Model training method and device based on longitudinal federated learning

Abstract

The invention relates to the technical field of machine learning, and discloses a longitudinal federated learning-based model training method and device, which are applied to a distributed network comprising a coordination node and a plurality of participating nodes, the method is executed by the coordination node, and the method comprises the following steps: acquiring encrypted data of each participating node; the encrypted data comprises an intermediate result; the intermediate result is generated by each participating node based on data after local preprocessing and current global model parameters; aggregating the encrypted data of each participating node, updating the global model, and generating updated global model parameters; transmitting the updated global model parameters to each participating node; and judging whether a termination condition is met at the moment, and if yes, completing the training of the global model. Through encrypted intermediate result transmission and global model iteration updating, the problems of data privacy leakage and low model training efficiency in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and particularly to a model training method and device based on vertical federated learning. Background Art

[0002] Vertical Federated Learning (VFL) collaboratively trains a model among multiple data participants to achieve privacy protection without data leaving the domain. However, existing methods require participants to directly upload raw intermediate results (such as gradients, loss values) without mandatory encrypted transmission. Attackers can use the intermediate results to reverse-engineer the original data, leading to user privacy leakage. Moreover, traditional methods (such as the FedAvg algorithm) adopt a synchronous update mechanism, requiring all participating nodes to upload data after completing local training. If some nodes fail to respond in a timely manner due to calculation delays or network fluctuations, the global model update will be forced to wait, significantly reducing the training efficiency.

[0003] Therefore, there is an urgent need for a new model training method based on vertical federated learning that can overcome the problems of data privacy leakage and low model training efficiency in the prior art. Summary of the Invention

[0004] In view of this, the present invention provides a model training method and device based on vertical federated learning to solve the problems of data privacy leakage and low model training efficiency in the prior art. The technical solution is as follows.

[0005] In a first aspect, the present invention provides a model training method based on vertical federated learning, which is applied to a distributed network including a coordination node and multiple participating nodes. This method is executed by the coordination node and includes:

[0006] Obtain the encrypted data of each participating node; the encrypted data includes intermediate results; the intermediate results are generated by each participating node based on the locally preprocessed data and the current global model parameters;

[0007] Aggregate the encrypted data of each participating node, update the global model, and generate updated global model parameters;

[0008] Transmit the updated global model parameters to each participating node;

[0009] Determine whether the termination condition is satisfied at this time. If it is satisfied, complete the training of the global model.

[0010] The model training method based on vertical federated learning provided by the present invention has the following beneficial effects.

[0011] The model training method based on vertical federated learning provided by the present invention first obtains the encrypted intermediate results of each participating node (generated based on local preprocessed data and current global model parameters). Since the intermediate results are preprocessed such as standardized and feature screened locally and encrypted, it avoids the risk of privacy leakage caused by direct transmission of raw data. At the same time, the data format after preprocessing is unified, reducing the computational overhead of cross-node data alignment and improving the training efficiency. Secondly, aggregate the encrypted data to update the global model parameters, perform secure calculations on the encrypted intermediate results in the ciphertext state through aggregation algorithms such as weighted average, prevent the exposure of gradients or parameter plaintext, and combine dynamic termination conditions (such as model convergence or maximum number of iterations), so that the model stops training immediately when reaching the optimal performance, avoiding resource waste or underfitting problems caused by fixed number of iterations. Finally, transmit the updated parameters to each participating node, ensure that all nodes can load the consistent model version in time through sharded encrypted transmission and version identification code synchronization, and prevent version conflicts caused by network latency or node offline. The synergistic effect of the above technical features realizes the comprehensive improvement of model training efficiency, resource utilization rate and system robustness on the premise of ensuring data privacy.

[0012] In an alternative embodiment, the intermediate result includes at least one of the parameter gradient of the global model, the loss value of the global model, and the intermediate activation value of the global model.

[0013] The model training method based on vertical federated learning provided by the present invention selectively transmits key intermediate results (for example, only transmits gradients instead of complete activation values), reducing the amount of data communicated each time; the loss value participates in dynamic weight allocation to enhance the training contribution of key samples.

