Federal learning optimization training method for multi-source heterogeneous privacy data

Abstract

This invention relates to the field of federated learning optimization technology, specifically a federated learning optimization training method for multi-source heterogeneous privacy data. The method includes: initializing the federated learning system and distributing unified global model parameters and model structure to each data holder. Each data holder, relying on an improved local update algorithm, adaptively adjusts the iterative update step size based on local data distribution differences, generating model parameter update amounts carrying integrity verification information. The parameter update amounts from multiple parties are collected and homomorphically encrypted and aggregated. After secure decryption, the global model is updated, and the iteration progress is determined through performance verification on a public dataset. This method can adapt to the differences in the distribution of multi-source heterogeneous data, balance the local training pace in different regions, and simultaneously achieve parameter aggregation processing under encrypted conditions, enhancing the security of federated learning parameter interaction and meeting the application requirements of collaborative training with privacy data.

Description

Technical Field

[0001] This invention relates to the field of federated learning optimization technology, and in particular to a federated learning optimization training method for multi-source heterogeneous privacy data. Background Technology

[0002] Federated learning enables collaborative modeling of privacy-preserving data across multiple nodes and is widely used in scenarios involving joint training of multi-source heterogeneous data. In a conventional federated learning training process, each data holder uses a fixed step size to iteratively update its local model parameters. When aggregating and integrating the parameter updates from multiple parties, a plaintext-based aggregation and fusion processing mode is commonly adopted to achieve iterative optimization of the global model.

[0003] The privacy data collected and stored by different institutions exhibits significant distributional differences. A uniform update mode with a fixed step size cannot match the differentiated data distribution characteristics of each node. Heterogeneous data can cause abnormal iteration speeds for some nodes and deviations in the local training fit of the model. The plaintext parameter aggregation processing mode exposes the updated parameters of multiple models directly during transmission and integration, making it easy for model information associated with privacy data to be leaked, and weakening the security protection capabilities of data collaborative training.

[0004] By adjusting the local parameter update rules based on the distributed differences of multi-source heterogeneous privacy data, and introducing an encrypted operation mechanism in the multi-party model parameter aggregation stage, the shortcomings of traditional federated learning in adaptive training control and parameter interaction security protection are filled, thus adapting to the joint training application needs of complex heterogeneous privacy data. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a federated learning optimization training method for multi-source heterogeneous privacy data.

[0006] To achieve the above objectives, the present invention employs the following technical solution: a federated learning optimization training method for multi-source heterogeneous privacy data, comprising: Initialize the federated learning system by sending unified global model parameters and model structure definitions to multiple data holders participating in the federated learning process; Each data holder trains the received global model parameters based on their locally stored private dataset using an improved model parameter local update algorithm. The improved model parameter local update algorithm adaptively adjusts the update step size of each iteration according to the degree of difference in the local data distribution of each data holder, and generates model parameter update amount with local training process integrity verification information. Collect the model parameter update amounts from all data holders, perform homomorphic encrypted aggregation on all the model parameter update amounts, and form an encrypted aggregated model parameter update; A secure decryption operation is performed on the aggregated model parameter update of the encrypted state to obtain the decrypted aggregated model parameter update; The global model parameters generated in the previous iteration are updated using the aggregated model parameters to obtain the global model parameters for the new iteration. Verify the performance of the new round of global model parameters on standard public datasets, and decide whether to start a new round of federated learning iterations based on the verification results.

[0007] As a further aspect of the present invention, the improved model parameter local update algorithm adaptively adjusts the update step size of each iteration based on the degree of difference in the local data distribution of each data holder, including: At the start of each local training iteration, the gradient of the loss function of the current local model parameters on a random sampled batch of the local private dataset is calculated. The covariance matrix of the feature vectors of all samples in the local private dataset is calculated, and the ratio of the largest eigenvalue to the smallest eigenvalue of the covariance matrix is ​​calculated. The ratio is used as a local heterogeneity index to characterize the degree of difference in the distribution of local data. An adaptive coefficient calculation rule that is negatively correlated with the local heterogeneity index is introduced. Based on the current iteration round and the local heterogeneity index, a positive number less than one is dynamically calculated as the adaptive step size coefficient for this iteration. The calculated gradient of the loss function is multiplied by the adaptive step size coefficient to obtain the gradient value after step size adjustment. Using the gradient values ​​adjusted by the step size, combined with momentum optimization techniques, update the current local copy of the model parameters to complete one local iteration update; After all preset local iterations are completed, the difference vector between the initial local model parameters and the final local model parameter copy is calculated. The difference vector is used as the model parameter update amount, and the norm sum of squared gradients of all loss functions calculated during local training is added to the model parameter update amount as part of the local training process integrity verification information.

[0008] As a further aspect of the present invention, each data holder trains the received global model parameters based on a locally stored private dataset using an improved local model parameter update algorithm, and generates model parameter update quantities with local training process integrity verification information, including: Upon receiving the global model parameters, each data holder loads them as the initial state of the local model parameters. During local training, the computation timestamp, computational resource consumption metric, and digest hash value of the final generated model parameter update amount are recorded each time the improved model parameter local update algorithm is used to update the parameters. The computation timestamp, the computational resource consumption metric, and the digest hash value, together with the norm sum of squares of the loss function gradient obtained from the improved local model parameter update algorithm, are jointly encoded into structured local training process integrity verification information. A differential privacy mechanism is applied to the model parameter update amount by adding random noise conforming to a specific Gaussian distribution to the difference vector to generate a noisy model parameter update amount that meets the differential privacy requirements. The updated parameters of the noisy model are digitally signed and packaged with the structured integrity verification information of the local training process to form a complete update data package ready for uploading.

[0009] As a further aspect of the present invention, the step of collecting the model parameter update amounts from all data holders, and homomorphically encrypting and aggregating all the model parameter update amounts to form an encrypted aggregated model parameter update includes: During the initialization phase of the federated learning system, public-private key pairs generated based on a homomorphic encryption scheme are distributed to all participating data holders and the central server. Each data holder uses the public key of the central server to perform homomorphic encryption encoding on each floating-point parameter in the noisy model parameter update, generating an encrypted model parameter update vector. After receiving the encrypted model parameter update vectors from all data holders, the central server performs element-wise addition on all the encrypted model parameter update vectors based on the additive homomorphic property of homomorphic encryption. After completing the element-by-element addition operation, the central server obtains a result vector with the same dimension as the single encrypted model parameter update vector. The result vector is the aggregated model parameter update of the encrypted state, and its content is the ciphertext of the sum of the update amounts of all data holders. During the aggregation process, the central server does not perform any decryption operations and cannot parse any original update information of individual data holders from the encrypted aggregated model parameter updates.

