A secure sharing method for telemedicine privacy data based on heterogeneous federated learning

Abstract

The present invention relates to the field of data processing technology, and in particular to a remote medical privacy data security sharing method based on heterogeneous federated learning, comprising the following steps: S10, a first terminal obtains a local model update based on local medical privacy data training, and signs the local model update based on an encryption algorithm, and sends the signature to a second terminal; S20, the second terminal verifies the signature, and after the verification, aggregates each local model update to obtain a global model, and decomposes the global model into a plurality of local models, one local model corresponds to one local model, and the local model is broadcasted to each first terminal accordingly, so that the first terminal obtains the local model update based on local medical privacy data training after receiving the local model; S10-S20 is executed cyclically until the global model converges. On the basis of federated learning, the present invention further enhances data security by signing the local model update before transmitting it.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method for securely sharing telemedicine privacy data based on heterogeneous federated learning. Background Art

[0002] Deep learning has been widely used in the medical field, for example, using pre-trained deep learning models to classify diseases. Traditional deep learning is centralized, consolidating all data in a single location and using deep learning models for centralized learning. This can lead to serious privacy and security issues. For example, institutions such as finance and healthcare, which hold highly private and secure data, often refuse to transfer their data due to privacy and security concerns. This prevents data sharing and creates data silos.

[0003] Federated learning does not require each institution to transmit its own data. Instead, it only requires each institution to train the deep model locally using local private data, and then transmit the model parameters (which can be understood as the deep model itself) to the central institution. The central institution aggregates all local models into a global model through an algorithm. Therefore, federated learning can effectively solve the problem that private data is inconvenient to transmit. For example, the Chinese invention with publication number CN202311800401.0 discloses "A federated construction method for a medical task model and a medical image classification method", which is based on federated learning to construct a medical task model. It can share multi-party medical data while ensuring the privacy of medical data. However, this solution also has defects, that is, it is not enough to guarantee the security of medical data. In other words, the security of medical data needs to be further improved. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for securely sharing private telemedicine data based on heterogeneous federated learning, which can improve the security of data sharing.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides a method for securely sharing private telemedicine data based on heterogeneous federated learning, which is applied to a sharing system including a plurality of first terminals and a second terminal, and includes the following steps:

[0007] S100, the first terminal obtains a local model update based on local medical privacy data training, signs the local model update based on an encryption algorithm, and sends the signed data to the second terminal;

[0008] S200: The second terminal verifies the signature and, after verification, aggregates the updates of the local models to obtain a global model. The global model is then decomposed into a plurality of local models, each corresponding to a local model. The local models are broadcast to each first terminal, so that the first terminal can obtain the local model update based on the local medical privacy data after receiving the local model.

[0009] S100 - S200 are executed in a loop until the global model converges.

[0010] On the one hand, the above solution avoids the easy leakage of data caused by directly transmitting local medical privacy data through federated learning. On the other hand, by signing and authenticating local model updates, it avoids the problem of local model leakage during federated learning, further strengthening the security of privacy data sharing.

[0011] In a further optimized solution, in S100, the process of signing the local model update based on the encryption algorithm includes:

[0012] The first terminal sends a first terminal identifier to the third terminal and signature component g=e(P1,P spub ), so that the third terminal is based on the first terminal identifier Generate two private key shards and And based on cp h Generate r, b1, b2; rd1 is a random number, P1 is the generator of the cyclic group G1, P spub is the signature master public key, r represents the hash value of the message, and e is the number of bytes from G1×G2 to G T Bilinear pairings of ;

[0013] The first terminal receives the private key fragment d h i , the hash value r of the message, the signature components b1 and b2, generate the signature (r, s), s = d h i ·[b1+r1·b2].

[0014] This solution combines the national secret SM9 algorithm with threshold cryptography for signature authentication, addressing issues of unauthorized access and data leakage in telemedicine and protecting the security of telemedicine systems and data. This decentralized authentication mechanism requires collaboration between the primary and secondary terminals to construct a complete key. Cryptographic operations cannot be performed with only the key of either terminal. This enables trustless key distribution and management in a distributed environment, implementing bidirectional permission restrictions for both primary and secondary terminals while safeguarding the interests of both parties.

[0015] In a further optimized solution, the third terminal is based on the first terminal identifier ID h i Generate two private key shards and And based on cp h The process of generating r, b1, and b2 includes:

[0016] q2=q1 -1 ·sk, and rd2, rd3, rn and sk are random numbers, N is the order of the cyclic group G1, H is the hash function, mod represents the modular operation, cp t1 、 All are intermediate parameters.

