the strict privacy guarantee with a small probability in some cases; the mean of 0 ensures that the expected impact of the noise on the gradient update is zero, which does not produce systematic bias; represents a unit matrix; the noise standard deviation , ensures that the data update of a single node cannot be inferred in reverse. For example, the parameters of the local AI consultation model , the objective function is Each time a batch of data B is selected from the data set to calculate the gradient and update the parameters. Its gradient representation is In order to meet differential privacy, noise is injected into the gradient, that is, At this time, the update formula of the model parameters is:

[0068]

[0069] Among them, represents the learning rate.

[0070] Step 102, when the zero-knowledge proof is verified, the reliability of the model parameters of each medical node is verified in turn through a three-dimensional reputation evaluation model, and the medical node that passes the verification is obtained.

[0071] In the embodiments of the present application, a three-dimensional reputation evaluation model is designed in the consensus verification layer, which is used for reliability verification of each medical node, so as to determine the suspected abnormal node.

[0072] In one example, step 102 can include the following sub-steps:

[0073] S21, when the zero-knowledge proof is verified, the reliability of the model parameters of each medical node is verified in turn through a three-dimensional reputation evaluation model, and the suspected abnormal node is obtained.

[0074] In specific implementation, the three-dimensional reputation evaluation model is:

[0075]

[0076] Among them, is the credibility of the model parameters, represents the historical update accuracy of the medical node, is the time sequence mean of the gradient verification pass rate, quantifies the storage and computing power contribution of the node, that is, the size of the local data of each medical node and how many rounds of local training iterations the medical node has performed for the AI consultation global model, is the model parameter, and the parameter ensures weight normalization.

[0077] By calculating the credibility of each node, suspected abnormal nodes in the medical nodes can be found.

[0078] S22, detecting a Byzantine node in the suspected abnormal node;

[0079] The Byzantine node is a node that may appear arbitrary errors or malicious behavior.

[0080] In the embodiments of the present application, the step of detecting the Byzantine node in the suspected abnormal node can include the following sub-steps:

[0081] S221, analyzing the gradient distribution difference of the suspected abnormal node by KL divergence, and determining the suspicious node according to the gradient distribution difference;

[0082] S222, performing explainability analysis on the suspicious node to determine the Byzantine node.

[0083] The KL divergence analysis is used to measure the difference between two probability distributions.

[0084] In a specific implementation, the detection of the Byzantine node adopts a hybrid verification strategy: first, analyze the gradient distribution difference of the suspected abnormal node by KL divergence (Kullback-Leibler Divergence), and the formula is as follows:

[0085]

[0086] Wherein, represents the real distribution (target), represents the approximate distribution (model), and the threshold for judging the gradient distribution difference of the suspicious node can be determined according to actual requirements, which is not limited in the embodiments of the present application.

[0087] If a suspicious node is detected, it is updated and submitted to the TEE environment (an independent secure computing space created by hardware isolation technology) for LIME (Local Interpretable Model-agnostic Explanations, Local Interpretable Model-agnostic Explanations) explainability analysis, to verify whether the feature importance ranking is consistent with the global model. This double verification mechanism converts subjective experience judgment into quantifiable mathematical verification.

[0088] S23, removing the Byzantine node from the medical node to obtain a verified medical node.

[0089] After the Byzantine node is determined, the Byzantine node can be removed from the medical node to obtain a verified medical node.

[0090] Step 103, store the model parameters of the verified medical node in the block to obtain the AI consultation global model, and the block is generated by the verification node.

[0091] The model parameters of the medical node after verification are stored in the block. The AI consultation global model is obtained. Then the model fragments are transmitted to different medical nodes through polynomial interpolation, the model parameters are exchanged, and the nodes indicate that the model updating is completed through the return signal. Then, the nodes can use the parameters of the global model to test the new local data set, and use the updated model to perform AI consultation.

[0092] Further, in the embodiments of the present application, the contribution of each medical node to the AI consultation global model is calculated by a contribution quantification model, and rewards and penalties are allocated to the medical nodes according to the contribution.