[0010] As a further aspect of the present invention, a secure decryption operation is performed on the aggregated model parameter update of the encrypted state to obtain the decrypted aggregated model parameter update, including: After completing the homomorphic encryption aggregation, the central server updates the aggregated model parameters of the encrypted state and sends them to one or more authorized decryption nodes. The decryption node uses a pre-allocated and securely stored decryption private key to perform decryption calculations on each ciphertext element of the aggregated model parameter update of the encrypted state; The decryption calculation generates a plaintext vector with the same dimension as the updated parameters of the aggregated model after the encryption state. Each element of the plaintext vector corresponds to the algebraic sum of the update amounts of all data holders on the corresponding parameters of the model. Verify whether the format and dimension of the plaintext vector obtained by decryption meet the expectations, and define the plaintext vector that passes the verification as the updated parameters of the decrypted aggregation model; The decrypted aggregated model parameters are updated and sent back to the central server for subsequent global model update steps.

[0011] As a further aspect of the present invention, the global model parameters generated in the previous iteration are updated using the aggregated model parameter update to obtain a new round of global model parameters, including: After receiving the decrypted aggregation model parameter update, the central server counts the number of valid data holders who successfully participated in the aggregation in this round. Divide each element of the decrypted aggregate model parameter update by the number of valid data holders to calculate the average value of the model parameter update amount for all participants in this round; A weighted average update rule is adopted, which weights and fuses the global model parameters generated in the previous iteration with the average value of the calculated model parameter update. The sum of the weight coefficients used for the global model parameters in the previous iteration and the weight coefficients used for the average value of the model parameter update is one. Perform weighted fusion calculation to generate a new set of model parameter values, which are the global model parameters for the new round; The global model parameters of the new round are compared with those of the global model parameters generated in the previous iteration. If the change in parameters is less than the preset convergence threshold, the global model is marked as having reached convergence.

[0012] As a further aspect of the present invention, the step of verifying the performance of the new round of global model parameters on a standard public dataset, and deciding whether to initiate a new round of federated learning iterations based on the verification results, includes: After receiving the new round of global model parameters, the central server loads them into a validation model with the same structure as the federated learning model. The validation model is run on a public benchmark dataset that is independent of the private datasets of all data holders, and the accuracy metrics and loss function values ​​of the model on the test task are calculated. The calculated accuracy metric and loss function value are compared with the pre-recorded historical best accuracy metric and historical lowest loss function value. If the current accuracy metric is better than the historical best accuracy metric, or the current loss function value is lower than the historical lowest loss function value, then the model performance is determined to be improved. The new round of global model parameters is archived as the current best global model parameters, and the instruction is broadcast to all data holders to start a new round of federated learning iteration. If the current accuracy metric is not better than the historical best accuracy metric and the current loss function value is not lower than the historical lowest loss function value, then the model performance is determined not to have improved. The central server will then pause initiating a new round of iterations and analyze the trend of model parameter changes to determine the subsequent strategy.

[0013] As a further aspect of the present invention, before initiating a new round of federated learning iterations, the method further includes a dynamic participant selection step based on model performance feedback: The central server evaluates the quality and reliability of each data holder's contribution based on the local training process integrity verification information submitted by each data holder in the historical multi-round federated learning. In specific evaluation, the changing trend of the norm sum of squares of the loss function gradients submitted by each data holder in history, the regularity of the calculated timestamps, and the consistency of the digest hash values ​​are analyzed. For data holders whose contribution quality and reliability assessment results are lower than the preset standards, they will be marked as pending observation in the invitation list for the next round of federated learning iterations; When forming the set of participants for the next iteration, data holders with high contribution quality and reliability assessment results are given priority. Data holders in the observation state are temporarily excluded from participation with a certain probability. The final list of participants is distributed along with the new set of global model parameters, thereby dynamically optimizing the composition of nodes participating in federated learning in each iteration.

[0014] As a further aspect of the present invention, the method further includes the step of inserting a model watermark during the federated learning process to verify model ownership: When initializing the federated learning system, the central server generates a unique secret watermark key associated with the model structure; After generating a new round of global model parameters in each round, the central server uses the unique secret watermark key and the current model parameters to calculate a hidden digital watermark signal through an irreversible watermark embedding function. The hidden digital watermark signal is superimposed onto a specific subset of the new round of global model parameters in a covert manner that does not affect the main function of the model, thereby generating watermarked global model parameters. The watermarked global model parameters will replace the original global model parameters and be sent to the data holder for the next round of local training, or released as the final model. When it is necessary to verify the ownership of the model, the corresponding watermark extraction function and the unique secret watermark key are used to attempt to extract the hidden digital watermark signal from the parameters of the model being verified. The ownership is confirmed by verifying whether the extracted signal matches the expectation.

[0015] As a further aspect of the present invention, the method also includes a fault-tolerance and recovery mechanism for handling data holders who exit midway: At the beginning of each federated learning iteration, the central server records a list of data holder identities confirmed to participate in that round. During the model parameter update collection phase, the central server sets a receiving timer for each data holder in the identity list; If the model parameter update of a data holder has not arrived after the receive timer expires, the data holder will be marked as a node that has exited this round, and a partial data summary of the data it has transmitted will be recorded. Using the model parameter updates from other data holders that have been successfully received, continue homomorphic encryption aggregation and subsequent decryption and global update operations; In the next round of iterative invitations, the central server will attempt to re-contact the nodes marked as withdrawn. If the re-contact is successful and the node provides reasonable proof of delay, its participation will be restored. Otherwise, it will be removed from the subsequent identity list, and its contributed parameter updates will be archived based on its historical contributions.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The model iteration update step size is adjusted in real time based on the degree of difference in local data distribution, abandoning fixed and uniform parameter update rules and adapting to the different private data characteristics of different data holders. This balances the pace of model iteration across data nodes, mitigating training bias caused by uneven heterogeneous data distribution and maintaining stable operation of the local model parameter iteration process. It reduces the negative impact of data distribution differences on the model training process, ensuring that the parameter update rhythm aligns with the operating characteristics of the local dataset, improving the adaptability of unified training modes, continuously stabilizing the independent training state of each node, smoothly advancing the multi-node distributed model training process, and reducing the phenomenon of training rhythm imbalance in heterogeneous data environments.

[0017] Homomorphic encryption aggregation is performed on all model parameter updates, ensuring that parameter fusion operations are conducted entirely in encrypted form, thus avoiding plaintext parameter transmission and integration. This reduces the scope of exposed core model parameters, lowers the potential risk of sensitive information leakage during parameter interaction, and strengthens information protection capabilities in multi-party collaborative training. Encrypted computation is used to complete multi-party parameter fusion, standardizing the operation of cross-node data interaction, improving the distributed federated learning process, enhancing the security level of joint training with heterogeneous privacy data, stabilizing the operating environment for multi-party collaborative model training, and optimizing information protection conditions in distributed training scenarios. Attached Figure Description

[0018] Figure 1 This is a flowchart of the federated learning optimization training method for multi-source heterogeneous privacy data described in this invention; Figure 2 A flowchart for generating model parameter update quantities with integrity verification information for the data holder; Figure 3 This is a flowchart illustrating the model parameter update process after homomorphic encrypted aggregation forms an encrypted state aggregation. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the 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 merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1 This invention provides a federated learning optimization training method for multi-source heterogeneous privacy data, the overall implementation scheme of which is as follows: The federated learning system is initialized by sending unified global model parameters and model structure definitions to multiple data holders participating in the federated learning process. Each data holder trains the received global model parameters using an improved local model parameter update algorithm based on its locally stored private dataset. This improved algorithm adaptively adjusts the update step size for each iteration based on the differences in the local data distribution of each data holder, and generates model parameter updates with information verifying the integrity of the local training process. Model parameter updates from all data holders are collected and homomorphically encrypted and aggregated to form an encrypted aggregated model parameter update. A secure decryption operation is performed on the encrypted aggregated model parameter update to obtain the decrypted aggregated model parameter update. The aggregated model parameter update is used to update the global model parameters generated in the previous iteration to obtain the new global model parameters. The performance of the new global model parameters on a standard public dataset is verified, and the decision to start a new round of federated learning iterations is based on the verification results.