[0014] In an alternative embodiment, the preprocessing includes data standardization, feature dimensionality reduction, feature selection based on feature importance scoring, and sample screening.

[0015] The model training method based on vertical federated learning provided by the present invention eliminates the difference in feature dimensions through standardization, improving the convergence stability of the model; reduces the feature dimension through dimensionality reduction, shortening the local training time; and eliminates irrelevant features based on feature selection using information gain to prevent overfitting.

[0016] In an alternative embodiment, the encrypted data is generated by a participating node through symmetric encryption or asymmetric encryption of the intermediate result.

[0017] The model training method based on vertical federated learning provided by the present invention uses symmetric encryption (such as AES) for the intermediate result to resist brute-force cracking attacks; or uses asymmetric encryption (such as RSA), taking into account both encryption efficiency and security.

[0018] In an alternative embodiment, before the step of aggregating the encrypted data of each participating node, the method further includes:

[0019] Performing shard integrity verification, digital signature authentication, and zero - knowledge proof verification on each encrypted data;

[0020] If the verification is successful, based on the current loss change rate of the global model, the sample gradient contribution degree, and the feature importance score, determine the priority weights of each participating node; allocate the aggregation weights of each participating node according to the priority weights of each participating node; and aggregate the encrypted data of each participating node according to the aggregation weights;

[0021] If the verification fails, trigger data re - transmission.

[0022] The model training method based on vertical federated learning provided by the present invention ensures the integrity of data shards through hash verification; prevents malicious nodes from disguising participation through digital signature authentication; and accelerates model convergence by dynamically adjusting the aggregation weights.

[0023] In an alternative embodiment, the performing shard integrity verification, digital signature authentication, and zero - knowledge proof verification on each encrypted data includes:

[0024] Compare the hash values of each encrypted data with the hash values provided by the corresponding participating nodes to perform shard integrity verification;

[0025] Match the decrypted signatures of the public keys of each participating node with the pre - stored certificates of the coordination node to perform digital signature authentication;

[0026] Confirm whether the encrypted data meets the preset format and privacy compliance requirements to perform zero - knowledge proof verification.

[0027] The model training method based on vertical federated learning provided by the present invention quickly locates damaged shards through hash value comparison (for example, re - transmit shards when the hash values do not match); ensures the credibility of data sources through public key certificate matching (for example, only allows registered nodes to participate in training); and verifies that the data does not contain sensitive fields through zero - knowledge proof.

[0028] In an alternative embodiment, the method further includes:

[0029] Generate a corresponding identification code for the global model parameters after each update; the identification code includes the current training round, timestamp, and the hash value of the global model parameters;

[0030] After receiving the global model parameters, the participating node compares the identification codes to confirm version consistency, and if they are inconsistent, trigger a parameter synchronization request.

[0031] The model training method based on vertical federated learning provided by the present invention accurately marks the model version through a unique identification code to avoid confusion of historical versions; when the identification codes do not match, parameter rollback is triggered (for example, rolling back to the previous round of parameters) to prevent model collapse.

[0032] In an alternative embodiment, the method further includes:

[0033] Dividing the global model parameters into multiple shards;

[0034] Each participating node only downloads the parameter shards related to its own features for local training.

[0035] The model training method based on vertical federated learning provided by the present invention reduces the download of invalid data through sharded transmission; only the relevant shards need to be loaded for local training, reducing memory occupancy.

[0036] In an alternative embodiment, the termination condition includes global model convergence or reaching the maximum number of iterations.

[0037] The model training method based on vertical federated learning provided by the present invention has a dynamic convergence determination (such as the loss change rate < threshold) to avoid premature termination or overtraining; the maximum number of iterations can be adjusted according to the hardware configuration.

