Data processing system of medical model federal learning framework

Abstract

The embodiment of the invention relates to a data processing system of a medical model federated learning framework. The system comprises a central network and a plurality of first-class and second-class mechanism networks, the central network comprises a central server and a base model, and the central server is respectively connected with the base model, the first mechanism server and the second mechanism server; the first-class institution network comprises a first institution server, a first sample set and a first local model, and the first institution server is connected with the first sample set and the first local model; the second-class mechanism network comprises a second mechanism server, a second sample set and a second local model, and the second mechanism server is respectively connected with the second sample set and the second local model and is also interconnected with the other one or more first and second mechanism servers. According to the system, the training sample range can be expanded, safe transmission guarantee can be provided for longitudinally transmitted model parameters, safe transmission guarantee can be provided for shared information when a transverse mechanism shares training data, and privacy protection can be provided for non-shared information.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a data processing system for a medical model federated learning framework. Background Art

[0002] Artificial Intelligence (AI) models used in the medical field are collectively referred to as medical models. Common types include: text object processing models for medical document classification and recognition, visual object processing models for tissue segmentation or lesion identification in medical images, feature prediction models for disease risk prediction based on medical features, and so on. Various medical models need to be fully trained before use. Insufficient training data will reduce the generalization ability of the model. Currently, a large amount of medical data is mainly distributed in various medical institutions (such as hospitals, medical teaching / research institutions, etc.). If the traditional centralized training mechanism is adopted, it is necessary to first collect and store massive amounts of medical data from all institutions, and then configure large computing resources to centrally train the model. This will undoubtedly bring great challenges in terms of storage resource allocation, computing resource allocation, and privacy data protection.

[0003] Federated Learning is a technical framework for distributed model training. This framework not only reduces the risk of privacy data leakage but also effectively prevents central servers from overdrawing storage and / or computing resources. This technical framework can be broken down into two categories: the Vertical Federated Learning (VFL) framework and the Horizontal Federated Learning (HFL) framework. Comparing the two frameworks, we can see that they share the same principles: a central server is set up as the central node in the central network, and an institutional server is set up as a distributed node in the institutional network of each medical institution. Each distributed node trains the medical model locally based on the institutional sample data, while the central server is only responsible for aggregating the model parameters of all nodes and publishing the aggregated new model parameters to the entire network. The differences between the two are: 1) In the vertical federated learning framework, it is assumed that the sample data of each distributed node has all the feature types of the model input features, so only local sample data needs to be trained. Each distributed node only transfers model parameters with the central node, and no cross-node data transfer is performed between nodes; 2) In the horizontal federated learning framework, it is assumed that the sample data of each distributed node does not have the complete model input feature types, and the sample features of different distributed nodes can complement each other. Therefore, each distributed node not only needs to transfer model parameters with the central node, but also needs to transfer training data across nodes.

[0004] Most of the current common distributed training frameworks for medical models are implemented based on a vertical federated learning framework. This requires that the sample features of all institutions where distributed nodes are located within the framework must be sufficiently rich and complete. However, in reality, this data requirement can only be met by higher-level hospitals (such as tertiary hospitals). In other words, the distributed training framework for medical models implemented based on the vertical federated learning mechanism can only establish distributed nodes in high-level hospitals, which invisibly limits the scope of node access (medium and low-level hospitals cannot access it) and the scope of training samples (medical data from medium and low-level hospitals cannot be included in the full set of training samples).

[0005] To solve this problem, we constructed an integrated learning framework by adding a horizontal federated learning framework to the original vertical federated learning framework. This integrated learning framework regards high-level medical institution networks as first-class institutional networks, and keeps the vertical federated learning framework of all first-class institutional networks unchanged; in addition, the medium and low-level hospital institution networks are regarded as second-class institutional networks, and access to the second-class institutional networks is achieved by constructing a horizontal federated learning framework, and a training data transmission channel is set up between the second-class institutional networks. Based on this integrated learning framework, the node access range and the training sample range can be further expanded. However, it also increases the management complexity of the data transmission of the entire system: 1) It is necessary to ensure the security of the model parameter transmission between each first- and second-class institution and the central network (the network where the central server is located); 2) When the horizontal institutions share training data, it is necessary to ensure the transmission security of the shared data and to ensure that the non-shared data will not be leaked. This is exactly the technical problem that the present invention needs to solve. Summary of the Invention

[0006] The purpose of the present invention is to address the defects of the existing technology and provide a data processing system for a medical model federated learning framework, which includes: a central network, multiple Class I institution networks, and multiple Class II institution networks; the central network is located in a data center, the Class I institution networks are located in high-level medical institutions, and the Class II institution networks are located in medium and low-level hospital institutions; the federated learning framework composed of the central network and all Class I institution networks belongs to a vertical federated learning framework, and the Class I institution servers under the vertical framework have no interconnection relationship with each other and are only connected to the central server upward; the federated learning framework composed of the central network and all Class II institution networks belongs to a horizontal federated learning framework, and the Class II institution servers under the horizontal framework can be horizontally interconnected with the services of other Class I or Class II institutions. The central server of the central network transmits data to each institution's server through an independent secure data transmission channel. The central server is used to regularly synchronize the model feature set of the medical model and the public parameter set required for the Oblivious Transfer (OT) protocol to all institution servers. It is also used to regularly distribute the latest version of the base model's model parameters to all institution servers to initiate a round of model training tasks. It is also used to aggregate and decrypt the ciphertext of the institution parameter set fed back by all institution servers using a homomorphic decryption method to obtain the latest version of the model parameters of the base model. In each round of model training tasks, the server of a type of institution trains the local model based on the sample set of the institution, encrypts the trained institution model parameters using a homomorphic encryption method, and sends the obtained institution parameter set ciphertext to the central server. In each round of model training, the servers of the second type of institutions first obtain supplementary features from other horizontally connected institutional servers based on the Private Set Intersection (PSI) method implemented by the OT protocol to improve the sample set of the institution, and then train the local model based on the improved sample set. The trained institutional model parameters are then encrypted based on the homomorphic encryption method, and the obtained institutional parameter set ciphertext is sent to the central server. On the one hand, the system of the present invention expands the node access range and the training sample range through the added horizontal federated learning framework; on the other hand, it improves the transmission security of institutional model parameters through the homomorphic encryption mechanism; on the other hand, the PSI method implemented by the OT protocol not only ensures the transmission security of horizontal institutional shared information, but also improves the privacy protection of non-shared information.

[0007] To achieve the above-mentioned purpose, an embodiment of the present invention provides a data processing system for a medical model federated learning framework, the system comprising: a central network, a plurality of first-class institution networks, and a plurality of second-class institution networks;

[0008] The central network corresponds to a data center; the central network includes a central server and a base model; the central server is connected to the base model; the central server is also connected to the first and second institution servers of each type I and II institution network respectively;

[0009] The base model is a type of preset medical model;

[0010] The central server is used to establish a secure data transmission channel with each of the first and second institution servers according to a preset key negotiation protocol, which is recorded as a corresponding first channel;

[0011] The central server is further configured to periodically synchronize the model feature set of the medical model and the public parameter set required by the oblivious transfer protocol to all institutional servers through all the first channels;

[0012] The central server is also used to regularly use the latest version model parameters of the base model as the corresponding previous version base parameter set; and based on the preset parameter aggregation mode, generate a pair of public and private key pairs for homomorphic encryption and decryption calculations for the previous version base parameter set, which are recorded as corresponding public and private keys. and all the first channels will carry the public key The model training notification of the previous version base parameter set is distributed to all institutional servers; the parameter aggregation mode includes a non-weighted mode and a weighted mode;

[0013] The central server is further configured to receive the ciphertext of the institution parameter set sent back by all institution servers through all the first channels, based on the parameter aggregation mode, the private key The previous version base parameter set and all the mechanism parameter set ciphertexts are subjected to new version parameter aggregation processing to obtain a corresponding new version base parameter set; and the new version base parameter set is used as the latest version model parameter of the base model and saved;