[0022] In one embodiment of the present invention, see [reference] Figure 2 In the improved local model parameter update algorithm, the update step size of each iteration is adaptively adjusted based on the degree of difference in the local data distribution among data holders. The process involves calculating the gradient of the loss function of the current local model parameters on a random sample batch of the local private dataset at the beginning of each local training iteration. The covariance matrix of the feature vectors of all samples in the local private dataset is calculated, and the ratio of the largest to the smallest eigenvalue of this covariance matrix is ​​used as a local heterogeneity index to characterize the degree of difference in local data distribution. An adaptive coefficient calculation rule negatively correlated with the local heterogeneity index is introduced. Based on the current iteration and the local heterogeneity index, a positive number less than one is dynamically calculated as the adaptive step size coefficient for this iteration. The calculated loss function gradient and the adaptive step size coefficient are multiplied by a scalar to obtain the gradient value after step size adjustment. Using the gradient value after step size adjustment, combined with momentum optimization techniques, the current local model parameter copy is updated, completing one local iteration update. After all preset local iterations are completed, the difference vector between the initial local model parameters and the final local model parameter copy is calculated. This difference vector is used as the model parameter update amount, and the norm sum of squared gradients of all loss functions calculated during local training is added to the model parameter update amount as part of the local training process integrity verification information.

[0023] Each data holder, based on its locally stored private dataset, trains the received global model parameters using an improved local model parameter update algorithm, generating model parameter updates with local training process integrity verification information. The process involves each data holder loading the received global model parameters as the initial state of their local model parameters. During local training, the timestamp of each parameter update using the improved local model parameter update algorithm, the metric of consumed computational resources, and the digest hash value of the final generated model parameter update are recorded. These timestamps, computational resource metrics, and digest hash values, along with the norm sum of squared gradients of the loss function obtained from the improved local model parameter update algorithm, are encoded into structured local training process integrity verification information. Differential privacy is applied to the model parameter update, adding random noise conforming to a specific Gaussian distribution to the difference vector to generate a noisy model parameter update that meets differential privacy requirements. The noisy model parameter update and the structured local training process integrity verification information are digitally signed and packaged to form a complete update data package ready for upload.

[0024] In practice, multiple consecutive steps are involved. At the start of local training, each data holder calculates the gradient of the loss function on a random sample batch of their local model parameters within their local private dataset. This gradient calculation is based on the standard backpropagation algorithm. The covariance matrix of the feature vectors of all samples in the local private dataset is calculated, and the ratio of the largest to smallest eigenvalue is used as a local heterogeneity index, representing the degree of dispersion in the local data distribution. An adaptive coefficient calculation rule negatively correlated with the local heterogeneity index is introduced. A positive number less than one is dynamically calculated based on the current iteration and the local heterogeneity index as the adaptive step size coefficient for this iteration. This adaptive step size coefficient decreases as the local heterogeneity index increases. The calculated loss function gradient and the adaptive step size coefficient are multiplied by a scalar to obtain the step-size-adjusted gradient value, which is used for subsequent parameter updates. Finally, the step-size-adjusted gradient value is combined with momentum optimization techniques to update the current local model parameter copy, completing one local iteration update. Momentum optimization accelerates the convergence process. After all preset local iterations are completed, the difference vector between the initial local model parameters and the final local model parameter copy is calculated. The difference vector is used as the model parameter update amount. The norm sum of squared gradients of all loss function gradients calculated during local training is added to the model parameter update amount as part of the local training process integrity verification information. The norm sum of squared gradients of loss function gradients is used to verify the continuity of the training process.

[0025] In some embodiments, the adaptive coefficient calculation rule of the improved model parameter local update algorithm is specified by a mathematical formula, which is expressed as follows: in: Indicates the adaptive step size coefficient. This represents the basic learning rate hyperparameter. Represents the attenuation factor constant. The formula represents the local heterogeneity index, ensuring a negative correlation between the adaptive step size coefficient and the local heterogeneity index. The calculation rule for the adaptive step size coefficient relies on pre-defined base learning rate hyperparameters and decay factor constants. The base learning rate hyperparameters are uniformly configured during the initialization of the federated learning system, while the decay factor constant is adjusted according to the model type. In each iteration, the local heterogeneity index is substituted into the formula to calculate the adaptive step size coefficient, which is directly applied to the scaling of the loss function gradient.

[0026] In practice, each data holder uses an improved local model parameter update algorithm to train the received global model parameters based on their locally stored private dataset. This generates model parameter updates with local training process integrity verification information. Upon receiving the global model parameters, each data holder loads them as the initial state of their local model parameters. During local training, the timestamp of each parameter update using the improved local model parameter update algorithm, the computational resource consumption metric, and the digest hash value of the final generated model parameter update are recorded. The timestamp records the start and end times of each iteration, and the computational resource consumption metric includes memory usage and floating-point operations. The timestamp, computational resource consumption metric, and digest hash value, along with the norm sum of squared gradients of the loss function obtained from the improved local model parameter update algorithm, are encoded into structured local training process integrity verification information using a serialized binary string encoding format. A differential privacy mechanism is applied to the model parameter update by adding random noise conforming to a specific Gaussian distribution to the difference vector, generating a noisy model parameter update that satisfies differential privacy requirements. The mean and variance of the Gaussian distribution are set according to the privacy budget. The updated parameters of the noisy model are digitally signed and packaged with the structured local training process integrity verification information to form a complete update data package ready for upload. The digital signature is generated using the private key of the data holder.

[0027] In some embodiments, the effectiveness of the improved local update algorithm for model parameters is illustrated by combining an example scenario and data comparison. The example scenario is a medical image classification task, with multiple data holders from different hospitals, each hospital's local private dataset containing different categories of medical images. Hospital A's local private dataset shows a higher ratio of the maximum to minimum eigenvalues ​​of the covariance matrix of sample feature vectors, indicating a larger local heterogeneity index and significant differences in data distribution. Hospital B's local private dataset shows a lower ratio of the maximum to minimum eigenvalues ​​of the covariance matrix of sample feature vectors, indicating a smaller local heterogeneity index and a more uniform data distribution. When calculating the adaptive step size coefficient, Hospital A's larger local heterogeneity index leads to a smaller adaptive step size coefficient, while Hospital B's smaller local heterogeneity index leads to a larger adaptive step size coefficient. This difference in the adaptive step size coefficient adjusts the magnitude of the gradient update. Data comparison shows that the values ​​of the local heterogeneity index differ among different data holders, and the adaptive step size coefficient changes accordingly, thus adapting to the local distribution characteristics of multi-source heterogeneous privacy data.