[0017] Accordingly, in S200, the second terminal verifies the signature, including:

[0018] The second terminal verifies whether r∈[1,N-1] and s∈G1 hold;

[0019] The second terminal requests the third terminal to calculate F=P spub +[r1]P2,l=e(s,F),g=e(P1,P spub ), m=g r , cp k =l·m,

[0020] Verify whether r2 and r are equal. If they are equal, the verification passes, otherwise the verification fails.

[0021] In a more optimized solution, in S200, the processing of aggregating the updates of the local models to obtain the global model after the verification is passed includes:

[0022] Align the updated filters of each local model according to a fixed number of dimensions and fill the empty positions with zero values;

[0023] The aligned local models are updated and aggregated at corresponding positions in the global model according to the model mapping matrix, wherein the positional relationship of each local model in the global model is recorded in the model mapping matrix.

[0024] Because the capacity of each medical institution varies, as does the computing and storage capabilities of the primary terminal, the structures of each local model will also vary. The above solution effectively addresses the problem of heterogeneous local models by first aligning the local models according to a fixed pattern to form a unified structure, and then aggregating the unified local models into the global model.

[0025] In a more optimized solution, before S100, step S000 is further included: the second terminal selects a first terminal participating in this round of training from all first terminals; the first terminal in S100 is the first terminal selected in S000.

[0026] Different applications have different requirements for local models, so federated learning isn't always suitable for all medical institutions. In the aforementioned solution, the second terminal selects the first terminal participating in the training before initiating the training task. This targeted selection of federated learning participants makes the resulting global model more valuable and stable, avoiding training failures or reduced predictive power due to insufficient resources on one or more participants.

[0027] In a more optimized solution, in S100, the step of the first terminal obtaining a local model update based on local medical privacy data training includes:

[0028] The first terminal preprocesses the acquired local medical privacy data, wherein the preprocessing includes data cleaning, data normalization, and data enhancement;

[0029] The first terminal uses the pre-processed local medical privacy data for training to obtain a local model update.

[0030] In the above solution, the collected local medical privacy data is first preprocessed to avoid obstruction of local model training or impact on the predictive ability of the local model due to missing or abnormal data.

[0031] In a second aspect, the present invention provides a computer program product comprising computer-readable instructions, characterized in that when the computer-readable instructions are executed by a processor, the steps of the method for securely sharing telemedicine privacy data based on heterogeneous federated learning of the present invention are implemented.

[0032] In a third aspect, the present invention provides a computer-readable storage medium comprising computer-readable instructions, characterized in that when the computer-readable instructions are executed by a processor, the steps of the method for securely sharing telemedicine privacy data based on heterogeneous federated learning of the present invention are implemented.

[0033] In a fourth aspect, the present invention provides an electronic device comprising: a memory for storing program instructions; a processor connected to the memory for executing the program instructions in the memory to implement the steps of the present invention's method for securely sharing telemedicine privacy data based on heterogeneous federated learning.

[0034] Compared with the prior art, the present invention has the following technical advantages:

[0035] (1) On the one hand, federated learning avoids the data leakage caused by directly transmitting local medical privacy data. On the other hand, by signing and authenticating the local model updates, the problem of local model leakage during the federated learning process is avoided, further strengthening the security of privacy data sharing.

[0036] (2) The SM9 threshold authentication scheme is proposed, which is a decentralized authentication mechanism. It requires the medical institution and the cryptographic institution to work together to construct a complete key. Only the key of either party can perform cryptographic operations, thereby realizing a trustless key distribution and management in a distributed environment and implementing two-way authority restrictions on medical institutions and key institutions.

[0037] (3) It removes the restriction that all participants in classic federated learning need to use the same model, allowing each medical institution to determine a personalized local model based on its own capabilities, and using a heterogeneous aggregation algorithm to aggregate local models of different structures and scales into a global large model, which promotes fairness in the field of telemedicine and allows medical institutions with different "capacities" to benefit from the reuse of medical privacy data. Compared with other heterogeneous federated learning solutions, the solution of the present invention allows medical institutions to specify their local models independently, and only needs to record their structures by the competent authorities after initialization. There is no need to specify or restrict the structure and scale of each institution's local model, which greatly enhances the autonomy of medical institutions, helps them design personalized models that are more suitable for their "capacity", and enhances their control over their local models, increasing their sense of security and trust. In addition, smaller local models can also reduce the number of parameters when transferring model parameters, reduce possible communication bottlenecks, and are more conducive to medical institutions with lagging in information construction and low bandwidth to participate in the federated learning process.