[0014] The type of institution network corresponds to a medical structure; the type of institution network includes the first institution server, the first sample set and the first local model; the first institution server is connected to the first sample set and the first local model respectively;

[0015] The first local model and the base model have the same model structure but different model parameters;

[0016] The first organization server is used to store the received model feature set, the public parameter set and the model training notification;

[0017] The first institution server is further configured to perform a round of local training on the first local model based on the model training notification and the first sample set each time a new model training notification is received, and obtain the corresponding ciphertext of the institution parameter set and send it back to the central server;

[0018] The second type of institution network corresponds to a medical structure; the second type of institution network includes the second institution server, the second sample set and the second local model; the second institution server is connected to the second sample set and the second local model respectively; each second institution server is also interconnected with one or more other first and second institution servers;

[0019] The second local model and the base model have the same model structure but different model parameters;

[0020] The second organization server is used to store the received model feature set, the public parameter set and the model training notification;

[0021] The second agency server is also used to extract the corresponding public key from the current model training notification every time a new model training notification is received. and the front-version base parameter set; and based on the front-version base parameter set, the second local model is initialized with parameters; and based on the full set of model features and the second sample set, a missing feature check is performed to obtain the corresponding missing feature set X; and based on the public parameter set and the missing feature set X, a round of interaction is performed with other institutional servers connected to itself, and through this round of interaction, the sample feature completion processing of the second sample set is performed to obtain the corresponding third sample set; and based on the third sample set, a round of local training is performed on the second local model to obtain the corresponding institutional parameter set ciphertext and send it back to the central server.

[0022] Preferably, the model feature set includes a plurality of model input / output features, each of the model input / output features corresponds to a type of medical feature information;

[0023] The oblivious transfer protocol is implemented based on the discrete logarithm problem; the public parameter set of the oblivious transfer protocol includes the cyclic group parameter G of the discrete logarithm problem, the generator parameter g of the cyclic group, the order parameter q of the cyclic group, and the hash function H for mapping the patient identity to the cyclic group element. ID , hash function H for key derivation KEY ; The hash function H ID The function input is a patient's social identity identifier and the output is a group element in the cyclic group parameter G; the hash function H KEYThe input is the result of the exponentiation operation of a cyclic group element, and the output is a derived key of a specified key length; the social identity identifier includes at least an identity card identifier;

[0024] The first sample set includes multiple first sample records; each first sample record is a historical medical record of a patient, including the social identity identifier of the current patient and all the medical feature information of the current patient in the current medical structure; the corresponding patients of all the first sample records are not repeated; the types of medical feature information include at least medical history text, physical examination report text, medical examination report text, medical test report text, medical examination image, medical examination audio, treatment plan text, and diagnosis record text; all the medical feature information of the first sample records covers all the model input / output features of the model feature set;

[0025] The second sample set includes multiple second sample records; each second sample record is a historical medical record of a patient, including the social identity identifier of the current patient and all the medical feature information of the current patient in the current medical structure; the corresponding patients of all the second sample records are not repeated; all the medical feature information of the second sample records partially or completely covers all the model input / output features of the model feature set; in the second sample set, it is assumed that the types of medical feature information of all the second sample records are fully aligned, and there is no difference in the type range of medical feature information between any two second sample records;

[0026] The third sample set includes a plurality of third sample records; each third sample record is a medical record of a patient, including the social identity identifier of the current patient and all the medical feature information of the current patient in one or more medical institutions; the corresponding patients of all the third sample records are non-repeated; all the medical feature information of the third sample records covers all the model input / output features of the model feature set;

[0027] When the missing feature set X is not empty, it consists of one or more missing features x j Composition, 1≤indexj≤N X , N X is the total number of missing features in the current missing feature set; each missing feature x j Corresponding to one of the model input / output features in the model feature set.

[0028] Preferably, the key negotiation protocol includes at least DH protocol, ECDH protocol, IKEv2 protocol, and TLS-PSK protocol;

[0029] The two servers on both sides of each first channel have a pair of shared keys;

[0030] The transmission mechanism of each first channel is: the central server or the first and second institutional servers serving as the data sender of the current channel encrypts the original plaintext based on the shared key and sends the obtained ciphertext to the data receiver on the other side of the current channel; the first and second institutional servers or the central server serving as the data receiver of the current channel decrypts the ciphertext based on the shared key to obtain the corresponding original plaintext.

[0031] Preferably, the central server is specifically used to generate a pair of public and private key pairs for homomorphic encryption and decryption calculations for the front-end base parameter set based on the preset parameter aggregation mode, which are recorded as corresponding public and private keys. hour:

[0032] Identifying the parameter aggregation pattern;

[0033] If the parameter aggregation mode is a non-weighted mode, the public-private key pair is generated based on the preset semi-homomorphic algorithm to obtain the corresponding public key and the private key The semi-homomorphic algorithm includes at least the Paillier algorithm;

[0034] If the parameter aggregation mode is weighted mode, the public-private key pair is generated based on the preset fully homomorphic algorithm to obtain the corresponding public key and the private key The fully homomorphic algorithm supports at least the BFV algorithm, the BGV algorithm, the CKKS algorithm, the FHEW algorithm, and the TFHE algorithm.

[0035] Preferably, the central server is specifically used for When the previous version base parameter set and all the mechanism parameter set ciphertexts are aggregated to obtain the corresponding new version base parameter set:

[0036] The total number of the ciphertexts of the organization parameter set is counted to obtain the corresponding total number N; and each ciphertext of the organization parameter set is recorded as the corresponding ciphertext EW i , 1≤index i≤N; and the front base parameter set is recorded as the corresponding parameter set W old ; and record the new version base parameter set as the corresponding parameter set W new ;

[0037] and identifying the parameter aggregation pattern;

[0038] If the parameter aggregation mode is non-weighted mode, then press Generate the corresponding new version of the differential parameter set ciphertext EW new ;

[0039] If the parameter aggregation mode is non-weighted mode, then press Generate the corresponding new version of the differential parameter set ciphertext EW new ; α i is the ciphertext EW i The corresponding preset weight coefficient;

[0040] And based on the private key The new differential parameter set ciphertext EW new Perform homomorphic decryption to obtain the corresponding new differential parameter set △W new ; is a preset homomorphic decryption function used to use the private key The new differential parameter set ciphertext EW new Decryption;

[0041] And based on the parameter set W old and the new differential parameter set △W new Calculate the corresponding parameter set W new , W new =W old +ΔW new .

[0042] Preferably, the first institution server is specifically configured to, when performing a round of local training on the first local model based on the model training notification and the first sample set to obtain the corresponding institution parameter set ciphertext and sending it back to the central server:

[0043] Extract the corresponding public key from the model training notification and said front version base parameter set;

[0044] and initializing parameters of the first local model based on the front-plate base parameter set;

[0045] and performing a round of training on the initialized first local model based on the first sample set;

[0046] At the end of this round of training, the current model parameter set of the first local model is recorded as parameter set W now , record the front plate base parameter set as parameter set W old , and based on the parameter set W now 、W old Calculate the latest differential parameter set △W=W now -W old ;

[0047] And based on the public key Encrypting the differential parameter set ΔW to obtain the corresponding mechanism parameter set ciphertext; is a preset homomorphic encryption function for using the public key Encrypting the latest differential parameter set ΔW;

[0048] And send the current encrypted text of the mechanism parameter set to the central server through the corresponding first channel.

[0049] Preferably, the second mechanism server is specifically used to identify all medical features covered by the second sample set to obtain a corresponding medical feature set when the missing feature check is performed based on the complete model feature set and the second sample set to obtain a corresponding missing feature set X; and identify the feature intersection of the medical feature set and the complete model feature set to obtain a corresponding current feature intersection; and use the feature set in the complete model feature set except the current feature intersection as the corresponding missing feature set X.