[0028] In one embodiment of the present invention, model parameter update amounts from all data holders are collected, and all model parameter update amounts are homomorphically encrypted and aggregated to form an encrypted aggregated model parameter update. The process is as follows: [Refer to...] Figure 3 During the initialization phase of the federated learning system, public-private key pairs generated based on a homomorphic encryption scheme are distributed to all participating data holders and the central server. Each data holder uses the central server's public key to homomorphically encode each floating-point parameter in the noisy model parameter update, generating an encrypted model parameter update vector. After receiving the encrypted model parameter update vectors from all data holders, the central server performs element-wise addition on all encrypted model parameter update vectors based on the additive homomorphic property of homomorphic encryption. After completing the element-wise addition, the central server obtains a result vector with the same dimension as a single encrypted model parameter update vector. This result vector is the encrypted aggregated model parameter update, containing the ciphertext of the sum of all data holders' updates. During the aggregation process, the central server does not perform any decryption operations and cannot parse any individual data holder's original update information from the encrypted aggregated model parameter update.

[0029] In practice, the process involves collecting model parameter updates from all data holders and homomorphically aggregating them to form an encrypted aggregated model parameter update. During the initialization phase of the federated learning system, public-private key pairs generated based on a homomorphic encryption scheme are distributed to all participating data holders and the central server. These key pairs are generated using the Paillier homomorphic encryption algorithm. Each data holder uses the central server's public key to homomorphically encode each floating-point parameter in the noisy model parameter update, generating an encrypted model parameter update vector. The homomorphic encryption process converts the floating-point parameters into large integers before encryption calculation. After receiving the encrypted model parameter update vectors from all data holders, the central server performs element-wise addition on all encrypted model parameter update vectors based on the additive homomorphic property of homomorphic encryption. This element-wise addition operation adds the ciphertext values ​​at each position in the model parameters. After completing the element-wise addition operation, the central server obtains a result vector with the same dimension as the single encrypted model parameter update vector. This result vector is the encrypted aggregated model parameter update, and its content is the ciphertext sum of the updates from all data holders. The central server does not perform any decryption during the aggregation process and cannot parse any individual data holder's original update information from the encrypted aggregated model parameter update; the original update information of each data holder always exists in ciphertext form.

[0030] In some embodiments, the mathematical expression of the homomorphic encryption aggregation process can be described by a formula, and the element-wise addition operation of the encrypted model parameter update vector is expressed as: in: The updated model parameters representing the aggregated encrypted state are at the [number]th [period]. Aggregated ciphertext at each parameter position, Represents a homomorphic encryption function. This represents the public key of the central server. Indicates the first The noise model parameter update of each data holder in the th... Plaintext values ​​at each parameter position This represents the total number of data holders participating in the aggregation. This represents the summation operation on the ciphertext. The homomorphic encryption function in the formula... It satisfies the additive homomorphic property, meaning the encrypted result of the sum of the plaintext values ​​equals the sum of the ciphertext values. The model parameters are updated after the aggregation of the encrypted states. It is made by all The resulting vector has the same dimensions as the single encrypted model parameter update vector.

[0031] In practical implementation, the example scenario uses the federated training of credit risk assessment models from multiple financial institutions. The data holders are banks, and each bank's noisy model parameter update is a floating-point vector containing weights and bias parameters. Bank A uses the central server's public key to homomorphically encode its generated noisy model parameter update, producing an encrypted model parameter update vector. Each element in Bank A's encrypted model parameter update vector is a Paillier ciphertext. Bank B uses the same central server's public key to homomorphically encode its generated noisy model parameter update, producing its own encrypted model parameter update vector. Bank B's encrypted model parameter update vector has the same dimension as Bank A's vector. After receiving the encrypted model parameter update vectors from Bank A and Bank B, the central server adds the ciphertext at the first position of each vector, adds the ciphertext at the second position, and so on, completing the ciphertext addition for all corresponding positions. Data comparison shows that the ciphertext value at a certain position in Bank A's encrypted model parameter update vector is... Bank B's ciphertext value at the same position is The central server calculates the aggregated encrypted state, and then updates the ciphertext value at that location using the resulting model parameters. (In the Paillier encryption scheme, ciphertext multiplication is equivalent to plaintext addition). After decryption, this ciphertext value corresponds to the sum of the plaintext values ​​of the update amounts of Bank A and Bank B on the model parameters.

[0032] In one embodiment of the present invention, a secure decryption operation is performed on the aggregated model parameter update in the encrypted state to obtain the decrypted aggregated model parameter update. The process involves the central server, after completing homomorphic encrypted aggregation, sending the aggregated model parameter update in the encrypted state to one or more authorized decryption nodes. The decryption nodes use pre-allocated and securely stored decryption private keys to perform decryption calculations on each ciphertext element of the aggregated model parameter update in the encrypted state. The decryption calculation produces a plaintext vector with the same dimension as the aggregated model parameter update in the encrypted state. Each element of this plaintext vector corresponds to the algebraic sum of the update amounts for the corresponding model parameters by all data holders. The format and dimension of the decrypted plaintext vector are verified to meet expectations, and the verified plaintext vector is defined as the decrypted aggregated model parameter update. The decrypted aggregated model parameter update is then sent back to the central server for subsequent global model update steps.

[0033] The global model parameters generated in the previous iteration are updated using aggregated model parameter updates to obtain new global model parameters. The process involves the central server receiving the decrypted aggregated model parameter updates and counting the number of valid data holders who successfully participated in the aggregation in this round. Each element of the decrypted aggregated model parameter update is divided by the number of valid data holders to calculate the average update amount of all participants in this round. A weighted average update rule is used to weight and fuse the global model parameters generated in the previous iteration with the calculated average update amount of the model parameters. The sum of the weight coefficients used for the global model parameters in the previous round and the weight coefficients used for the average update amount of the model parameters is one. This weighted fusion calculation generates a new set of model parameter values, which constitutes the new global model parameters. The new global model parameters are compared with those generated in the previous iteration. If the change in parameters is less than a preset convergence threshold, the global model is marked as converged.

[0034] In practice, the process involves performing a secure decryption operation on the aggregated model parameter updates in the encrypted state to obtain the decrypted aggregated model parameter updates. After completing homomorphic encryption aggregation, the central server sends the aggregated model parameter updates in the encrypted state to one or more authorized decryption nodes. Each decryption node is a pre-designated independent computing unit with secure storage capabilities. The decryption nodes use pre-allocated and securely stored decryption private keys to decrypt each ciphertext element of the aggregated model parameter updates in the encrypted state. The decryption calculation is based on the decryption algorithm of the homomorphic encryption scheme. The decryption calculation produces a plaintext vector with the same dimension as the aggregated model parameter updates in the encrypted state. Each element of the plaintext vector corresponds to the algebraic sum of the updates made by all data holders to the corresponding model parameters. The format and dimension of the decrypted plaintext vector are verified to ensure they meet expectations. The verification process includes checking whether the vector length matches the total number of model parameters and whether each element is a valid floating-point number. The verified plaintext vector is defined as the decrypted aggregated model parameter update, and the decrypted aggregated model parameter update is sent back from the decryption nodes to the central server for subsequent global model update steps.