[0093] The incentive clearing layer measures the contribution of the nodes through the contribution quantification model , wherein, is the square of the gradient norm, reflecting the model updating strength of node i, the more significant the update, the greater the gradient, and the stronger the "reform power" to the global model, is the node gradient vector, indicating the gradient vector (obtained by the back propagation algorithm) calculated by the i-th medical node in the local model training on the model parameters w (such as neural network weights, bias), i is the unique identifier of the i-th medical node (such as "Hospital A" and "Clinic B", etc. Participating institutions, taking values (i=1,2,…,n) ) ; is the target node for current contribution calculation, associated with its gradient and the Shapley value ShapleyValue(i), the Shapley value is used to evaluate the marginal contribution, and the denominator realizes global normalization, and j in the denominator is used to traverse the temporary index of all medical nodes. By traversing j, the sum of the gradient norm squares of all nodes is calculated to realize global normalization (so that the contribution degrees of different nodes are comparable). n represents the total number of medical nodes participating in the construction of the AI consultation global model (i.e. how many institutions contribute model parameters). The role of the formula: limit the range of summation (j from 1 to n), reflecting the scale of the collaboration network. Due to the high complexity of the exact calculation of the Shapley value, the system uses the Monte Carlo method for approximate estimation. The smart contract counteracts free-riding behavior by dynamically adjusting the reward weight, for example, giving an additional reward of 12% to nodes with a reputation value higher than 0.8, and imposing an exponentially growing penalty function on nodes that submit updates with a delay , with a penalty increase of 5% per hour of overtime.

[0094] The reward and penalty mechanism is designed to encourage medical nodes to actively participate and ensure efficient operation of the network.

[0095] Specifically, rewards refer to encouraging nodes to complete tasks with high quality. The system measures the contribution of nodes through two indicators: the gradient norm, which reflects the work intensity of the node (for example, the magnitude of parameter updates when training a model), and the Shapley value, which evaluates the "marginal contribution" of the node (similar to the actual role of each person in team cooperation). The combination of the two (the numerator in the formula) determines the contribution ratio of the node.

[0096] Additional rewards: nodes with high reputation values (e.g., reputation score > 0.8) will receive additional rewards (e.g., +12%). When multiple nodes jointly train an AI model, nodes with greater contributions (large update magnitude and key role) will receive more rewards.

[0097] The application receives model slices of the last round of AI diagnosis global model through polynomial interpolation. When medical nodes collect medical data, the local AI diagnosis model is trained through medical data and model slices to obtain model parameters and generate zero-knowledge proof. When the zero-knowledge proof is verified, the reliability of the model parameters of each medical node is verified in turn through a three-dimensional reputation evaluation model to obtain the medical nodes that pass the verification. The model parameters of the medical nodes that pass the verification are stored in a block to obtain an AI diagnosis global model, and the block is generated by the verification node. Thus, while ensuring patient privacy, effective sharing and analysis of data are realized.

[0098] After the delay penalty, GPU power verification and feature watermarking can also be performed. GPU power verification is a physical side measurement that is a verifiable and chainable "energy receipt" that prevents the occurrence of fake gradients, such as "large gradient + low power consumption", which is unreasonable. In addition, it also encourages efficient GPU settings. Feature watermarking is to engrave a small batch of "trigger features" known only to the owner into the model, so that it outputs a special verifiable signal when called, thereby providing auditable semantic evidence for node contribution, model ownership, and on-chain incentives / punishments in blockchain federated learning. In addition, subsequent medical image analysis / electronic medical record modeling can also be performed through the AI diagnosis global model.

[0099] Please refer to Figure 3 , Figure 3 A structure block diagram of a trusted AI diagnosis model generation device provided by an embodiment of the application.

[0100] An embodiment of the application provides a trusted AI diagnosis model generation device, which comprises:

[0101] The model parameter generation module 301 is configured to receive model slices of the last round of AI diagnosis global model through polynomial interpolation, train a local AI diagnosis model through medical data and the model slices when medical nodes collect medical data, obtain model parameters, and generate zero-knowledge proof.

[0102] The verification module 302 is configured to, when the zero-knowledge proof is verified, perform reliability verification on the model parameters of each medical node in sequence through a three-dimensional reputation evaluation model, and obtain a medical node that passes the verification.