[0050] Preferably, the second institution server is specifically configured to, when performing a round of interaction with other institution servers connected to itself based on the public parameter set and the missing feature set X, and performing sample feature completion processing on the second sample set through this round of interaction to obtain a corresponding third sample set:

[0051] Perform a round of traversal on all the second sample records of the second sample set; and in this round of traversal, use the second sample record currently traversed as the corresponding current sample record; and increase N in the current sample record X The missing features x of the missing feature set X j The medical feature information that corresponds one to one and is preset as empty information is used to obtain a corresponding third sample record; and at the end of this round of traversal, all the third sample records obtained in this round of traversal form an initialized third sample set;

[0052] And perform a round of traversal on all the first and second institution servers connected to itself; and in this round of traversal, use the first or second institution server currently traversed as the current sender server; and based on the privacy set intersection method implemented by the oblivious transfer protocol, obtain the sample missing features from the sender server according to the public parameter set and the missing feature set X to fill the third sample set; and at the end of this round of traversal, output the latest third sample set as the processing result of this sample feature filling processing.

[0053] Furthermore, the second institution server is specifically configured to, when the privacy set intersection method based on the oblivious transfer protocol is used to obtain the sample missing features from the sender server according to the public parameter set and the missing feature set X to fill the third sample set,

[0054] Sending an interactive request carrying the missing feature set X to the sending server;

[0055] The total number of records in the third sample set is counted to obtain the corresponding total number N U ; and bring each of the social identity identifiers of the third sample set into the hash function H ID Perform calculations and use the calculation results as the corresponding identification element h u =H ID (ID u ), 1≤indexu≤N U , ID u is the u-th social identity identifier in the third sample set, h u ∈G;

[0056] And based on the order parameter q, the corresponding remainder ring parameter Z is set q {0,1,2,…,q-1};

[0057] And the remainder ring parameter Z q Perform N U The random number is selected and each selected random number is used as a corresponding patient private key S u , S u ∈Z q ; and based on the generator parameter g and each of the patient's private keys S u and its corresponding identification element h u Generate a corresponding patient public key C u , C u ∈G;

[0058] After receiving the sender's public key H sent back by the sender server, N U The patient's public key C u The patient public key set {C u}Send to the sending server; and receive the ciphertext matrix {ED u,v}; the ciphertext matrix {ED u,v} by N U ×N V Ciphertext data ED u,v composition;

[0059] And based on the hash function HKEY , the sender's public key H and each of the patient's private keys S u Generate a corresponding patient shared key K u ,

[0060] And based on the shared key K of each patient u Decrypted corresponding N V The ciphertext data ED u,v Get the corresponding N V Plaintext data DED u,v , DED u,v =f DE (ED u,v ,K u,v ), f DE The decryption function of the preset symmetric encryption and decryption algorithm;

[0061] And N U ×N V The plaintext data DED u,v Perform a round of traversal; and in this round of traversal, the plaintext data DED currently traversed u,v As the corresponding current plaintext; and according to the preset patient missing information set and the check code assembly rule, extract the corresponding patient missing information set D from the current plaintext u,v Sum check code MAC u,v ; and based on the preset check code algorithm, the patient missing information set D u,v Perform check code calculation to obtain the corresponding current check code MAC now ; and in the current check code MAC now With the MAC u,v When matching, the current patient missing information set D u,v As the patient missing information set D corresponding to the current plaintext u , and based on the current patient missing information set D u For the N in the u-th third sample record of the third sample set X The medical feature information corresponding to the missing feature set X is set.

[0062] Furthermore, after receiving the interaction request, the sending server uses the second organization server that sends the current interaction request as the current receiving server;

[0063] and extracting the corresponding missing feature set X from the current interaction request;

[0064] and using the first or second local sample set as the corresponding current sample set;

[0065] And the total number of records in the current sample set is counted to obtain the corresponding total number N V ;

[0066] And bring each social identity identifier of the current sample set into the hash function H ID Perform calculations and use the calculation results as the corresponding identification element h v =H ID (ID v ), 1≤indexv≤N V , ID v is the vth social identity identifier in the current sample set, h v ∈G;

[0067] And based on the order parameter q, the corresponding remainder ring parameter Z is set q {0,1,2,…,q-1};

[0068] And the remainder ring parameter Z q Select a random number and use the selected random number as a corresponding sender private key S, S∈Z q ; and based on the generator parameter g, generate a corresponding sender public key H for the sender private key S, H = g S , H∈G;

[0069] And send the sender's public key H back to the receiving server; and after sending the sender's public key H, receive the patient's public key set {C u}; The patient public key set {C u} by N U The patient's public key C u composition;

[0070] And based on the hash function H KEY , the patient public key set {C u}, the sender's private key S and all the identification elements h v Generate the corresponding patient shared key K u,v ,

[0071] And each missing feature x in the first or second sample record corresponding to the vth social identity identifier in the current sample set and the missing feature set X j The corresponding medical feature information is extracted as the corresponding missing feature information d v,j , if the sample records of the current sample set do not have the missing feature x j The corresponding medical feature information corresponds to the missing feature information d v,jSet to empty; and N corresponding to each social identity X The missing feature information d v,j The corresponding patient missing information set D v {d v,j};

[0072] And based on the preset check code algorithm, each patient missing information set D v {d v,j}Calculate the check code to get the corresponding check code MAC v ;

[0073] And each patient missing information set D v {d v,j} and the corresponding verification code MAC v Form a corresponding patient data DM v ;

[0074] And based on the shared key K of each patient u,v Encrypt the corresponding patient data DM v Get the corresponding ciphertext data ED u,v , ED u,v =f EN (D v ,K u,v ), f EN The encryption function of the preset symmetric encryption and decryption algorithm, f EN With f DE They are inverse operation functions;

[0075] And the obtained N U ×N V The ciphertext data ED u,v The corresponding ciphertext matrix {ED u,v}Send back to the receiving server.

[0076] Preferably, the second institution server is specifically configured to, when performing a round of local training on the second local model based on the third sample set to obtain the corresponding institution parameter set ciphertext and sending it back to the central server:

[0077] Performing a round of training on the initialized second local model based on the third sample set;

[0078] At the end of this round of training, the current model parameter set of the second local model is recorded as parameter set W' now , record the front plate base parameter set as parameter set W' old , and based on the parameter set W' now 、W' oldCalculate the latest differential parameter set △W'=W' now -W' old ;

[0079] And based on the public key Encrypting the differential parameter set ΔW' to obtain the corresponding mechanism parameter set ciphertext; is a preset homomorphic encryption function for using the public key Encrypting the latest differential parameter set ΔW';

[0080] And send the current encrypted text of the mechanism parameter set to the central server through the corresponding first channel.

[0081] An embodiment of the present invention provides a data processing system for a medical model federated learning framework. The system comprises: a central network, multiple Class I institution networks, and multiple Class II institution networks. The central network is located in a data center, with Class I institution networks located in high-level medical institutions and Class II institution networks located in mid- and low-level hospitals. The federated learning framework formed by the central network and all Class I institution networks is a vertical federated learning framework. Class I institution servers within the vertical framework are not interconnected and are only connected upward to the central server. The federated learning framework formed by the central network and all Class II institution networks is a horizontal federated learning framework. Class II institution servers within the horizontal framework can interconnect horizontally with services from other Class I or Class II institutions. The central server of the central network communicates with each institution server via independent secure data transmission channels. The central server is responsible for regularly synchronizing the model feature set of the medical model and the common parameter set required by the OT protocol with all institution servers. It is also responsible for regularly distributing the latest version of the model parameters of the base model to all institution servers to initiate a round of model training tasks. It is also responsible for aggregating and decrypting the ciphertext of the institution parameter set feedback from all institution servers using a homomorphic decryption method to obtain the latest version of the model parameters of the base model. In each round of model training tasks, the servers of the first type of institutions train the local model based on the sample set of the institution, encrypt the institutional model parameters obtained through training based on the homomorphic encryption method, and send the obtained institutional parameter set ciphertext to the central server. In each round of model training tasks, the servers of the second type of institutions first obtain supplementary features from other horizontally connected institutional servers based on the PSI method implemented by the OT protocol to improve the sample set of the institution, and then train the local model based on the improved sample set, and then encrypt the institutional model parameters obtained through training based on the homomorphic encryption method, and send the obtained institutional parameter set ciphertext to the central server. On the one hand, the system of the embodiment of the present invention expands the node access range and the training sample range through the added horizontal federated learning framework; on the other hand, it improves the transmission security of institutional model parameters through the homomorphic encryption mechanism; on the other hand, the PSI method implemented by the OT protocol not only ensures the transmission security of horizontal institutional shared information, but also improves the privacy protection of non-shared information. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 A module structure diagram of a data processing system for a medical model federated learning framework provided by an embodiment of the present invention;