[0035] In some embodiments, the mathematical expression of the secure decryption operation can be described by a decryption formula, which is expressed as follows: in: This indicates that the parameters of the decrypted aggregation model are updated at the [number]th [time]. Plaintext values ​​at each parameter position This represents a homomorphic decryption function. This indicates the decryption private key securely stored by the decryption node. The updated model parameters after the aggregation of the received encrypted states are in the [number]th [period]. The aggregated ciphertext at each parameter position. Decryption function. It is an encryption function. The inverse operation, using the decryption private key. For ciphertext Calculations were performed to obtain the number of all data holders in the first... Plaintext of the sum of updates on each parameter Update the parameters of the decrypted aggregation model. It is made by all The vector formed, vector It was sent back to the central server.

[0036] In practice, the process involves updating the global model parameters generated in the previous iteration using the decrypted aggregated model parameter updates to obtain the new global model parameters. After receiving the decrypted aggregated model parameter updates, the central server counts the number of valid data holders who successfully participated in the aggregation in this round. The number of valid data holders is the total number of data holders who successfully submitted the encrypted model parameter update vector and were successfully aggregated. The average value of the model parameter updates for all participants in this round is calculated by dividing each element of the decrypted aggregated model parameter updates by the number of valid data holders; this calculation is an element-wise scalar division. A weighted average update rule is used to weight and fuse the global model parameters generated in the previous iteration with the calculated average value of the model parameter updates. The sum of the weight coefficients used for the global model parameters in the previous round and the weight coefficients used for the average value of the model parameter updates is one. This weighted fusion calculation generates a new set of model parameter values, which are the new global model parameters. The differences between the new global model parameters and the global model parameters generated in the previous iteration are compared. If the parameter change is less than a preset convergence threshold, the global model is marked as having converged.

[0037] In some embodiments, the decryption and update process is illustrated by comparing an example scenario and data from a cross-regional sales forecasting model. The example scenario includes three data holders: sales data centers in regions A, B, and C. After receiving the encrypted aggregated model parameter update, the central server sends it to the sole authorized decryption node. The decryption node uses its private key to calculate the decrypted aggregated model parameter update. Table 1 shows the plaintext values ​​of some parameters in the decrypted aggregated model parameter update, the number of valid data holders, and the calculated average values. Table 1: Parameter Update and Average Calculation Table for Decrypted Aggregation Model The data comparison illustrates that the updated values ​​of the decrypted aggregated model parameters are the sum of the updates from the three regional data centers. For example, the aggregated value for parameter index 1 is 0.045. The central server calculates the average based on the number of valid data holders being 3, resulting in an average update value of 0.015 for parameter index 1. The central server employs a weighted average update rule, setting the weight coefficient for the previous round of global model parameters to 0.7 and the weight coefficient for the average value of model parameter updates to 0.3. For parameter index 1, assuming the previous round's global model parameter value was 0.100, the new round's global model parameter calculation is 0.100. 0.7 + 0.015 0.3 = 0.0745. After performing this weighted fusion calculation on all parameters, a complete new set of global model parameters is obtained.

[0038] Optionally, the weight coefficients in the weighted average update rule can be dynamically adjusted with each training epoch. In the early stages of training, the average value of the model parameter updates is given higher weight, while in the later stages, the global model parameters from the previous epoch are given higher weight. Optionally, the convergence threshold can be set to a vector norm; for example, convergence is indicated when the Euclidean norm between the global model parameters of the new epoch and those of the previous iteration is less than the threshold ε. This can be understood as the decryption node needing to process the plaintext vector after completing the decryption calculation. Format validation is performed, and if validation fails, an error code is returned to the central server. It's understandable that the number of valid data holders may differ in each federated learning iteration; the central server counts the number of successfully participating data holders from the integrity verification information.

[0039] In one embodiment of the present invention, the performance of the new round of global model parameters on a standard public dataset is verified, and a decision is made on whether to initiate a new round of federated learning iterations based on the verification results. The process involves the central server, after obtaining the new round of global model parameters, loading them into a verification model with the same structure as the federated learning model. The verification model is run on a public benchmark dataset independent of the private datasets of all data holders, and the accuracy metric and loss function value of the model on the test task are calculated. The calculated accuracy metric and loss function value are compared with the pre-recorded historical best accuracy metric and historical lowest loss function value. If the current accuracy metric is better than the historical best accuracy metric, or the current loss function value is lower than the historical lowest loss function value, the model performance is determined to have improved. The new round of global model parameters is archived as the current best global model parameters, and a command is broadcast to all data holders to initiate a new round of federated learning iterations. If the current accuracy metric is not better than the historical best accuracy metric and the current loss function value is not lower than the historical lowest loss function value, the model performance is determined not to have improved. The central server will pause initiating a new round of iterations and analyze the trend of model parameter changes to determine subsequent strategies.

[0040] Before initiating a new round of federated learning iterations, this method includes a dynamic participant selection step based on model performance feedback. The process involves the central server evaluating the contribution quality and reliability of each data holder based on the integrity verification information of their local training processes submitted in previous federated learning rounds. Specifically, the evaluation analyzes the changing trends of the norm sum of squared gradients of the loss function submitted by each data holder historically, the regularity of timestamp calculations, and the consistency of the summary hash values. Data holders whose contribution quality and reliability evaluation results are below a preset standard are marked as pending observation in the invitation list for the next round of federated learning iterations. When constructing the participant set for the next iteration, data holders with high contribution quality and reliability evaluation results are prioritized, while data holders in the pending observation state are temporarily excluded from participation with a certain probability. The final participant set list is distributed along with the new round of global model parameters, thereby dynamically optimizing the node composition participating in federated learning in each iteration.

[0041] In practice, the process involves verifying the performance of the new round of global model parameters on a standard public dataset and deciding whether to initiate a new round of federated learning iterations based on the verification results. After receiving the new round of global model parameters, the central server loads them into a verification model with the same structure as the federated learning model. This verification model is a temporary instance used for performance evaluation. The verification model is run on a publicly available benchmark dataset, independent of the private datasets of all data holders. This publicly available benchmark dataset is a widely accepted standardized dataset. The model's accuracy metric and loss function value on the test task are calculated. The accuracy metric is the proportion of correctly classified samples, and the loss function value reflects the overall error between the model's prediction and the true label. The calculated accuracy metric and loss function value are compared with pre-recorded historical best accuracy metric and historical lowest loss function value. The historical best accuracy metric is the highest accuracy obtained in all previous rounds, and the historical lowest loss function value is the lowest loss obtained in all previous rounds. If the current accuracy metric is better than the historical best accuracy metric, or the current loss function value is lower than the historical lowest loss function value, then the model performance is considered improved. The new set of global model parameters is then archived as the current optimal global model parameters, and a new round of federated learning iterations is initiated by broadcasting a command to all data holders. If the current accuracy metric is not better than the historical best accuracy metric and the current loss function value is not lower than the historical lowest loss function value, then the model performance is considered not improved. The central server will pause initiating a new round of iterations and analyze the trend of model parameter changes to determine the subsequent strategy.