[0038] For other advantages of the present invention, please refer to the relevant description in the embodiment section. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 This is an architectural diagram of a telemedicine privacy data security sharing system based on heterogeneous federated learning in an embodiment of the present invention.

[0041] Figure 2 This is a detailed flow chart of the method for securely sharing private telemedicine data based on heterogeneous federated learning provided in an embodiment of the present invention.

[0042] Figure 3 Schematic diagram of the model mapping matrix in an embodiment of the present invention.

[0043] Figure 4 Flowchart of signature verification in an embodiment of the present invention.

[0044] Figure 5 This is a simplified flowchart of the method for securely sharing private telemedicine data based on heterogeneous federated learning provided in an embodiment of the present invention.

[0045] Figure 6 It is a block diagram of the composition of the electronic device provided in the embodiment. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0047] See also Figure 1 In this embodiment, a hardware system for secure sharing of telemedicine privacy data based on heterogeneous federated learning is provided, including several first terminals, a second terminal and a third terminal. The first terminal is a terminal device located in a medical institution, the second terminal is a terminal device located in a competent authority, and the third terminal is a terminal device located in a key institution. The terminal device can be any electronic device capable of performing data processing and remote communication, such as a desktop computer, a laptop computer, a PDA, or even an electronic device such as a smart phone.

[0048] See Figure 2 This embodiment provides a method for securely sharing private telemedicine data based on heterogeneous federated learning, including the following steps:

[0049] System initialization step S1: The second terminal initializes the software system and sets the security parameters to (n, t) = (2, 2). This means the key is split into two parts. Only by collecting these two key fragments can a correct signature be constructed.

[0050] Institutional registration step S2: The first terminal sends the first terminal identifier to Then, the third terminal generates a key pair and splits the private key into private key fragments. and Stored separately by the first terminal and the third terminal. In order to generate key fragments, the third terminal needs to generate the master public key P spub And the function identifier hid used to generate the private key, and calculate q1, q2, and Subsequently, the third terminal shares the private key fragment with the first terminal through a secure channel At the same time, the third terminal stores the registration information in a list form.

[0051] See Figure 4 The first terminal will negotiate the communication key with the second terminal and the third terminal in advance, and all key information in the subsequent process will be transmitted through the encrypted secure communication channel.

[0052] More specifically, the third terminal uses the first terminal identifier The process of generating a private key shard is as follows:

[0053] 01: The third terminal generates a random number sk∈[1,N-1]; where sk is the signature master private key, and N is the cyclic group G1, G2 and G T A cyclic group is a group with a specific algebraic structure, where all elements in the group can be represented by the power of an element.

[0054] 02: The third terminal calculates P in G1 spub =[sk]P2; where P spub is the signature master public key, G1 is an additive cyclic group of order prime number N, and P2 is the generator of the cyclic group G2.

[0055] 03: The third terminal saves sk and publishes P spub .

[0056] 04: The third terminal generates a HID. HID is the private key generation function identifier, represented by a single character. Different private keys are generated depending on the HID value. For example, when HID is 1, a signature private key is generated; when HID is 3, an encryption private key is generated.

[0057] 05: The third terminal calculates the finite field F N in H is the hash function, mod represents the modular operation, All are intermediate parameters.

[0058] 06: The third terminal generates a random number rn.

[0059] 07: Third terminal calculation and

[0060] Telemedicine practice step S3: Medical institutions provide telemedicine services to patients in telemedicine practice and accumulate medical privacy data, which are collected and recorded as x i This step can also be understood as the collection of local medical privacy data. For example, electronic medical records, medical imaging data, and laboratory test data. Electronic medical records include basic patient information, medical history, diagnosis records, treatment plans, etc.; medical imaging data includes X-rays, CT scans, MRIs, and other medical imaging images; laboratory test data includes test results of blood, urine, and tissue samples.