[0083] Figure 2 A schematic diagram of the interaction between a receiving server and a sending server provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0084] To make the objectives, technical solutions, and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the embodiments described herein are merely some, rather than all, of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0085] The embodiment of the present invention provides a data processing system for a medical model federated learning framework, such as Figure 1 As shown in the module structure diagram of a data processing system of a medical model federated learning framework provided in an embodiment of the present invention, it mainly includes: a central network 1, multiple first-class institution networks 2, and multiple second-class institution networks 3.

[0086] (1) Central Network 1:

[0087] The central network 1 in the embodiment of the present invention corresponds to a data center.

[0088] The central network 1 includes a central server 11 and a base model 12; the central server 11 is connected to the base model 12; the central server 11 is also connected to the first institution server 21 of each type I institution network 2 and the second institution server 31 of each type II institution network 3 respectively.

[0089] The central server 11 is used to establish a secure data transmission channel with each first institution server 21 and second institution server 31 according to a preset key negotiation protocol, which is recorded as a corresponding first channel.

[0090] Here, the key agreement protocol of the embodiment of the present invention includes at least the DH protocol, the ECDH protocol, the IKEv2 protocol, and the TLS-PSK protocol.

[0091] The secure data transmission channel (i.e., the first channel) built using the key agreement protocol is characterized by the presence of an identical pair of shared keys on the two servers on either side of each channel. The transmission mechanism for the first channel is as follows: the central server 11, first agency server 21, or second agency server 31, acting as the data sender in the current channel, encrypts the original plaintext using the shared key and sends the resulting ciphertext to the data receiver on the other side of the current channel. The first agency server 21, second agency server 31, or central server 11, acting as the data receiver in the current channel, then decrypts the ciphertext using the shared key to obtain the corresponding original plaintext.

[0092] The central server 11 is further configured to periodically synchronize the model feature set of the medical model and the public parameter set required by the oblivious transfer protocol to all institutional servers through all first channels.

[0093] Here, the model feature set of the embodiment of the present invention includes multiple model input / output features, and each model input / output feature corresponds to a type of medical feature information.

[0094] The OT protocol of the embodiment of the present invention is an OT protocol implemented based on the discrete logarithm problem; the public parameter set of the OT protocol includes the cyclic group parameter G of the discrete logarithm problem, the generator parameter g of the cyclic group, the order parameter q of the cyclic group, and the hash function H for mapping the patient identity to the cyclic group element. ID , hash function H for key derivation KEY ; Among them, the hash function H ID The function input is a patient's social identity identifier and the output is a group element in a cyclic group parameter G; the hash function H KEY The input is the result of the power operation of a cyclic group element, and the output is a derived key of a specified key length; the social identity identifier in the embodiment of the present invention includes at least an identity card identifier.

[0095] The central server 11 is also used to regularly use the latest version model parameters of the base model 12 as the corresponding previous version base parameter set; and based on the preset parameter aggregation mode, generate a pair of public and private key pairs for homomorphic encryption and decryption calculations for the previous version base parameter set, which are recorded as the corresponding public key and private key And all first channels will carry the public key Model training notifications for the previous version of the base parameter set are distributed to all institutional servers.

[0096] Here, the parameter aggregation mode of the embodiment of the present invention includes a non-weighted mode and a weighted mode.

[0097] The central server 11 is also used to receive the ciphertext of the organization parameter set sent back by all the organization servers through all the first channels, based on the parameter aggregation mode, private key The previous version of the base parameter set and all the mechanism parameter set ciphertexts are subjected to new version parameter aggregation processing to obtain the corresponding new version of the base parameter set; and the new version of the base parameter set is used as the latest version of the model parameters of the base model 12 and saved.

[0098] The base model 12 in the embodiment of the present invention is a type of preset medical model.

[0099] In a specific implementation of an embodiment of the present invention, the central server 11 is specifically used to generate a pair of public and private key pairs for homomorphic encryption and decryption calculations for the front-end base parameter set based on a preset parameter aggregation mode, which are recorded as corresponding public keys. and private key hour:

[0100] Step A1, identifying parameter aggregation patterns;

[0101] Step A2: If the parameter aggregation mode is non-weighted mode, the public-private key pair is generated based on the preset semi-homomorphic algorithm to obtain the corresponding public key. and private key

[0102] Among them, the semi-homomorphic algorithm includes at least the Paillier algorithm;

[0103] Step A3: If the parameter aggregation mode is weighted mode, the public-private key pair is generated based on the preset fully homomorphic algorithm to obtain the corresponding public key. and private key

[0104] Among them, the fully homomorphic algorithm supports at least the BFV algorithm, BGV algorithm, CKKS algorithm, FHEW algorithm, and TFHE algorithm.

[0105] In another specific implementation of the embodiment of the present invention, the central server 11 is specifically configured to: When the previous version of the base parameter set and the ciphertext of all mechanism parameter sets are aggregated to obtain the corresponding new version of the base parameter set:

[0106] Step B1: Count the total number of the organization parameter set ciphertexts to obtain the corresponding total number N; and record each organization parameter set ciphertext as the corresponding ciphertext EW i , 1≤index i≤N; and the previous version base parameter set is recorded as the corresponding parameter set W old ; and record the new version of the base parameter set as the corresponding parameter set W new ;

[0107] Step B2, and identify the parameter aggregation mode; if the parameter aggregation mode is non-weighted mode, then press Generate the corresponding new version of the differential parameter set ciphertext EW new ; If the parameter aggregation mode is non-weighted mode, then press Generate the corresponding new version of the differential parameter set ciphertext EW new ;

[0108] Among them, α i is the i-th ciphertext EW i The corresponding preset weight coefficient;

[0109] Step B3, and based on the private key For the new differential parameter set ciphertext EW new Perform homomorphic decryption to obtain the corresponding new differential parameter set △W new ;

[0110] in,

[0111] It is a preset homomorphic decryption function used to use the private key For the new differential parameter set ciphertext EW new Decryption;

[0112] Step B4, and based on parameter set W old and the new differential parameter set △W new Calculate the corresponding parameter set W new ;

[0113] Among them, W new =W old +ΔW new .

[0114] (2) Category 1 Institutional Network 2:

[0115] The type 1 institutional network 2 of the embodiment of the present invention corresponds to a high-level medical structure, such as a tertiary hospital.

[0116] A type of institution network 2 includes a first institution server 21, a first sample set 22 and a first local model 23; the first institution server 21 is connected to the first sample set 22 and the first local model 23 respectively.

[0117] The first organization server 21 is used to store the received model feature set, public parameter set and model training notification.

[0118] The first organization server 21 is also used to perform a round of local training on the first local model 23 based on the model training notification and the first sample set 22 every time a new model training notification is received, and obtain the corresponding organization parameter set ciphertext and send it back to the central server 11.

[0119] The first sample set 22 of the embodiment of the present invention includes multiple first sample records; each first sample record is a patient's historical medical record, including the current patient's social identity identifier and all medical feature information of the current patient in the current medical structure. It should be noted that the corresponding patients of all first sample records are not repeated. It should also be noted that the types of medical feature information in the embodiment of the present invention include at least medical record text, physical examination report text, medical examination report text, medical test report text, medical examination image, medical examination audio, treatment plan text, and diagnosis record text. It should also be noted that all medical feature information of the first sample record covers all model input / output features of the model feature set.