[0042] In some embodiments, the evaluation logic of the dynamic participant selection step based on model performance feedback can be expressed by a formula, and the calculation of contribution quality and reliability evaluation scores is expressed as follows: in: Indicates the first The contribution quality and reliability assessment score of each data holder. The stability score is calculated based on the trend of the variation of the norm sum of squares of the gradient of the loss function. This represents a reliability score calculated based on the regularity of the timestamps. This represents the deviation score calculated based on the consistency of the digest hash value. , and These are preset non-negative weighting coefficients. (Score) The higher the number, the better. The better the contribution quality and reliability assessment results of each data holder, the better the central server will be based on the score. Sort and filter.

[0043] In practice, before initiating a new round of federated learning iterations, a dynamic participant selection step based on model performance feedback is performed. The central server evaluates the contribution quality and reliability of each data holder based on the integrity verification information of the local training process submitted by each data holder in previous rounds of federated learning. Specifically, the evaluation analyzes the trend of the norm sum of squares of the loss function gradients submitted by each data holder historically, the regularity of the timestamp calculations, and the consistency of the summary hash values. The trend of the norm sum of squares of the loss function gradients is measured by calculating the variance of the submitted values ​​across multiple rounds. Data holders whose contribution quality and reliability evaluation results are below a preset standard are marked as pending observation in the invitation list for the next round of federated learning iterations. The preset standard is a minimum evaluation score threshold. When forming the participant set for the next round of iterations, data holders with high contribution quality and reliability evaluation results are prioritized, and data holders in the pending observation state are temporarily excluded from participation with a certain probability. The final participant set list is distributed along with the new round of global model parameters, thereby dynamically optimizing the node composition participating in federated learning in each iteration.

[0044] In some embodiments, the verification and dynamic selection process is illustrated by comparing example scenarios and data with an IoT device anomaly detection model. The example scenario involves four data holders: equipment management platforms of factories A, B, C, and D. The standard public dataset is a publicly available dataset of industrial equipment sensor readings. The central server performs verification after obtaining the new round of global model parameters. Table 2 shows a comparison of the verification results from the most recent three iterations with the historical best record. Table 2: Model Performance Validation Results and Iteration Decision Table Data comparison shows that the accuracy metric of the 5th iteration (92.5%) is better than the historical best accuracy metric of 92.3%, indicating a performance improvement, and the decision is to start the 6th iteration. After the 6th iteration, the historical best accuracy metric is updated to 92.5%. The accuracy metric of the 7th iteration is not better than the historical best value and the loss function value is not lower than the historical lowest value, indicating no performance improvement, and the decision is to pause the iteration. In the dynamic participant selection, the central server calculates the contribution quality and reliability evaluation score of each factory, for example, the score of factory A. The score for factory B is 85. The score for factory C is 92. The score of factory D is 45. The score is 78. The preset standard score is 60. Factory C's score is lower than the standard and is marked as an observation state. When forming the participant set for the 8th round, factories A, B, and D with higher scores are selected first. Factory C is excluded from participation with a 50% probability. If the random decision result is exclusion, the participant set for the 8th round is {A,B,D}.

[0045] Optionally, the regularity of timestamp calculation can be scored by analyzing the stability of the time intervals between consecutive rounds of update packet submissions; the more stable the time intervals, the higher the regularity score. The higher the score, the better. Optionally, the consistency of the digest hash value can be assessed by checking whether the original data format used to generate the digest hash value conforms to predetermined specifications in the integrity verification information of the local training process submitted by the same data holder in multiple consecutive rounds. It is understandable that analyzing the trend of model parameter changes may include calculating the norm of the update magnitude vector of the global model parameters in recent rounds. If the norm remains below a very small threshold, training may be terminated prematurely. It is also understandable that after pausing the initiation of a new round of iterations, the central server's subsequent strategies may include adjusting the weight coefficients in the weighted average update rule, requiring some data holders to retrain, or switching to a new public benchmark dataset for validation.

[0046] In one embodiment of the invention, the method further includes a step of inserting a model watermark during the federated learning process to verify model ownership. The process involves the central server generating a unique secret watermark key associated with the model structure during the initialization of the federated learning system. After generating a new round of global model parameters in each round, the central server uses the unique secret watermark key and the current model parameters to calculate a hidden digital watermark signal using an irreversible watermark embedding function. The hidden digital watermark signal is then subtly superimposed onto a specific subset of the new round of global model parameters in a manner that does not affect the model's main functionality, generating watermarked global model parameters. These watermarked global model parameters replace the original global model parameters and are distributed to the data holder for the next round of local training or released as the final model. When model ownership verification is required, the corresponding watermark extraction function and the unique secret watermark key are used to attempt to extract the hidden digital watermark signal from the model parameters being verified. Ownership is confirmed by verifying whether the extracted signal matches the expected signal.

[0047] The method also includes a fault tolerance and recovery mechanism for handling data holders who withdraw midway. The process is as follows: at the beginning of each federated learning iteration, the central server records a list of data holder identities confirmed for participation in that round. During the model parameter update collection phase, the central server sets a receiving timer for each data holder in the identity list. If a data holder's model parameter update has not arrived after the receiving timer expires, the data holder is marked as a node that has withdrawn from this round, and a partial data digest of its transmission is recorded. Homomorphic encryption aggregation and subsequent decryption and global update operations are continued using the model parameter updates successfully received from other data holders. In the next iteration invitation, the central server attempts to re-contact the nodes marked as withdrawn. If re-contact is successful and the node provides reasonable proof of delay, its participation is restored; otherwise, it is removed from the subsequent identity list, and its contributed parameter updates are archived based on historical contributions.

[0048] In practice, the steps of inserting model watermarks to verify model ownership during federated learning include the initialization phase and operations after each iteration. During system initialization, the central server generates a unique secret watermark key related to the model structure. This key is a random sequence derived from the hash values ​​of the initial model parameters. After generating new global model parameters in each iteration, the central server uses the unique secret watermark key and the current model parameters to calculate a hidden digital watermark signal using an irreversible watermark embedding function. This function employs a modulation method based on pseudo-random sequences. The hidden digital watermark signal is then subtly superimposed onto a specific subset of the new global model parameters in a manner that does not affect the model's main functionality, generating watermarked global model parameters. This superposition operation adds the watermark signal with a small perturbation to the selected model parameters. The watermarked global model parameters then replace the original global model parameters and are distributed to the data holder for the next round of local training or as the final model release. The performance of the watermarked global model parameters on the main tasks is essentially the same as that of the original model parameters. When it is necessary to verify the ownership of a model, the corresponding watermark extraction function and the unique secret watermark key are used to try to extract the hidden digital watermark signal from the parameters of the model being verified. The ownership is confirmed by verifying whether the extracted signal matches the expected signal. If the match is successful, it proves the ownership of the model.