[0061] Telemedicine data processing step S4: The first terminal processes x i Preprocessing is performed to facilitate subsequent steps. Data preprocessing mainly includes data cleaning, data normalization, and data enhancement. Data cleaning includes supplementing missing values, removing outliers, and correcting inconsistent or erroneous data, ensuring the quality and integrity of the data and laying the foundation for subsequent processing. The purpose of data normalization is to ensure that different features can influence the model fairly. The feature values ​​need to be normalized to the same standard range, such as [0,1] or [-1,1]. This helps prevent certain features from dominating the training process due to their large value range. The main operation of data enhancement is to transform the data, such as rotation, scaling, shearing, etc., to increase the diversity of the data and improve the generalization ability of the model. The data set of all medical institutions is represented as X.

[0062] Training task initialization step S5: The second terminal initiates a training task as needed and generates a global model w based on the model mapping matrix M t ,The model mapping matrix M records the positions of all local models in the ,global model, and is generated by the second terminal when the ,software system is initialized. t is the global model, which is a superset of all models and includes all local models. The height and width of the global model are equal to the maximum value of each layer of all local models.

[0063] Participant Selection Step S6: The second terminal selects the first terminal participating in this round according to a specific logic. The selection logic can be an average selection or selection based on the capabilities of the first terminal. This step includes three key parameters: 1) the number of first terminals; 2) the participation rate of first terminals; and 3) the number of first terminals participating in this round. Let n represent the number of first terminals and c represent the participation rate of first terminals, then the number of first terminals participating in this round is t = c × n. The participant selection scheme is determined based on the participant's strength relative to the task domain. For example, assume that medical data from higher-ranked medical institutions is relatively more valuable. Registered institutions are scored for the task domain, with stronger institutions receiving higher scores. A Softmax operation is then performed on the scores, mapping them to a probability distribution between 0 and 1. The resulting mapped probabilities are used as the selection probabilities at the start of the task.

[0064] For a specific task, selecting the first terminal with strong capabilities in the task field to participate in federated learning will help improve the predictive ability and stability of the final global model.

[0065] Global sub-model extraction step S7: The second terminal generates a global sub-model based on the first terminal participating in this round and the model mapping matrix M The global submodel is the global model W t The subset of is responsible for carrying the knowledge aggregated from each local model in this round; when c = 1, When c≠1,

[0066] In other words, the structure of the local model to be trained by each first terminal is predetermined, and the model mapping matrix M is determined during system initialization. Therefore, once the second terminal selects the first terminal participating in this round of federated learning training, the structure of the global sub-model is determined. For each round of federated learning, the global sub-model is actually the global model for that round, and can therefore also be understood as the global model.

[0067] Model Pruning and Broadcasting Step S8: The knowledge in the global model is back-propagated to the local model in each local model. This process is not limited to the first terminals participating in this round of aggregation, so each first terminal can benefit from this process. In this process, the second terminal decomposes the global model into local models corresponding to each first terminal based on the model mapping matrix M (this process is generally called model pruning) and transmits it to each first terminal via an encrypted communication channel. After receiving the decomposed local model, the first terminal applies it to the local model, completing the feature regression process and benefiting from this process.

[0068] The second terminal "crops" the global sub-model into a shape that adapts to each local model according to the mapping matrix M. The model mapping matrix M is the core of the global aggregation of heterogeneous models, recording the position of each local model in the global model, that is, which part of the global model the model parameters of the local model correspond to. The deep model is regarded as a set of matrices carrying model parameters, and the Cartesian coordinate system is used to identify the relative position of the filters in the local model in the global model. Figure 3 , the first terminal uses the Cartesian coordinate system as a reference to align the local model to calibrate and demarcate the relative position of each filter therein. The model mapping matrix M records the mapping relationship between the relative positions of the local model and the global model. The first element records which layer of the global model the filter is located in will act on, that is, the "X-axis" in the Cartesian coordinate system corresponds to the height dimension of the depth model; the second element records which filter of the corresponding layer in the global model the filter will act on, that is, the "Y-axis" in the Cartesian coordinate system corresponds to the width dimension of the depth model. The global aggregation of heterogeneous models can be regarded as the superposition of all local models in the third dimension. In this process, filters at the same position are aggregated into that position. Both the global model and the local model are deep learning models (or deep models for short). The deep model is composed of filters distributed in several layers. Therefore, local model aggregation is achieved by aggregating filters at corresponding positions.

[0069] After determining the structure of the global model, the corresponding local models in the global model are broadcast to all first terminals to obtain model parameters. The purpose of broadcasting the global model is to allow the first terminals to obtain the relevant parameters of their corresponding local models. Therefore, broadcasting the global model here can also be understood as broadcasting the model parameters of the global model.