[0103] The AI inquiry global model generation module 303 is configured to store the model parameters of the medical node that passes the verification in a block, and obtain an AI inquiry global model, and the block is generated through a verification node.

[0104] In the embodiment of the application, the model parameter generation module 301 comprises:

[0105] The initial model parameter generation submodule is configured to, when the medical node collects medical data, train a local AI inquiry model by using the medical data and model fragments, and obtain initial model parameters.

[0106] The gradient calculation submodule is configured to calculate the gradient of the loss function of the local AI inquiry model.

[0107] The update gradient calculation submodule is configured to calculate an update gradient by using the gradient and a preset injection noise.

[0108] The update model parameter calculation submodule is configured to calculate an update model parameter by using the update gradient and the initial model parameter.

[0109] The return submodule is configured to determine whether the loss function meets a convergence condition, and if not, return to the step of calculating the gradient of the loss function of the local AI inquiry model.

[0110] The model parameter determination submodule is configured to, if yes, take the update model parameter as the model parameter of the trained local AI model.

[0111] In the embodiment of the application, the verification module 302 comprises:

[0112] The reliability verification submodule is configured to, when the zero-knowledge proof is verified, perform reliability verification on the model parameters of each medical node in sequence through a three-dimensional reputation evaluation model, and obtain a suspected abnormal node.

[0113] The Byzantine node detection submodule is configured to detect a Byzantine node in the suspected abnormal node.

[0114] The medical node that passes the verification determination submodule is configured to remove the Byzantine node from the medical node, and obtain a medical node that passes the verification.

[0115] In the embodiment of the application, the Byzantine node detection submodule comprises:

[0116] The suspicious node determination unit is configured to analyze the gradient distribution difference of the suspected abnormal node through KL divergence, and determine a suspicious node according to the gradient distribution difference.

[0117] The Byzantine node determination unit is configured to perform interpretability analysis on the suspicious node to determine the Byzantine node.

[0118] In the embodiment of the present application, the three-dimensional reputation evaluation model is:

[0119]

[0120] wherein, is the credibility of the model parameters, represents the accuracy rate of the historical update of the medical node, is the time sequence average of the gradient verification pass rate, quantifies the storage and computing power contribution of the node, is the model parameter, and the parameter ensures the weight normalization.

[0121] In the embodiment of the present application, further comprising:

[0122] The reward and punishment distribution module is configured to calculate the contribution degree of each medical node to the AI diagnosis global model through the contribution quantification model, and distribute rewards and punishments to the medical nodes according to the contribution degree.

[0123] The embodiment of the present application also provides an electronic device, which comprises a processor and a memory:

[0124] The memory is configured to store program code and transmit the program code to the processor.

[0125] The processor is configured to execute the trusted AI diagnosis model generation method according to the instructions in the program code.

[0126] The embodiment of the present application also provides a computer readable storage medium, which is configured to store program code, and the program code is configured to execute the trusted AI diagnosis model generation method.

[0127] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0128] Each embodiment in the specification adopts a progressive manner for description, and each embodiment focuses on the difference from other embodiments, and the same and similar parts between each embodiment can be referred to.

[0129] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, apparatus, or computer program product. Accordingly, embodiments of the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, embodiments of the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, and the like) embodying computer program instructions.

[0130] Embodiments of the present application are described herein with reference to the drawings, which are as follows: Figure 1 Figure 1 The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks.

[0131] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 Figure 1 The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks.

[0132] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 Figure 1 The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks.

[0133] While preferred embodiments of the present application have been described, additional variations and modifications can be made to these embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to cover all such modifications and variations as fall within the scope of the present application.

[0134] ​​​It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant national and regional laws, regulations and standards, and provide corresponding operation portal for user to choose authorization or refusal.

[0135] Finally, it should be noted that in this document, relational terms such as first and second and the like can merely be used to distinguish one entity or action from another, without necessarily requiring or implying that any such entity or action is in fact prior or subsequent in time to the other. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0136] The above-described and above-mentioned embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for generating a reliable AI-based medical consultation model, characterized in that, The method comprises the steps of: receiving model fragments of a previous round of AI diagnosis global model through polynomial interpolation, training a local AI diagnosis model through medical data collected by a medical node and the model fragments, obtaining model parameters, and generating zero-knowledge proof; when the zero-knowledge proof is verified, performing reliability verification on the model parameters of each medical node in turn through a three-dimensional reputation evaluation model to obtain a medical node that passes the verification; storing the model parameters of the medical node that passes the verification in a block to obtain an AI diagnosis global model, and the block is generated by a verification node.