[0038] In summary, the model training method based on vertical federated learning provided by the present invention realizes a comprehensive improvement in privacy protection, training efficiency, and system stability through the synergistic effect of multi-dimensional technical features: First, by obtaining the encrypted intermediate results (such as gradients, loss values) of participating nodes and restricting their types, while avoiding the leakage of original data, it reduces invalid communication and improves transmission efficiency; Second, combined with the standardization, feature dimensionality reduction, and screening operations in local preprocessing, it eliminates data noise and redundant features, ensures the reliability of model input, and reduces the risk of overfitting; Further, symmetric / asymmetric encryption and multi-layer security verification mechanisms are adopted, including shard hash verification, digital signature authentication, and zero-knowledge proof, to block tampering attacks and malicious node access, ensuring data integrity and the legitimacy of the identities of participating parties; At the same time, based on dynamic aggregation weight allocation and model identification code generation, it optimizes the parameter update priority according to the real-time state of the model, and realizes accurate version traceability and abnormal rollback through unique identification codes to prevent the training process from being interrupted due to version conflicts; In addition, the combination of sharded parameter transmission and dynamic termination conditions significantly reduces the node computing load and bandwidth consumption, and adaptively balances the number of training rounds and model accuracy. The organic integration of the above technical features systematically solves the core defects of existing vertical federated learning solutions in data security, dynamic scheduling, and version management, providing an efficient, robust, and compliant technical foundation for cross-institutional collaborative training.

[0039] Second aspect, the present invention provides a model training device based on vertical federated learning, which is applied to a distributed network including a coordination node and multiple participating nodes. This device is executed by the coordination node and includes:

[0040] An acquisition module, configured to acquire the encrypted data of each participating node; the encrypted data includes intermediate results; the intermediate results are generated by each participating node based on the locally preprocessed data and the current global model parameters;

[0041] An update module, configured to aggregate the encrypted data of each participating node, update the global model, and generate updated global model parameters;

[0042] A distribution module, configured to transmit the updated global model parameters to each participating node;

[0043] A termination module, configured to determine whether the termination condition is satisfied at this time. If so, the training of the global model is completed.

[0044] Third aspect, the present invention provides a computer device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to execute the model training method based on vertical federated learning according to the first aspect or any corresponding embodiment thereof.

[0045] Fourth aspect, the present invention provides a computer-readable storage medium, on which computer instructions are stored. The computer instructions are used to cause a computer to execute the model training method based on vertical federated learning according to the first aspect or any corresponding embodiment thereof.

[0046] Fifth aspect, the present invention provides a computer program product, including computer instructions, which are used to cause a computer to execute the model training method based on vertical federated learning according to the first aspect or any corresponding embodiment thereof. Description of the Drawings

[0047] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required to be used in the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0048] Figure 1 is a schematic flowchart of the model training method based on vertical federated learning according to an embodiment of the present invention;

[0049] Figure 2It is a structural block diagram of a secure data management and model training system based on vertical federated learning according to an embodiment of the present invention;

[0050] Figure 3 It is a schematic flowchart of a method for secure data management and model training based on vertical federated learning according to an embodiment of the present invention;

[0051] Figure 4 It is a structural block diagram of a model training device based on vertical federated learning according to an embodiment of the present invention;

[0052] Figure 5 It is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed implementation manners

[0053] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0054] First, the terms involved in this application are introduced.

[0055] VFL (Vertical Federated Learning): A federated learning architecture applicable to datasets with the same users but different features held by different institutions. Each party only shares intermediate calculation results instead of raw data, thus protecting data privacy.

[0056] Encrypted Storage: Use encryption algorithms such as AES and RSA to protect data storage and prevent data leakage. Key management adopts a distributed key exchange protocol (such as Diffie-Hellman) to ensure security.

[0057] Message Authentication Code (MAC): By attaching a digital signature and an integrity verification mechanism, it ensures that the data has not been tampered with during the transmission process and improves data security.

[0058] Zero-Knowledge Proof (ZKP): A cryptographic technology that verifies the authenticity of data without revealing the specific data content, ensuring the compliance and privacy of training data.

[0059] End-to-End Encryption (E2EE): Ensures the encryption protection of data throughout the transmission link. Only the sender and the receiver can decrypt it, preventing man-in-the-middle attacks.