[0070] By decomposing the entire global model into corresponding local models and distributing them to all first terminals, that is, only sending the corresponding local model to the first terminal, the leakage of other model parameters that do not belong to the local model can be avoided, and these parameters that do not belong to the local model can be avoided from interfering with the training and updating of the local model.

[0071] Global model parameter reception step S9: The first terminal receives the local model decomposed from the global model and applies it to the local model. This means that the global model output from the previous training round overwrites the local model from the previous round and serves as the input for the next training round. This step allows the knowledge accumulated in the global model from the previous round to be returned to the local model. If this is the first round of federated learning, the global model corresponding to the layout model received by the first terminal is the initially initialized global model.

[0072] Local training step S10: Based on the pre-processed local medical privacy data, a local model update is obtained by training. The pre-processed local medical privacy data is used to perform further gradient descent on the local model to form a gradient update. That is, for each round of training for each first terminal, the following process occurs: w represents the parameters of the local model, which can also be understood as the local model, η represents the learning rate, l(w; b) represents the loss function, Represents the gradient of the loss function l with respect to the parameter w.

[0073] Local model signing step S11: The first terminal uses the private key to shard Sign the local model update and generate a signature (r,s). To generate the signature, the first terminal and the third terminal need to use and To generate the signature component, the first terminal then calculates the signature (r, s).

[0074] See Figure 4 ,More specifically, the model signature process is as follows:

[0075] 01: The first terminal calculates G T g=e(P1,P spub ), and generate a random number rd1∈[1,N-1]. T is a multiplicative cyclic group of order prime number N, g is G T The elements in are the intermediate values ​​of the calculation, e is the value from G1×G2 to G T Bilinear pairing, P1 is the generator of cyclic group G1, P spub The master public key for signature.

[0076] 02: The first terminal calculates the signature component

[0077] 04: Third terminal calculation Where r represents the hash value of the message and H is the hash function.

[0078] 05: The third terminal calculates the signature component and

[0079] 06: The first terminal receives r, b1 and b2.

[0080] 07: First Terminal Computing Then we get the signature (r, s).

[0081] Local model parameter receiving step S12: The first terminal signs the local model update and sends it to the second terminal. The second terminal then receives the local model updates sent by each first terminal. The purpose of receiving the local model updates is to obtain the parameters of the updated local model. Therefore, receiving the local model updates here can also be understood as receiving the updated local model parameters.

[0082] The local model update that has not been uploaded within the set time will be abandoned. The return timeout indicates that the first terminal cannot obtain the local model update for some reason, or the local model update However, for some reason, the local model update cannot be transmitted to the second terminal.

[0083] Signature Verification Step S13: After receiving the signature, the second terminal verifies the signature. To verify the local model update, the second terminal needs to verify the legitimacy of r and s, calculate r2 using r, and compare it with r to obtain the verification result.

[0084] More specifically, the signature verification process is as follows:

[0085] 01: The second terminal verifies that r∈[1,N-1] and s∈G1 hold. The following steps are executed only if both conditions hold. If not, the two values ​​are theoretically illegal and no further verification is required; the verification is directly judged as failed.

[0086] 03: The second terminal requests the third terminal

[0087] 04: The second terminal calculates g=e(P1,P spub ) and m=g r .

[0088] 05: Second terminal calculation F=P spub +[r1]P2,l=e(s,F).

[0089] 06: The second terminal calculates cp k =l·m,

[0090] 07: The second terminal verifies whether r2 and r are equal. If they are equal, the verification succeeds; otherwise, the verification fails.

[0091] Heterogeneous model global aggregation step S14: The second terminal aligns the filters updated by each local model according to a fixed number of dimensions and fills the empty positions with zero values; according to the model mapping matrix M, the aligned local model updates are aggregated at the corresponding positions in the global model.

[0092] Global aggregation is the process of aggregating several local models with different structures and scales into a consistent global large model. The global aggregation process of heterogeneous models can be regarded as a traversal process performed according to the global model. The second terminal traverses all local model updates at each position of the global model and aggregates the filters existing at that position in all local model updates into the global model. In this process, there may be situations where the filter structures and scales are different. Here, an alignment method is used to deal with it. That is, the filters (generally regarded as matrices of the same dimension) are aligned according to a fixed number of dimensions (just like the operation in the model mapping matrix M), and the empty positions are filled with zero values. The global filters to be aggregated are filled with the same scale and then aggregated. During aggregation, the amount of data used by the first terminal corresponding to each local model will be used as a weight to determine the importance of each local model. Through aggregation, all knowledge carried in the local models (that is, the data "texture" captured by the filters) will be gathered into the global model.