[0120] The first local model 23 of the embodiment of the present invention has the same model structure as the base model 12 but has different model parameters.

[0121] In another specific implementation of the embodiment of the present invention, the first organization server 21 is specifically configured to perform a round of local training on the first local model 23 based on the model training notification and the first sample set 22 to obtain the corresponding organization parameter set ciphertext and send it back to the central server 11:

[0122] Step C1: Extract the corresponding public key from the model training notification and the previous version of the base parameter set;

[0123] Step C2, initializing parameters of the first local model 23 based on the previous version base parameter set;

[0124] Step C3: performing a round of training on the initialized first local model 23 based on the first sample set 22;

[0125] Step C4, at the end of this round of training, the current model parameter set of the first local model 23 is recorded as parameter set W now , record the previous version base parameter set as parameter set W old , and based on the parameter set W now 、W old Calculate the latest differential parameter set △W=W now -W old ;

[0126] Step C5, and based on the public key Encrypt the differential parameter set △W to obtain the corresponding mechanism parameter set ciphertext;

[0127] in,

[0128] It is a preset homomorphic encryption function used to use the public key Encrypt the latest differential parameter set △W;

[0129] Step C6, and send the current mechanism parameter set ciphertext to the central server 11 through the corresponding first channel.

[0130] (3) Category II Institutional Network 3:

[0131] The second-class institutional network 3 corresponds to a medium-to-low-level medical structure, such as Class II hospitals and community hospitals.

[0132] The second type of institutional network 3 includes a second institutional server 31, a second sample set 32 ​​and a second local model 33; the second institutional server 31 is connected to the second sample set 32 ​​and the second local model 33 respectively; each second institutional server 31 is also interconnected with another one or more first institutional servers 21 and second institutional servers 31.

[0133] The second mechanism server 31 is used to store the received model feature set, public parameter set and model training notification.

[0134] The second organization server 31 is also used to extract the corresponding public key from the current model training notification every time a new model training notification is received. and the previous version base parameter set; and based on the previous version base parameter set, the second local model 33 is initialized with parameters; and based on the full set of model features and the second sample set 32, a missing feature check is performed to obtain the corresponding missing feature set X; and based on the public parameter set and the missing feature set X, a round of interaction is performed with other institutional servers connected to itself and the sample feature completion processing of the second sample set 32 ​​is performed through this round of interaction to obtain the corresponding third sample set; and based on the third sample set, a round of local training is performed on the second local model 33 to obtain the corresponding institutional parameter set ciphertext and send it back to the central server 11.

[0135] Here, when the missing feature set X of the embodiment of the present invention is not empty, one or more missing features x j Composition, 1≤indexj≤N X , N X is the total number of missing features in the current missing feature set; each missing feature x j Corresponds to a model input / output feature in the model feature set.

[0136] The second sample set 32 ​​of the embodiment of the present invention includes multiple second sample records; each second sample record is a patient's historical medical record, including the current patient's social identity identifier and all medical feature information of the current patient in the current medical structure; the corresponding patients of all second sample records are not repeated; all medical feature information of the second sample records partially or completely covers all model input / output features of the model feature set; the second sample set 32 ​​defaults to fully aligning the types of medical feature information of all second sample records, and there is no difference in the type range of medical feature information between any two second sample records.

[0137] The third sample set of an embodiment of the present invention includes multiple third sample records; each third sample record is a medical record of a patient, including the current patient's social identity identifier and all medical feature information of the current patient in one or more medical structures; the corresponding patients of all third sample records are not repeated; all medical feature information of the third sample records covers all model input / output features of the model feature set.

[0138] The second local model 33 of the embodiment of the present invention has the same model structure as the base model 12 but has different model parameters.

[0139] In another specific implementation of an embodiment of the present invention, the second mechanism server 31 is specifically used to identify all medical features covered by the second sample set 32 ​​to obtain a corresponding medical feature set when performing a missing feature check based on the complete set of model features and the second sample set 32 ​​to obtain a corresponding missing feature set X; and to identify the feature intersection of the medical feature set and the complete set of model features to obtain a corresponding current feature intersection; and to use the feature set in the complete set of model features except the current feature intersection as the corresponding missing feature set X.

[0140] In another specific implementation of the embodiment of the present invention, the second organization server 31 is specifically configured to perform a round of interaction with other organization servers connected to itself based on the public parameter set and the missing feature set X, and to perform sample feature completion processing on the second sample set 32 ​​through this round of interaction to obtain a corresponding third sample set:

[0141] Step D1, perform a round of traversal on all second sample records of the second sample set 32; and in this round of traversal, use the second sample record currently traversed as the corresponding current sample record; and increase N in the current sample record. X The missing features x of the missing feature set X j One-to-one correspondence and preset empty medical feature information are used to obtain a corresponding third sample record; and at the end of this round of traversal, all the third sample records obtained in this round of traversal form an initialized third sample set;

[0142] Step D2, and perform a round of traversal on all the first institution servers 21 and second institution servers 31 connected to itself; and in this round of traversal, the currently traversed first institution server 21 or second institution server 31 is used as the current sending server; and based on the privacy set intersection method implemented by the oblivious transfer protocol, the sample missing features are obtained from the sending server according to the public parameter set and the missing feature set X to fill the third sample set; and at the end of this round of traversal, the latest third sample set is output as the processing result of this sample feature completion processing.

[0143] In another specific implementation of the embodiment of the present invention, the second organization server 31 is specifically configured to, when filling the third sample set with missing features of samples obtained from the sender server according to the public parameter set and the missing feature set X, using the privacy set intersection method implemented based on the oblivious transfer protocol:

[0144] Step E1, sending an interactive request carrying the missing feature set X to the sending server;

[0145] Step E2: Count the total number of records in the third sample set to obtain the corresponding total number N U ; and bring each social identity identifier of the third sample set into the hash function HID Perform calculations and use the calculation results as the corresponding identification element h u ;

[0146] Here, h u =H ID (ID u ),

[0147] Where 1≤index u≤N U , ID u is the u-th social identity identifier of the third sample set, h u ∈G;

[0148] Step E3, and set the corresponding remainder ring parameter Z based on the order parameter q q {0,1,2,…,q-1};

[0149] Step E4, and the remainder ring parameter Z q Perform N U The random number is selected and each selected random number is used as a corresponding patient private key S u , S u ∈Z q ; and based on the generator parameter g and each patient's private key S u and its corresponding identification element h u Generate a corresponding patient public key C u ;

[0150] here, C u ∈G;

[0151] Step E5, after receiving the sender's public key H sent back by the sender server, N U Patient public key C u The patient public key set {C u}Send to the sending server;

[0152] Step E6, and receive the ciphertext matrix {ED u,v};

[0153] Among them, the ciphertext matrix {ED u,v} by N U ×N V Ciphertext data ED u,v composition;

[0154] Step E7, and based on the hash function H KEY , the sender's public key H and each patient's private key S u Generate a corresponding patient shared key K u ;

[0155] here,

[0156] Step E8, and based on the shared key K of each patient u Decrypted corresponding N V Ciphertext data ED u,v Get the corresponding N V Plaintext data DED u,v ;

[0157] Here, DED u,v =f DE (ED u,v ,K u,v ),

[0158] Among them, f DE The decryption function of the preset symmetric encryption and decryption algorithm;

[0159] Step E9, and N U ×N V Plaintext data DED u,v Perform a round of traversal; and in this round of traversal, the currently traversed plaintext data DED u,v As the corresponding current plaintext; and according to the preset patient missing information set and check code assembly rules, extract the corresponding patient missing information set D from the current plaintext u,v Sum check code MAC u,v ; and based on the preset check code algorithm, the patient missing information set D u,v Perform check code calculation to obtain the corresponding current check code MAC now ; and in the current check code MAC now With MAC u,v When matching, the current patient missing information set D u,v As the patient missing information set D corresponding to the current plaintext u , and based on the current patient missing information set D u For the N in the u-th third sample record of the third sample set X The medical feature information corresponding to the missing feature set X is set.