[0060] With the rapid development of big data and artificial intelligence, the need for cross-organization data sharing and collaborative modeling is growing. However, data privacy and security issues limit the application of traditional centralized model training, especially in highly sensitive fields such as healthcare and finance. As an emerging technology, VFL (Vertical Federated Learning) allows multiple parties to keep data locally and achieve collaboration through distributed model training, which not only improves model performance but also protects data privacy, and gradually becomes an important method to solve the data silo problem. The core of federated learning lies in distributed model update and aggregation. Participants upload intermediate results (such as model weights or gradients) to the central server for global model training. However, this process faces challenges such as data leakage risk, high communication cost, and difficulties in model conflict and version management. Therefore, how to ensure the security, stability, and efficiency of model training while protecting data privacy has become a key technical problem in the application of vertical federated learning.

[0061] Current mainstream vertical federated learning solutions are mainly based on the Federated Averaging (FedAvg) algorithm. The client uploads model weights, and the server performs global aggregation and model update. However, FedAvg lacks end-to-end encryption and privacy protection mechanisms and is vulnerable to man-in-the-middle attacks and data theft. In addition, the lack of a model conflict and version management mechanism may lead to data inconsistency and model corruption. To improve security, some enhanced solutions introduce techniques such as differential privacy, homomorphic encryption, and model perturbation, which improve the level of data protection. However, these methods usually bring high computational costs and synchronization delays, weakening the real-time performance and scalability of the system. In addition, existing solutions mostly focus on data encryption and privacy protection and do not fully address key issues in practical applications such as model identity management, version control, and automatic fault recovery, which limits the wide application of federated learning in complex distributed environments.

[0062] Existing vertical federated learning solutions mainly face multiple challenges in data security and model management. First, the lack of an end-to-end data encryption and privacy protection mechanism poses a risk of sensitive data leakage during transmission and storage. Second, the model conflict detection and version management mechanisms are imperfect, resulting in incorrect model parameter overwriting, data loss, and unstable training processes. In addition, the system lacks an automatic fault recovery and rollback mechanism in unstable network environments, making it difficult to ensure the continuity of model training and the integrity of data processing. These drawbacks severely limit the practical application of federated learning in security-sensitive and distributed environments. At the same time, existing solutions are insufficient in model synchronization and conflict repair in multi-party collaboration. The issues of version inconsistency and parameter conflicts in model training have not been effectively resolved, making it difficult for the system to support long-term cross-organizational model training. In addition, the lack of model full-life cycle management functions, including model identification management, version control, and historical record tracing mechanisms, leads to low transparency and auditability in model development. These problems limit the application potential of federated learning systems in large-scale complex environments.

[0063] Therefore, to address the core issues of data security, model conflicts, and fault recovery in existing vertical federated learning systems, an embodiment of the present invention provides a model training method based on vertical federated learning, which is applied to a distributed network including a coordination node and multiple participating nodes. This method is executed by the coordination node, and the process of this method is as Figure 1 shown, and includes the following steps:

[0064] S101. Obtain the encrypted data of each participating node; the encrypted data includes intermediate results; the intermediate results are generated by each participating node based on the locally preprocessed data and the current global model parameters.

[0065] Specifically, the coordination node (such as a central server) is responsible for aggregating data, updating the global model, and coordinating the training process. The participating nodes are data holders (such as local clients), which only locally process data and upload encrypted intermediate results without exposing the original data. The participating nodes encrypt and transmit the locally calculated intermediate results (such as gradients, loss values) to prevent theft or tampering during transmission. Local data preprocessing includes operations such as data cleaning (filling missing values, removing outliers), feature standardization (normalization), and feature selection (filtering high-value features). The generation of intermediate results is based on the current global model parameters (issued by the coordination node) to perform forward / backward propagation on the local data to generate intermediate variables required for model update.

[0066] S102. Aggregate the encrypted data of each participating node, update the global model, and generate updated global model parameters.

[0067] Specifically, the coordination node performs mathematical operations (such as weighted average) on the received encrypted intermediate results and combines the local update information of each participant. The model parameters are adjusted according to the aggregation result. For example, the model is iteratively optimized by the gradient descent method. The updated parameters will be used as the benchmark for the next round of training to ensure the gradual convergence of the model.