[0049] In some embodiments, the mathematical expression of the watermark embedding and extraction process can be described by a formula, and the embedding expression of the hidden digital watermark signal is as follows: in: This represents the global model parameter vector with watermark. This represents the original global model parameter vector for the new round. This represents a tiny scalar coefficient that controls the watermark strength. This represents a pseudo-random sequence with the same dimension as the model parameter vector, generated by a unique secret watermark key. This represents a selection vector with the same dimension as the model parameter vector, whose elements are either 0 or 1, used to specify a specific subset of parameters. This represents element-wise multiplication. It is calculated when extracting hidden digital watermark signals. ,in This represents the claimed original model parameter vector, obtained by comparing the extracted sequences. as expected The correlation between them is used to verify ownership.

[0050] In practical implementation, the fault tolerance and recovery mechanism for handling data holders leaving midway involves state management across multiple iterations. At the start of each federated learning iteration, the central server records a list of confirmed data holder identities, containing unique codes for each confirmed participant. During the model parameter update collection phase, the central server sets a receiving timer for each data holder in the identity list. The timeout duration of the receiving timer is dynamically set based on network conditions and historical latency. If a data holder's model parameter update has not arrived after the receiving timer expires, the data holder is marked as a node that has left the round, and its partially transmitted data digest is recorded. This partially transmitted data digest is used to verify data consistency during subsequent recovery. Homomorphic encryption aggregation and subsequent decryption and global update operations are then performed using model parameter updates successfully received from other data holders. The global model update is based on the actual received update amounts. In the next round of iterative invitations, the central server will attempt to re-contact the nodes marked as withdrawn. If the re-contact is successful and the node provides reasonable proof of delay, its participation will be restored. Otherwise, it will be removed from the subsequent identity list and its contributed parameters will be archived based on its historical contributions.

[0051] In some embodiments, the watermark insertion and fault tolerance process is illustrated by comparing an example scenario and data from a cross-institutional medical research consortium. In the example scenario, three medical research institutions, X, Y, and Z, participate in a federated learning task for medical image analysis. During the initialization phase, the central server generates a binary sequence of length L as the unique secret watermark key K. After generating the new round of global model parameters Mt in the t-th iteration, the central server uses a formula to calculate the watermarked global model parameters Wt, where the selection vector P specifies 5% of the weight parameters in the fully connected layers of the model as a specific subset, and the scalar coefficient λ is set to 0.001. After watermark embedding, the initial model received by institutions X, Y, and Z for local training in the (t+1)-th iteration is Wt, not Mt. After the model is released, the copyright holder uses the key K and the selection vector P to extract signals from the model parameters, calculating a correlation coefficient with λ·K that reaches 0.98, far higher than the correlation coefficient of random noise (typically below 0.1), thus verifying ownership. Regarding the fault tolerance mechanism, it is assumed that in the (t+1)-th iteration, the initial record of the identity list is {X, Y, Z}. The central server sets a receiving timer to 30 minutes. Institution Y sent its model parameter update after 25 minutes, Institution X after 28 minutes, and Institution Z failed to send its update after 35 minutes. Institution Z's receiving timer expired, and the central server marked Institution Z as a node that exited this round. The central server continued subsequent aggregation and updates using the successfully received model parameter updates from Institution X and Institution Y. During the (t+2)th round of iteration invitations, the central server attempted to re-contact Institution Z. If Institution Z responded within 5 minutes and provided network interruption logs as proof, Institution Z's participation was restored and it was added to the identity list for the (t+2)th round; if Institution Z did not respond, it was removed from the invitation list for subsequent rounds.

[0052] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A federated learning optimization training method for multi-source heterogeneous privacy data, characterized in that, The method includes: Initialize the federated learning system by sending unified global model parameters and model structure definitions to multiple data holders participating in the federated learning process; Each data holder trains the received global model parameters based on their locally stored private dataset using an improved model parameter local update algorithm. The improved model parameter local update algorithm adaptively adjusts the update step size of each iteration according to the degree of difference in the local data distribution of each data holder, and generates model parameter update amount with local training process integrity verification information. Collect the model parameter update amounts from all data holders, perform homomorphic encrypted aggregation on all the model parameter update amounts, and form an encrypted aggregated model parameter update; A secure decryption operation is performed on the aggregated model parameter update of the encrypted state to obtain the decrypted aggregated model parameter update; The global model parameters generated in the previous iteration are updated using the aggregated model parameters to obtain the global model parameters for the new iteration. Verify the performance of the new round of global model parameters on standard public datasets, and decide whether to start a new round of federated learning iterations based on the verification results.

2. The federated learning optimization training method for multi-source heterogeneous privacy data according to claim 1, characterized in that, The improved model parameter local update algorithm adaptively adjusts the update step size of each iteration based on the degree of difference in the local data distribution of each data holder, including: At the start of each local training iteration, the gradient of the loss function of the current local model parameters on a random sampled batch of the local private dataset is calculated. The covariance matrix of the feature vectors of all samples in the local private dataset is calculated, and the ratio of the largest eigenvalue to the smallest eigenvalue of the covariance matrix is ​​calculated. The ratio is used as a local heterogeneity index to characterize the degree of difference in the distribution of local data. An adaptive coefficient calculation rule that is negatively correlated with the local heterogeneity index is introduced. Based on the current iteration round and the local heterogeneity index, a positive number less than one is dynamically calculated as the adaptive step size coefficient for this iteration. The calculated gradient of the loss function is multiplied by the adaptive step size coefficient to obtain the gradient value after step size adjustment. Using the gradient values ​​adjusted by the step size, combined with momentum optimization techniques, update the current local copy of the model parameters to complete one local iteration update; After all preset local iterations are completed, the difference vector between the initial local model parameters and the final local model parameter copy is calculated. The difference vector is used as the model parameter update amount, and the norm sum of squared gradients of all loss functions calculated during local training is added to the model parameter update amount as part of the local training process integrity verification information.

3. The federated learning optimization training method for multi-source heterogeneous privacy data according to claim 2, characterized in that, Each data holder trains the received global model parameters based on their locally stored private dataset using an improved local model parameter update algorithm, and generates model parameter updates with information verifying the integrity of the local training process, including: Upon receiving the global model parameters, each data holder loads them as the initial state of the local model parameters. During local training, the computation timestamp, computational resource consumption metric, and digest hash value of the final generated model parameter update amount are recorded each time the improved model parameter local update algorithm is used to update the parameters. The computation timestamp, the computational resource consumption metric, and the digest hash value, together with the norm sum of squares of the loss function gradient obtained from the improved local model parameter update algorithm, are jointly encoded into structured local training process integrity verification information. A differential privacy mechanism is applied to the model parameter update amount by adding random noise conforming to a specific Gaussian distribution to the difference vector to generate a noisy model parameter update amount that meets the differential privacy requirements. The updated parameters of the noisy model are digitally signed and packaged with the structured integrity verification information of the local training process to form a complete update data package ready for uploading.