[0160] Here, the preset check code algorithm of the embodiment of the present invention includes at least SHA series hash algorithm, SM3 hash algorithm, MD series digital digest algorithm. It should also be noted that, in the case of N U ×N V Plaintext data DED u,v When performing traversal verification, each index u has at most one matching patient missing information set D u .

[0161] The second organization server 31 that performs the above steps E1-E9 is regarded as a receiving server, and the processing flow on the sending server side corresponding to the receiving server is as follows:

[0162] Step F1, after receiving an interaction request, the sending server uses the second organization server 31 that sent the current interaction request as the current receiving server;

[0163] Step F2, and extract the corresponding missing feature set X from the current interaction request;

[0164] Step F3, and use the local first or second sample set 32 ​​as the corresponding current sample set;

[0165] Step F4: Count the total number of records in the current sample set to obtain the corresponding total number N V ;

[0166] Step F5, and bring each social identity identifier of the current sample set into the hash function H ID Perform calculations and use the calculation results as the corresponding identification element h v ;

[0167] Here, h v =H ID (ID v ),

[0168] Where 1≤indexv≤N V , ID v is the vth social identity in the current sample set, h v ∈G;

[0169] Step F6, and set the corresponding remainder ring parameter Z based on the order parameter q q {0,1,2,…,q-1};

[0170] Step F7, and the remainder ring parameter Z q Select a random number and use the selected random number as a corresponding sender private key S, S∈Z q ; and generate a corresponding sender public key H for the sender private key S based on the generator parameter g;

[0171] Here, H = g S , H∈G;

[0172] Step F8, and send the sender's public key H back to the receiving server;

[0173] Step F9, after sending the sender's public key H, receives the patient's public key set {C u};

[0174] Here, the patient public key set {C u} by N U Patient public key C u composition;

[0175] Step F10, and based on the hash function H KEY , patient public key set {C u}, sender's private key S and all identification elements h v Generate the corresponding patient shared key K u,v ;

[0176] here,

[0177] Step F11, and each missing feature x in the first or second sample record corresponding to the vth social identity in the current sample set and the missing feature set X j The corresponding medical feature information is extracted as the corresponding missing feature information d v,j , if the sample records of the current sample set do not have a missing feature x j The corresponding medical feature information corresponds to the missing feature information d v,j Set to empty; and N corresponding to each social identity X Missing feature information d v,j The corresponding patient missing information set D v {d v,j};

[0178] Step F12, and based on the preset verification code algorithm, each patient missing information set D v {d v,j}Calculate the check code to get the corresponding check code MAC v ;

[0179] Step F13, and each patient missing information set D v {d v,j} and the corresponding check code MAC v Form a corresponding patient data DM v ;

[0180] Step F14, and based on the shared key K of each patient u,v Encrypt the corresponding patient data DM v Get the corresponding ciphertext data ED u,v ;

[0181] Here, ED u,v =f EN (D v ,K u,v ),

[0182] Among them, f ENThe encryption function of the preset symmetric encryption and decryption algorithm, f EN With f DE They are inverse operation functions;

[0183] Step F15, and by the obtained N U ×N V Ciphertext data ED u,v The corresponding ciphertext matrix {ED u,v}Send back to the receiving server.

[0184] The above-mentioned receiving server-sending server interaction process can be found in Figure 2 The interaction diagram between the receiving server and the sending server provided in the embodiment of the present invention is understood.

[0185] In another specific implementation of the embodiment of the present invention, the second organization server 31 is specifically configured to perform a round of local training on the second local model 33 based on the third sample set to obtain a corresponding organization parameter set ciphertext and send it back to the central server 11:

[0186] Step G1, performing a round of training on the initialized second local model 33 based on the third sample set;

[0187] Step G2, at the end of this round of training, the current model parameter set of the second local model 33 is recorded as parameter set W' now , record the previous version base parameter set as parameter set W' old , and based on the parameter set W' now 、W' ol d. Calculate the latest differential parameter set △W'=W' now -W' old ;

[0188] Step G3, and based on the public key Encrypt the differential parameter set △W' to obtain the corresponding mechanism parameter set ciphertext;

[0189] here,

[0190] It is a preset homomorphic encryption function used to use the public key Encrypt the latest differential parameter set △W';

[0191] Step G4, and send the current mechanism parameter set ciphertext to the central server 11 through the corresponding first channel.

[0192] It should be noted that the division of the various modules of the above system is merely a division of logical functions. In actual implementation, they can be fully or partially integrated into a single physical entity or physically separated. Furthermore, these modules can be implemented entirely in software invoked by a processing element; entirely in hardware; or partially in software invoked by a processing element and partially in hardware. For example, the central server can be a separate processing element or a chip integrated into a device, equipment, or server. Furthermore, it can be stored in the memory of a device, equipment, or server in the form of program code and invoked by a processing element within the device, equipment, or server to implement the corresponding processing end functions. The implementation of other modules is similar. Furthermore, these modules can be fully or partially integrated or implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. During implementation, the processing steps of each module of the aforementioned system can be performed by hardware integrated logic circuits within the processor element or by software instructions.

[0193] For example, these modules of the above system can be one or more integrated circuits configured to implement the aforementioned processing steps, such as one or more application-specific integrated circuits (ASICs), or one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs). For another example, when a module of the above system is implemented in the form of a processing element scheduling program code, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules can be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0194] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the above method embodiments are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The above-mentioned computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the above-mentioned computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, Bluetooth, microwave, etc.) means. The above-mentioned computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The above-mentioned available medium can be a magnetic medium (such as a floppy disk, hard disk, tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0195] An embodiment of the present invention provides a data processing system for a medical model federated learning framework. The system comprises: a central network, multiple Class I institution networks, and multiple Class II institution networks. The central network is located in a data center, with Class I institution networks located in high-level medical institutions and Class II institution networks located in mid- and low-level hospitals. The federated learning framework formed by the central network and all Class I institution networks is a vertical federated learning framework. Class I institution servers within the vertical framework are not interconnected and are only connected upward to the central server. The federated learning framework formed by the central network and all Class II institution networks is a horizontal federated learning framework. Class II institution servers within the horizontal framework can interconnect horizontally with services from other Class I or Class II institutions. The central server of the central network communicates with each institution server via independent secure data transmission channels. The central server is responsible for regularly synchronizing the model feature set of the medical model and the common parameter set required by the OT protocol with all institution servers. It is also responsible for regularly distributing the latest version of the model parameters of the base model to all institution servers to initiate a round of model training tasks. It is also responsible for aggregating and decrypting the ciphertext of the institution parameter set feedback from all institution servers using a homomorphic decryption method to obtain the latest version of the model parameters of the base model. In each round of model training tasks, the servers of the first type of institutions train the local model based on the sample set of the institution, encrypt the institutional model parameters obtained through training based on the homomorphic encryption method, and send the obtained institutional parameter set ciphertext to the central server. In each round of model training tasks, the servers of the second type of institutions first obtain supplementary features from other horizontally connected institutional servers based on the PSI method implemented by the OT protocol to improve the sample set of the institution, and then train the local model based on the improved sample set, and then encrypt the institutional model parameters obtained through training based on the homomorphic encryption method, and send the obtained institutional parameter set ciphertext to the central server. On the one hand, the system of the embodiment of the present invention expands the node access range and the training sample range through the added horizontal federated learning framework; on the other hand, it improves the transmission security of institutional model parameters through the homomorphic encryption mechanism; on the other hand, the PSI method implemented by the OT protocol not only ensures the transmission security of horizontal institutional shared information, but also improves the privacy protection of non-shared information.