[0068] S103. Transmit the updated global model parameters to each participating node.

[0069] Specifically, the coordination node distributes the newly generated global parameters to all participating nodes to ensure that each node conducts the next round of training based on the same version of the model. Only the model parameters (such as the weight matrix) are transmitted, without involving any plaintext of the original data or intermediate results.

[0070] S104. Determine whether the termination condition is met at this time. If it is met, the training of the global model is completed.

[0071] Specifically, the termination condition is a preset stop criterion for the global model training. For example: the change rate of the model loss function is lower than the threshold (such as the loss decrease rate < 1e-5 for 3 consecutive rounds) or the maximum number of training rounds is reached (such as 100 rounds). After termination, the final model is output for each participant to deploy and use. In addition, the termination condition also includes abnormal termination, such as detecting system errors (such as inconsistent parameters, node disconnection timeout). Abnormal termination means a temporary interruption of the training process, rather than a permanent termination. Through the version rollback and breakpoint resumption mechanism, the model can resume from the breakpoint and finally complete the training.

[0072] It should be noted that in the above steps, if the termination condition is not met, the training process continues (i.e., return to step S101 to obtain a new round of encrypted intermediate results and loop through the subsequent steps) until the termination condition is met.

[0073] The model training method based on vertical federated learning provided in this embodiment solves the problem of privacy leakage by having the participating nodes provide encrypted intermediate results; realizes the optimization of the global model by having the coordination node aggregate encrypted data to generate new parameters; ensures training consistency by distributing the new parameters to all nodes; and ends the training according to the preset conditions to balance efficiency and accuracy. Through the transmission of encrypted intermediate results and the dynamic update of the global model, the training efficiency is improved on the premise of ensuring data privacy; combined with sharded parameter synchronization and adaptive termination condition judgment, the resource utilization rate is optimized and version conflicts are prevented, realizing the coordinated improvement of privacy, efficiency and stability in vertical federated learning.

[0074] Optionally, in the above step S101, the intermediate result specifically refers to the model update-related data participating in the local calculation of the node. The model parameter gradient refers to the weight adjustment amount generated by backpropagation; the loss function value refers to the prediction error of the current model on the local data (such as cross-entropy loss); the intermediate activation value refers to the output result of the hidden layer in forward propagation (such as the feature map after ReLU activation). In local data preprocessing, data standardization performs Z-Score normalization on features to eliminate dimensional differences; feature dimensionality reduction uses principal component analysis (PCA) to retain the main feature directions; feature selection filters high-discrimination features based on information gain or chi-square test. The above encryption process uses a symmetric encryption algorithm or an asymmetric encryption algorithm for encryption. Specifically, for symmetric encryption, such as AES-256, the same key is used for encryption / decryption, which is suitable for high-frequency data transmission; for asymmetric encryption, such as RSA-2048, the public key is used for encryption and the private key is used for decryption, which is used for key exchange or signature verification.

[0075] Optionally, in the above step S102, before aggregating the encrypted data of each participating node, it further includes: performing shard integrity verification, digital signature authentication, and zero-knowledge proof verification on each encrypted data; if the verification is successful, determining the priority weights of each participating node based on the current loss change rate of the global model, the sample gradient contribution degree, and the feature importance score; allocating the aggregation weights of each participating node according to the priority weights of each participating node; aggregating the encrypted data of each participating node according to the aggregation weights; if the verification fails, triggering data retransmission. Shard integrity verification is to detect tampering by comparing shard hash values (such as SHA-256); digital signature authentication is to verify the signature using the sender's public key (such as the ECDSA algorithm); dynamic weight allocation is to adjust the aggregation weights according to the node data volume or contribution degree (such as gradient magnitude).