4. The federated learning optimization training method for multi-source heterogeneous privacy data according to claim 3, characterized in that, The process of collecting model parameter update data from all data holders, homomorphically aggregating all model parameter update data to form an encrypted aggregated model parameter update includes: During the initialization phase of the federated learning system, public-private key pairs generated based on a homomorphic encryption scheme are distributed to all participating data holders and the central server. Each data holder uses the public key of the central server to perform homomorphic encryption encoding on each floating-point parameter in the noisy model parameter update, generating an encrypted model parameter update vector. After receiving the encrypted model parameter update vectors from all data holders, the central server performs element-wise addition on all the encrypted model parameter update vectors based on the additive homomorphic property of homomorphic encryption. After completing the element-by-element addition operation, the central server obtains a result vector with the same dimension as the single encrypted model parameter update vector. The result vector is the aggregated model parameter update of the encrypted state, and its content is the ciphertext of the sum of the update amounts of all data holders. During the aggregation process, the central server does not perform any decryption operations and cannot parse any original update information of individual data holders from the encrypted aggregated model parameter updates.

5. The federated learning optimization training method for multi-source heterogeneous privacy data according to claim 4, characterized in that, Perform a secure decryption operation on the aggregated model parameter update of the encrypted state to obtain the decrypted aggregated model parameter update, including: After completing the homomorphic encryption aggregation, the central server updates the aggregated model parameters of the encrypted state and sends them to one or more authorized decryption nodes. The decryption node uses a pre-allocated and securely stored decryption private key to perform decryption calculations on each ciphertext element of the aggregated model parameter update of the encrypted state; The decryption calculation generates a plaintext vector with the same dimension as the updated parameters of the aggregated model after the encryption state. Each element of the plaintext vector corresponds to the algebraic sum of the update amounts of all data holders on the corresponding parameters of the model. Verify whether the format and dimension of the plaintext vector obtained by decryption meet the expectations, and define the plaintext vector that passes the verification as the updated parameters of the decrypted aggregation model; The decrypted aggregated model parameters are updated and sent back to the central server for subsequent global model update steps.

6. The federated learning optimization training method for multi-source heterogeneous privacy data according to claim 5, characterized in that, The aggregated model parameters are updated to update the global model parameters generated in the previous iteration, resulting in new global model parameters, including: After receiving the decrypted aggregation model parameter update, the central server counts the number of valid data holders who successfully participated in the aggregation in this round. Divide each element of the decrypted aggregate model parameter update by the number of valid data holders to calculate the average value of the model parameter update amount for all participants in this round; A weighted average update rule is adopted, which weights and fuses the global model parameters generated in the previous iteration with the average value of the calculated model parameter update. The sum of the weight coefficients used for the global model parameters in the previous iteration and the weight coefficients used for the average value of the model parameter update is one. Perform weighted fusion calculation to generate a new set of model parameter values, which are the global model parameters for the new round; The global model parameters of the new round are compared with those of the global model parameters generated in the previous iteration. If the change in parameters is less than the preset convergence threshold, the global model is marked as having reached convergence.

7. The federated learning optimization training method for multi-source heterogeneous privacy data according to claim 6, characterized in that, The process of verifying the performance of the new round of global model parameters on a standard public dataset, and deciding whether to start a new round of federated learning iterations based on the verification results, includes: After receiving the new round of global model parameters, the central server loads them into a validation model with the same structure as the federated learning model. The validation model is run on a public benchmark dataset that is independent of the private datasets of all data holders, and the accuracy metrics and loss function values ​​of the model on the test task are calculated. The calculated accuracy metric and loss function value are compared with the pre-recorded historical best accuracy metric and historical lowest loss function value. If the current accuracy metric is better than the historical best accuracy metric, or the current loss function value is lower than the historical lowest loss function value, then the model performance is determined to be improved. The new round of global model parameters is archived as the current best global model parameters, and the instruction is broadcast to all data holders to start a new round of federated learning iteration. If the current accuracy metric is not better than the historical best accuracy metric and the current loss function value is not lower than the historical lowest loss function value, then the model performance is determined not to have improved. The central server will then pause initiating a new round of iterations and analyze the trend of model parameter changes to determine the subsequent strategy.

8. The federated learning optimization training method for multi-source heterogeneous privacy data according to claim 7, characterized in that, Before initiating a new round of federated learning iterations, the method also includes a dynamic participant selection step based on model performance feedback: The central server evaluates the quality and reliability of each data holder's contribution based on the local training process integrity verification information submitted by each data holder in the historical multi-round federated learning. In specific evaluation, the changing trend of the norm sum of squares of the loss function gradients submitted by each data holder in history, the regularity of the calculated timestamps, and the consistency of the digest hash values ​​are analyzed. For data holders whose contribution quality and reliability assessment results are lower than the preset standards, they will be marked as pending observation in the invitation list for the next round of federated learning iterations; When forming the set of participants for the next iteration, data holders with high contribution quality and reliability assessment results are given priority. Data holders in the observation state are temporarily excluded from participation with a certain probability. The final list of participants is distributed along with the new set of global model parameters, thereby dynamically optimizing the composition of nodes participating in federated learning in each iteration.

9. The federated learning optimization training method for multi-source heterogeneous privacy data according to claim 1, characterized in that, The method also includes the step of inserting a model watermark during the federated learning process to verify model ownership: When initializing the federated learning system, the central server generates a unique secret watermark key associated with the model structure; After generating a new round of global model parameters in each round, the central server uses the unique secret watermark key and the current model parameters to calculate a hidden digital watermark signal through an irreversible watermark embedding function. The hidden digital watermark signal is superimposed onto a specific subset of the new round of global model parameters in a covert manner that does not affect the main function of the model, thereby generating watermarked global model parameters. The watermarked global model parameters will replace the original global model parameters and be sent to the data holder for the next round of local training, or released as the final model. When it is necessary to verify the ownership of the model, the corresponding watermark extraction function and the unique secret watermark key are used to attempt to extract the hidden digital watermark signal from the parameters of the model being verified. The ownership is confirmed by verifying whether the extracted signal matches the expectation.

10. The federated learning optimization training method for multi-source heterogeneous privacy data according to claim 1, characterized in that, The method also includes fault tolerance and recovery mechanisms to handle data holders leaving midway: At the beginning of each federated learning iteration, the central server records a list of data holder identities confirmed to participate in that round. During the model parameter update collection phase, the central server sets a receiving timer for each data holder in the identity list; If the model parameter update of a data holder has not arrived after the receive timer expires, the data holder will be marked as a node that has exited this round, and a partial data summary of the data it has transmitted will be recorded. Using the model parameter updates from other data holders that have been successfully received, continue homomorphic encryption aggregation and subsequent decryption and global update operations; In the next round of iterative invitations, the central server will attempt to re-contact the nodes marked as withdrawn. If the re-contact is successful and the node provides reasonable proof of delay, its participation will be restored. Otherwise, it will be removed from the subsequent identity list, and its contributed parameter updates will be archived based on its historical contributions.

Images