[0196] Professionals should also be further aware that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0197] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0198] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A data processing system for a medical model federated learning framework, characterized in that: The system includes: a central network, multiple first-class institution networks, and multiple second-class institution networks; The central network corresponds to a data center; the central network includes a central server and a base model; the central server is connected to the base model; the central server is also connected to the first and second institution servers of each type I and II institution network respectively; The base model is a type of preset medical model; The central server is used to establish a secure data transmission channel with each of the first and second institution servers according to a preset key negotiation protocol, which is recorded as a corresponding first channel; The central server is further configured to periodically synchronize the model feature set of the medical model and the public parameter set required by the oblivious transfer protocol to all institutional servers through all the first channels; The central server is also used to regularly use the latest version model parameters of the base model as the corresponding previous version base parameter set; and based on the preset parameter aggregation mode, generate a pair of public and private key pairs for homomorphic encryption and decryption calculations for the previous version base parameter set, which are recorded as corresponding public and private keys. and all the first channels will carry the public key The model training notification of the previous version base parameter set is distributed to all institutional servers; the parameter aggregation mode includes a non-weighted mode and a weighted mode; The central server is further configured to receive the ciphertext of the institution parameter set sent back by all institution servers through all the first channels, based on the parameter aggregation mode, the private key The previous version base parameter set and all the mechanism parameter set ciphertexts are subjected to new version parameter aggregation processing to obtain a corresponding new version base parameter set; and the new version base parameter set is used as the latest version model parameter of the base model and saved; The type of institution network corresponds to a medical structure; the type of institution network includes the first institution server, the first sample set and the first local model; the first institution server is connected to the first sample set and the first local model respectively; The first local model and the base model have the same model structure but different model parameters; The first organization server is used to store the received model feature set, the public parameter set and the model training notification; The first institution server is further configured to perform a round of local training on the first local model based on the model training notification and the first sample set each time a new model training notification is received, and obtain the corresponding ciphertext of the institution parameter set and send it back to the central server; The second type of institution network corresponds to a medical structure; the second type of institution network includes the second institution server, the second sample set and the second local model; the second institution server is connected to the second sample set and the second local model respectively; each second institution server is also interconnected with one or more other first and second institution servers; The second local model and the base model have the same model structure but different model parameters; The second organization server is used to store the received model feature set, the public parameter set and the model training notification; The second agency server is also used to extract the corresponding public key from the current model training notification every time a new model training notification is received. and the front-version base parameter set; and based on the front-version base parameter set, the second local model is initialized with parameters; and based on the full set of model features and the second sample set, a missing feature check is performed to obtain the corresponding missing feature set X; and based on the public parameter set and the missing feature set X, a round of interaction is performed with other institutional servers connected to itself, and through this round of interaction, the sample feature completion processing of the second sample set is performed to obtain the corresponding third sample set; and based on the third sample set, a round of local training is performed on the second local model to obtain the corresponding institutional parameter set ciphertext and send it back to the central server.

2. The data processing system of the medical model federated learning framework according to claim 1 is characterized in that: The model feature set includes a plurality of model input / output features, each of the model input / output features corresponding to a type of medical feature information; The oblivious transfer protocol is implemented based on the discrete logarithm problem; the public parameter set of the oblivious transfer protocol includes the cyclic group parameter G of the discrete logarithm problem, the generator parameter g of the cyclic group, the order parameter q of the cyclic group, and the hash function H for mapping the patient identity to the cyclic group element. ID , hash function H for key derivation KEY ; The hash function H ID The function input is a patient's social identity identifier and the output is a group element in the cyclic group parameter G; the hash function H KEY The input is the result of the exponentiation operation of a cyclic group element, and the output is a derived key of a specified key length; the social identity identifier includes at least an identity card identifier; The first sample set includes a plurality of first sample records; each first sample record is a historical medical record of a patient, including the social identity identifier of the current patient and all the medical characteristic information of the current patient in the current medical structure; the corresponding patients of all the first sample records are not repeated; the types of medical characteristic information include at least medical history text, physical examination report text, medical examination report text, medical test report text, medical examination image, medical examination audio, treatment plan text, and diagnosis record text; All the medical feature information of the first sample record covers all the model input / output features of the model feature set; The second sample set includes a plurality of second sample records; each second sample record is a historical medical record of a patient, including the social identity identifier of the current patient and all the medical characteristic information of the current patient in the current medical structure; All corresponding patients recorded in the second sample are not repeated; All of the medical feature information recorded in the second sample partially or completely covers all of the model input / output features of the model feature set; In the second sample set, it is assumed that the types of medical feature information of all second sample records are fully aligned, and there is no difference in the type range of medical feature information between any two second sample records; The third sample set includes a plurality of third sample records; each third sample record is a medical record of a patient, including the social identity identifier of the current patient and all the medical feature information of the current patient in one or more medical institutions; the corresponding patients of all the third sample records are not repeated; All of the medical feature information of the third sample record covers all of the model input / output features of the model feature set; When the missing feature set X is not empty, it consists of one or more missing features x j Composition, 1≤indexj≤N X , N X is the total number of missing features in the current missing feature set; each missing feature x j Corresponding to one of the model input / output features in the model feature set.

3. The data processing system of the medical model federated learning framework according to claim 1 is characterized in that: The key negotiation protocol includes at least DH protocol, ECDH protocol, IKEv2 protocol, and TLS-PSK protocol; The two servers on both sides of each first channel have a pair of shared keys; The transmission mechanism of each first channel is: the central server or the first and second institutional servers serving as the data sender of the current channel encrypts the original plaintext based on the shared key and sends the obtained ciphertext to the data receiver on the other side of the current channel; the first and second institutional servers or the central server serving as the data receiver of the current channel decrypts the ciphertext based on the shared key to obtain the corresponding original plaintext.

4. The data processing system of the medical model federated learning framework according to claim 1 is characterized in that: The central server is specifically used to generate a pair of public and private key pairs for homomorphic encryption and decryption calculations for the front-end base parameter set based on the preset parameter aggregation mode, which are recorded as corresponding public and private keys. hour: Identifying the parameter aggregation pattern; If the parameter aggregation mode is a non-weighted mode, the public-private key pair is generated based on the preset semi-homomorphic algorithm to obtain the corresponding public key and the private key The semi-homomorphic algorithm includes at least the Paillier algorithm; If the parameter aggregation mode is weighted mode, the public-private key pair is generated based on the preset fully homomorphic algorithm to obtain the corresponding public key and the private key The fully homomorphic algorithm supports at least the BFV algorithm, the BGV algorithm, the CKKS algorithm, the FHEW algorithm, and the TFHE algorithm.

5. The data processing system of the medical model federated learning framework according to claim 1 is characterized in that: The central server is specifically configured to: When the previous version base parameter set and all the mechanism parameter set ciphertexts are aggregated to obtain the corresponding new version base parameter set: The total number of the ciphertexts of the organization parameter set is counted to obtain the corresponding total number N; and each ciphertext of the organization parameter set is recorded as the corresponding ciphertext EW i , 1≤index i≤N; and the front base parameter set is recorded as the corresponding parameter set W old ; and record the new version base parameter set as the corresponding parameter set W new ; and identifying the parameter aggregation pattern; If the parameter aggregation mode is non-weighted mode, then press Generate the corresponding new version of the differential parameter set ciphertext EW new ; If the parameter aggregation mode is non-weighted mode, then press Generate the corresponding new version of the differential parameter set ciphertext EW new ; α i is the ciphertext EW i The corresponding preset weight coefficient; And based on the private key The new differential parameter set ciphertext EW new Perform homomorphic decryption to obtain the corresponding new differential parameter set △W new ; is a preset homomorphic decryption function used to use the private key The new differential parameter set ciphertext EW new Decryption; And based on the parameter set W old and the new differential parameter set △W new Calculate the corresponding parameter set W new , W new =W old +ΔW new .