2. The method of claim 1, wherein, The step of training a local AI diagnosis model through medical data collected by a medical node and the model fragments to obtain model parameters comprises the steps of: training a local AI diagnosis model through the medical data and the model fragments to obtain initial model parameters; calculating the gradient of the loss function of the local AI diagnosis model; calculating an update gradient using the gradient and a preset injected noise; calculating an update model parameter using the update gradient and the initial model parameter; determining whether the loss function meets a convergence condition, and if not, returning to the step of calculating the gradient of the loss function of the local AI diagnosis model; if yes, the update model parameter is taken as the model parameter of the trained local AI model.

3. The method of claim 1, wherein, The step of performing reliability verification on the model parameters of each medical node in turn through a three-dimensional reputation evaluation model to obtain a medical node that passes the verification when the zero-knowledge proof is verified comprises the steps of: performing reliability verification on the model parameters of each medical node in turn through a three-dimensional reputation evaluation model to obtain a suspected abnormal node when the zero-knowledge proof is verified; detecting a Byzantine node in the suspected abnormal node; removing the Byzantine node from the medical node to obtain a medical node that passes the verification.

4. The method of claim 3, wherein, The step of detecting a Byzantine node in the suspected abnormal node comprises the steps of: analyzing the gradient distribution difference of the suspected abnormal node through KL divergence, and determining a suspicious node according to the gradient distribution difference; performing explainability analysis on the suspicious node to determine a Byzantine node.

5. The method of claim 3, wherein, The three-dimensional reputation evaluation model is: wherein, is the credibility of the model parameters, represents the accuracy of the historical updates of the medical node, is the time series average of the gradient verification pass rate, quantifies the storage and computing power contribution of the node, is the model parameter, parameter ensures the weight normalization.

6. The method of claim 1, wherein, The method further comprises the steps of: calculating the contribution degree of each medical node to the AI diagnosis global model through a contribution quantification model, and distributing rewards and penalties to the medical nodes according to the contribution degree. 7.A trusted AI triage model generation apparatus, characterized by comprising: The method comprises the steps of: a model parameter generation module configured to receive model fragments of a previous round of AI diagnosis global model through polynomial interpolation, train a local AI diagnosis model through medical data collected by a medical node and the model fragments, obtain model parameters, and generate zero-knowledge proof; a verification module configured to perform reliability verification on the model parameters of each medical node in turn through a three-dimensional reputation evaluation model to obtain a medical node that passes the verification when the zero-knowledge proof is verified; an AI diagnosis global model generation module configured to store the model parameters of the medical node that passes the verification in a block to obtain an AI diagnosis global model, and the block is generated by a verification node.

8. The apparatus of claim 7, wherein, The model parameter generation module comprises: An initial model parameter generation submodule is configured to, when medical data is collected by the medical node, train a local AI diagnosis model by using the medical data and the model fragment, and obtain initial model parameters; A gradient calculation submodule is configured to calculate a gradient of a loss function of the local AI diagnosis model; An update gradient calculation submodule is configured to calculate an update gradient by using the gradient and a preset injection noise; An update model parameter calculation submodule is configured to calculate update model parameters by using the update gradient and the initial model parameters; A return submodule is configured to determine whether the loss function meets a convergence condition, and if not, return to the step of calculating the gradient of the loss function of the local AI diagnosis model; A model parameter determination submodule is configured to, if yes, take the update model parameters as model parameters of a trained local AI model.

9. An electronic device, comprising: The device comprises a processor and a memory: The memory is configured to store program code and transmit the program code to the processor; The processor is configured to execute the instructions in the program code to perform the trusted AI diagnosis model generation method in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is configured to store program code, and the program code is configured to execute the trusted AI diagnosis model generation method in any one of claims 1-6.

Images