[0093] Global model evaluation step S15: The second terminal estimates the global model w t , if w t If the model has not converged, the process starts from the participant selection step S6 and repeats S6-S14 until the global model converges. t Has converged, the second terminal to the global model w t The trimming is performed, and each local model obtained after the trimming is distributed to all the first terminals, and the heterogeneous federated learning process is completed.

[0094] After learning, medical institutions can use local models to implement smart healthcare applications and improve telemedicine practices. Smart healthcare applications can include assisted diagnosis and treatment, assisted disease prediction, assisted drug development, clinical decision support, assisted remote monitoring, automated document processing, and health management. For example, assisted diagnosis and treatment uses deep models to automatically analyze medical images to help doctors diagnose diseases faster and reduce the possibility of misdiagnosis. Assisted disease prediction uses deep models to analyze patients' electronic health records to predict disease trends and patient risk levels. Assisted drug development uses deep models to analyze chemical compounds and their bioactivity data to predict potentially effective drug candidates, accelerating the development of new drugs, shortening drug development cycles, and reducing R&D costs. Clinical decision support uses deep models to analyze clinical data, provide insights into patient conditions, recommend treatment options, and assist doctors in making diagnostic and treatment decisions. Assisted remote monitoring uses deep models to analyze health data from wearable devices and mobile apps to monitor patients' health in real time and alert doctors to take necessary treatment actions. Automated document processing uses deep models to automatically process medical documents, extract key information, and improve the efficiency of medical institutions. Health management uses deep models to analyze people's health data and lifestyle habits from wearable devices and mobile applications to provide users with personalized health advice and lifestyle improvement plans.

[0095] Figure 2 The process shown is from system initialization to the final completion of the entire federated learning process. From the perspective of practical application, please refer to Figure 5 ,The secure sharing method of telemedicine privacy data based on heterogeneous federated learning can also be simplified and summarized into the following steps:

[0096] S100, the first terminal obtains a local model update based on local medical privacy data training, signs the local model update based on an encryption algorithm, and sends the signed data to the second terminal;

[0097] S200: The second terminal verifies the signature and, after verification, aggregates the updates of the local models to obtain a global model. The global model is then decomposed into a plurality of local models, each corresponding to a local model. The local models are broadcast to each first terminal, so that the first terminal can obtain the local model update after receiving the local model and training based on the local medical privacy data.

[0098] S100 - S200 are executed in a loop until the global model converges.

[0099] like Figure 6As shown, this embodiment also provides an electronic device, which may include a processor 41 and a memory 42, wherein the memory 42 is coupled to the processor 41. It is worth noting that this figure is exemplary, and other types of structures may be used to supplement or replace this structure to implement data extraction, report generation, communication or other functions.

[0100] like Figure 6 As shown, the electronic device may further include: an input unit 43, a display unit 44 and a power supply 45. It is worth noting that the electronic device does not necessarily have to include Figure 6 In addition, electronic devices may also include Figure 6 For components not shown, reference may be made to the prior art.

[0101] The processor 41 is sometimes also called a controller or an operation control, and may include a microprocessor or other processor devices and / or logic devices. The processor 41 receives inputs and controls the operations of various components of the electronic device.

[0102] The memory 42 may be, for example, one or more of a cache, flash memory, a hard drive, a removable medium, a volatile memory, a non-volatile memory, or other suitable devices, and may store information such as configuration information of the processor 41 and instructions executed by the processor 41. The processor 41 may execute programs stored in the memory 42 to implement information storage or processing. In one embodiment, the memory 42 also includes a buffer memory to store intermediate information.

[0103] The input unit 43 is used to input MRI images to the processor 41. The display unit 44 is used to display various results of the processing process. The display unit can be, for example, an LCD display, but the present invention is not limited thereto. The power supply 45 is used to provide power to the electronic device.

[0104] An embodiment of the present invention further provides a computer program product, comprising computer-readable instructions. When the computer-readable instructions are executed in an electronic device, the program product enables the electronic device to perform the operating steps included in the method of the present invention.

[0105] An embodiment of the present invention further provides a storage medium storing computer-readable instructions, wherein the computer-readable instructions enable an electronic device to execute the operation steps included in the method of the present invention.