[0076] In the above steps, performing shard integrity verification, digital signature authentication, and zero-knowledge proof verification on each encrypted data includes: comparing the hash values of each encrypted data with the hash values provided by the corresponding participating nodes for shard integrity verification; matching the decrypted signatures of each participating node with the pre-stored certificates of the coordination node for digital signature authentication; and confirming whether the encrypted data meets the preset format and privacy compliance requirements for zero-knowledge proof verification.

[0077] Optionally, in the above step S102, after updating the global model parameters, a corresponding identification code is generated for the updated global model parameters each time; the identification code includes the current training round, timestamp, and hash value of the global model parameters; after receiving the global model parameters, the participating nodes compare the identification codes to confirm the version consistency, and if they are inconsistent, a parameter synchronization request is triggered. The identification code generation rule is to record the iteration progress based on the training round (round_id); the timestamp marks the parameter update time; the parameter hash value (such as SHA-256) verifies the data consistency.

[0078] Optionally, in the above step S103, before distributing the updated global model parameters, the global model parameters are divided into multiple shards; each participating node only downloads the parameter shards related to its own features for local training. The sharding is performed on the feature dimensions. For example, node A is responsible for the feature subset F1 - Fk and only transmits the corresponding parameter shards; the distributed shards are also encrypted during transmission, such as using AES-256 to encrypt the shards and attaching the version identification code.

[0079] In summary, the model training method based on vertical federated learning provided by the embodiments of the present invention realizes a comprehensive improvement in privacy protection, training efficiency, and system stability of vertical federated learning through the synergistic effect of multi-dimensional technical features: First, by obtaining the encrypted intermediate results (such as gradients, loss values) of the participating nodes and restricting their types, it reduces ineffective communication while avoiding the leakage of raw data, improving the transmission efficiency; second, combined with the standardization, feature dimensionality reduction, and screening operations in local preprocessing, it eliminates data noise and redundant features, ensures the reliability of the model input, and reduces the risk of overfitting; further, it adopts symmetric / asymmetric encryption and multi-layer security verification mechanisms, including shard hash verification, digital signature authentication, and zero-knowledge proof, to block tampering attacks and malicious node access, ensuring data integrity and the legitimacy of the participating party identities; at the same time, based on dynamic aggregation weight distribution and model identification code generation, it optimizes the parameter update priority according to the real-time state of the model, and realizes accurate version tracing and abnormal rollback through the unique identification code, preventing the training process from being interrupted due to version conflicts; in addition, the combination of sharded parameter transmission and dynamic termination conditions significantly reduces the node computing load and bandwidth consumption, and adaptively balances the training rounds and model accuracy. The organic integration of the above technical features systematically solves the core defects of existing vertical federated learning solutions in data security, dynamic scheduling, and version management, providing an efficient, robust, and compliant technical foundation for cross-institutional collaborative training.

[0080] Based on the model training method based on vertical federated learning provided in the above embodiments, this embodiment also provides a secure data management and model training system based on vertical federated learning. The architecture of this system is as Figure 2As shown, the system is applied to a distributed network including a central server and multiple clients. The architecture includes: a data preparation module, a data transmission module, a data verification module, a model training module, a model synchronization module, and a fault recovery module. These modules are deployed in the central server and multiple clients. The functions of each module are as follows:

[0081] The data preparation module is responsible for loading, cleaning, feature selection, and encrypted storage of data from each client to improve the quality and security of model training. It includes four sub-modules: data loading, data cleaning, feature selection, and encrypted storage. The data loading sub-module extracts data from the local database; the data cleaning sub-module performs missing value filling and feature scaling operations; the feature selection sub-module optimizes data quality by screening important features and samples; the encrypted storage sub-module encrypts the preprocessed data using AES / RSA to protect privacy.

[0082] The data transmission module is used to send the data processed by the client to the server and ensure the security and integrity of the data during transmission. It includes three sub-modules: data sending, format checking, and retransmission request. The data sending sub-module batch-transmits the encrypted sharded data; the format checking sub-module verifies the packet structure and arrival order; the retransmission request sub-module re-sends the data when it is missing or damaged.