6. The data processing system of the medical model federated learning framework according to claim 1 is characterized in that: The first organization server is specifically configured to, when performing a round of local training on the first local model based on the model training notification and the first sample set to obtain the corresponding organization parameter set ciphertext and sending it back to the central server: Extract the corresponding public key from the model training notification and said front version base parameter set; and initializing parameters of the first local model based on the front-plate base parameter set; and performing a round of training on the initialized first local model based on the first sample set; At the end of this round of training, the current model parameter set of the first local model is recorded as parameter set W now , record the front plate base parameter set as parameter set W old , and based on the parameter set W now 、W old Calculate the latest differential parameter set △W=W now -W old ; And based on the public key Encrypting the differential parameter set ΔW to obtain the corresponding mechanism parameter set ciphertext; is a preset homomorphic encryption function for using the public key Encrypting the latest differential parameter set ΔW; And send the current encrypted text of the mechanism parameter set to the central server through the corresponding first channel.

7. The data processing system of the medical model federated learning framework according to claim 2, characterized in that: The second mechanism server is specifically used to identify all medical features covered by the second sample set to obtain a corresponding medical feature set when the missing feature check is performed based on the complete model feature set and the second sample set to obtain a corresponding missing feature set X; and identify the feature intersection of the medical feature set and the complete model feature set to obtain a corresponding current feature intersection; and use the feature set in the complete model feature set except the current feature intersection as the corresponding missing feature set X.

8. The data processing system of the medical model federated learning framework according to claim 2 is characterized in that: The second institution server is specifically configured to, when performing a round of interaction with other institution servers connected to itself based on the public parameter set and the missing feature set X and completing sample feature processing on the second sample set through this round of interaction to obtain a corresponding third sample set: Perform a round of traversal on all the second sample records of the second sample set; and in this round of traversal, use the second sample record currently traversed as the corresponding current sample record; and increase N in the current sample record X The missing features x of the missing feature set X j The medical feature information that corresponds one to one and is preset as empty information is used to obtain a corresponding third sample record; and at the end of this round of traversal, all the third sample records obtained in this round of traversal form an initialized third sample set; And perform a round of traversal on all the first and second institution servers connected to itself; and in this round of traversal, use the first or second institution server currently traversed as the current sender server; and based on the privacy set intersection method implemented by the oblivious transfer protocol, obtain the sample missing features from the sender server according to the public parameter set and the missing feature set X to fill the third sample set; and at the end of this round of traversal, output the latest third sample set as the processing result of this sample feature filling processing.

9. The data processing system of the medical model federated learning framework according to claim 8, characterized in that: The second institution server is specifically configured to obtain sample missing features from the sender server according to the public parameter set and the missing feature set X to fill the third sample set in the privacy set intersection method implemented based on the oblivious transfer protocol. Sending an interactive request carrying the missing feature set X to the sending server; The total number of records in the third sample set is counted to obtain the corresponding total number N U ; and bring each of the social identity identifiers of the third sample set into the hash function H ID Perform calculations and use the calculation results as the corresponding identification element h u =H ID (ID u ), 1≤indexu≤N U , ID u is the u-th social identity identifier in the third sample set, h u ∈G; And based on the order parameter q, the corresponding remainder ring parameter Z is set q {0,1,2,…,q-1}; And the remainder ring parameter Z q Perform N U The random number is selected and each selected random number is used as a corresponding patient private key S u , S u ∈Z q ; and based on the generator parameter g and each of the patient's private keys S u and its corresponding identification element h u Generate a corresponding patient public key C u , C u ∈G; After receiving the sender's public key H sent back by the sender server, N U The patient's public key C u The patient public key set {C u }Send to the sending server; and receive the ciphertext matrix {ED u,v }; the ciphertext matrix {ED u,v } by N U ×N V Ciphertext data ED u,v composition; And based on the hash function H KEY , the sender's public key H and each of the patient's private keys S u Generate a corresponding patient shared key K u , And based on the shared key K of each patient u Decrypted corresponding N V The ciphertext data ED u,v Get the corresponding N V Plaintext data DED u,v , DED u,v =f DE (ED u,v ,K u,v ), f DE The decryption function of the preset symmetric encryption and decryption algorithm; And N U ×N V The plaintext data DED u,v Perform a round of traversal; and in this round of traversal, the plaintext data DED currently traversed u,v As the corresponding current plaintext; According to the preset rules for assembling the missing patient information set and the check code, the corresponding missing patient information set D is extracted from the current plaintext. u,v Sum check code MAC u,v ; and based on the preset check code algorithm, the patient missing information set D u,v Perform check code calculation to obtain the corresponding current check code MAC now ; and in the current check code MAC now With the MAC u,v When matching, the current patient missing information set D u,v As the patient missing information set D corresponding to the current plaintext u , and based on the current patient missing information set D u For the N in the u-th third sample record of the third sample set X The medical feature information corresponding to the missing feature set X is set.

10. The data processing system of the medical model federated learning framework according to claim 9, characterized in that: After receiving the interaction request, the sending server uses the second organization server that sends the current interaction request as the current receiving server; and extracting the corresponding missing feature set X from the current interaction request; and using the first or second local sample set as the corresponding current sample set; And the total number of records in the current sample set is counted to obtain the corresponding total number N V ; And bring each social identity identifier of the current sample set into the hash function H ID Perform calculations and use the calculation results as the corresponding identification element h v =H ID (ID v ), 1≤indexv≤N V , ID v is the vth social identity identifier in the current sample set, h v ∈G; And based on the order parameter q, the corresponding remainder ring parameter Z is set q {0,1,2,…,q-1}; And the remainder ring parameter Z q Select a random number and use the selected random number as a corresponding sender private key S, S∈Z q ; and based on the generator parameter g, generate a corresponding sender public key H for the sender private key S, H = g s , H∈G; And send the sender's public key H back to the receiving server; After sending the sender's public key H, the patient's public key set {C u }; The patient public key set {C u } by N U The patient's public key C u composition; And based on the hash function H KEY , the patient public key set {C u }, the sender's private key S and all the identification elements h v Generate the corresponding patient shared key K u,v , And each missing feature x in the first or second sample record corresponding to the vth social identity identifier in the current sample set and the missing feature set X j The corresponding medical feature information is extracted as the corresponding missing feature information d v,j , if the sample records of the current sample set do not have the missing feature x j The corresponding medical feature information corresponds to the missing feature information d v,j Set to empty; and N corresponding to each social identity X The missing feature information d v,j The corresponding patient missing information set D v {d v,j }; And based on the preset check code algorithm, each patient missing information set D v {d v,j }Calculate the check code to get the corresponding check code MAC v ; And each patient missing information set D v {d v,j } and the corresponding verification code MAC v Form a corresponding patient data DM v ; And based on the shared key K of each patient u,v Encrypt the corresponding patient data DM v Get the corresponding ciphertext data ED u,v , ED u,v =f EN (D v ,K u,v ), f EN The encryption function of the preset symmetric encryption and decryption algorithm, f EN With f DE They are inverse operation functions; And the obtained N U ×N V The ciphertext data ED u,v The corresponding ciphertext matrix {ED u,v }Send back to the receiving server.

11. The data processing system of the medical model federated learning framework according to claim 1, characterized in that: The second institution server is specifically configured to, when performing a round of local training on the second local model based on the third sample set to obtain the corresponding institution parameter set ciphertext and sending it back to the central server: Performing a round of training on the initialized second local model based on the third sample set; At the end of this round of training, the current model parameter set of the second local model is recorded as parameter set W ' now , record the front plate base parameter set as parameter set W ’ old , and based on the parameter set W ' now 、W ’ old Calculate the latest differential parameter set △W ’ =W ' now -W ’ old ; And based on the public key For the differential parameter set △W ’ Encrypting to obtain the corresponding ciphertext of the mechanism parameter set; is a preset homomorphic encryption function for using the public key For the latest differential parameter set ΔW ’ Encryption; And send the current encrypted text of the mechanism parameter set to the central server through the corresponding first channel.

Images