[0106] Those skilled in the art will appreciate 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 composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel 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.

[0107] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0108] Those skilled in the art will appreciate that the modules of the various examples 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 constituent modules and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel 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.

[0109] In the several embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the module division is merely a logical functional division. In actual implementation, other division methods may be used, such as combining or integrating multiple modules or components into another system, or ignoring or not implementing certain features.

[0110] The above-described embodiments are merely specific implementations of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications, substitutions, and improvements within the technical scope disclosed in the present invention, and such modifications, substitutions, and improvements are intended to be encompassed within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for securely sharing private telemedicine data based on heterogeneous federated learning, applied to a sharing system including a plurality of first terminals and a second terminal, characterized in that: The following steps are involved: S100, the first terminal obtains a local model update based on local medical privacy data training, signs the local model update based on an encryption algorithm, and sends the signed local model to the second terminal; the structure of the local model to be trained by each first terminal is predetermined, and the structure of each local model is different; S200, the second terminal verifies the signature, and after the verification is passed, aggregates the updates of each local model to obtain a global model, and decomposes the global model into a plurality of local models, one local model corresponds to one local model, and the local models are broadcasted to each first terminal accordingly, so that after the first terminal receives the local model, it can obtain the local model update based on the local medical privacy data training; After the verification is passed, the local model updates are aggregated to obtain the global model, which includes: aligning the filters of the local model updates according to a fixed number of dimensions and filling the empty positions with zero values; aggregating the aligned local model updates at corresponding positions in the global model according to the model mapping matrix, wherein the model mapping matrix records the positional relationship of each local model in the global model and is recorded and generated by the second terminal when the software system is initialized; S100-S200 are executed cyclically until the global model converges; In S100, the process of signing the local model update based on the encryption algorithm includes: The first terminal sends a first terminal identifier to the third terminal and signature component g=e(P1,P spub ), so that the third terminal can use the first terminal identifier Generate two private key shards and And based on cp h Generate r, b1, b2; rd1 is a random number, P1 is the generator of the cyclic group G1, P spub is the signature master public key, r is the hash value of the message, and e is the T Bilinear pairings of ; The first terminal receives the private key fragment The hash value r of the message and the signature components b1 and b2 generate the signature (r, s). The third terminal is based on the first terminal identifier Generate two private key shards and And based on cp h The process of generating r, b1, and b2 includes: and rd2, rd3, rn and sk are random numbers, N is the order of the cyclic group G1, H is the hash function, mod represents the modular operation, All are intermediate parameters; In S200, the second terminal verifies the signature, including: The second terminal verifies whether r∈[1,N-1] and s∈G1 hold; If so, the second terminal requests the third terminal to calculate F=P spub +[r1]P2,l=e(s,F),g=e(P1,P spub ), m=g r , cp k = l·m, Verify whether r2 and r are equal. If they are equal, the verification passes, otherwise the verification fails.

2. According to claim 1, the method for secure sharing of telemedicine privacy data based on heterogeneous federated learning is characterized in that: Before S100, the process further includes step S000: the second terminal selects a first terminal participating in this round of training from all the first terminals; the first terminal in S100 is the first terminal selected in S000.

3. The method for securely sharing telemedicine privacy data based on heterogeneous federated learning according to claim 1 is characterized in that: In S100, the step of obtaining a local model update based on local medical privacy data training by the first terminal includes: The first terminal preprocesses the acquired local medical privacy data, wherein the preprocessing includes data cleaning, data normalization and data enhancement; The first terminal uses the preprocessed local medical privacy data for training to obtain a local model update.

4. A computer program product comprising computer readable instructions, characterized in that: When the computer-readable instructions are executed by a processor, the steps in the method for securely sharing telemedicine privacy data based on heterogeneous federated learning described in any one of claims 1 to 3 are implemented.

5. A computer-readable storage medium comprising computer-readable instructions, characterized in that: When the computer-readable instructions are executed by a processor, the steps in the method for securely sharing telemedicine privacy data based on heterogeneous federated learning described in any one of claims 1 to 3 are implemented.

6. An electronic device, characterized in that: include: Memory, which stores program instructions; A processor is connected to the memory and executes program instructions in the memory to implement the steps in the method for secure sharing of telemedicine privacy data based on heterogeneous federated learning as described in any one of claims 1 to 3.

Images