[0083] The data verification module is used to ensure the integrity, legality, and privacy protection of the data packets sent by the client on the server side. It includes three sub-modules: signature verification, integrity check, and privacy authentication. The signature verification sub-module checks the digital signature of the data packet to ensure a legitimate source; the integrity check sub-module uses a hash algorithm to verify data integrity; the privacy authentication sub-module protects data privacy through zero-knowledge authentication.

[0084] The model training module is used to perform global model training and model identification management to achieve model optimization and tracking. It includes four sub-modules: sample distribution, model training, identification generation, and record storage. The sample distribution sub-module assigns data priorities according to importance; the model training sub-module performs model weight update and aggregation; the identification generation sub-module generates a unique model identification through a hash algorithm; the record storage sub-module records the model training process in a distributed database.

[0085] The model synchronization module is used to synchronize model parameters and identifications between the client and the server to ensure consistent model versions. It includes three sub-modules: model broadcasting, conflict checking, and automatic repair. The model broadcasting sub-module sends the updated model parameters and identifications to the client; the conflict checking sub-module detects model conflicts based on timestamps and version numbers; the automatic repair sub-module performs model rollback and parameter repair when conflicts occur.

[0086] The fault recovery module is used to automatically recover the model and training process when the server encounters an exception, ensuring the stable operation of the system. It includes four sub-modules: model storage, life cycle management, recovery mechanism, and log recording. The model storage sub-module stores the model parameters and identification information in a distributed database; the life cycle management sub-module controls the activation, expiration, and destruction of the model; the recovery mechanism sub-module automatically rolls back the model version when a system fault is detected; the log recording sub-module saves the operation logs and recovery reports to support system auditing and fault analysis.

[0087] The model training process of this system is as Figure 3 shown. First, each client performs local data preprocessing. That is, the client standardizes the local data, selects features, filters samples, and encrypts and stores sensitive fields using AES / RSA. Calculate intermediate results such as the initial gradient to ensure data privacy. Subsequently, each client sends encrypted data such as gradients to the server through end-to-end encryption (E2EE) and sharded transmission. The server verifies the data integrity (MAC, hash value) and validates the identity (digital signature) to ensure secure transmission. After that, the server evaluates the data quality and feature importance, and dynamically adjusts the sample priority. High-priority samples enter the high-speed processing queue to accelerate model convergence. The server aggregates the encrypted gradients uploaded by each client and calculates the global model update. Generate a unique model identifier (SHA-256) and store it in the distributed database to ensure version management and security. After the global model update is completed, the server broadcasts the new model parameters to the clients and uses timestamps to ensure consistency. If parameter mismatches are found, a conflict resolution mechanism is triggered to roll back to the correct version. In addition, the trained model is stored in the distributed storage system, and a mapping between the model identifier and the storage location is established. Manage the life cycle (activation, expiration, destruction) according to the usage frequency to optimize storage resources. In addition, if the server detects training anomalies (such as client disconnection, calculation failure), an automatic rollback mechanism is triggered. Roll back to the most recent successful version, notify the clients to rejoin the training, and ensure system stability.

[0088] The specific process of the secure data management and model training mechanism of this system is as follows:

[0089] When the VFL model training starts, the client first performs standard preprocessing on the local dataset. This includes steps such as missing value filling, feature scaling, and categorical encoding to ensure that the data formats and distributions between different data sources are as consistent as possible, thereby improving the training effect and stability of the model. In addition, for high-dimensional feature data, the participating parties use dimensionality reduction algorithms (such as PCA) to reduce the data scale and improve the subsequent processing efficiency. During the data preprocessing process, the system will perform feature selection and sample screening operations. Feature selection removes irrelevant or redundant features through feature importance scoring (such as information gain, chi-square test, etc.), thereby reducing data noise and training complexity. Sample screening filters out unqualified samples based on preset rules (such as data quality scoring or anomaly detection algorithms) to ensure the accuracy and robustness of model training. The preprocessed data is stored locally in encrypted form, and sensitive fields are encrypted using data encryption standards (such as AES or RSA) to ensure data privacy protection during storage and operation. Key management uses a distributed key exchange protocol (such as Diffie-Hellman) to make the data exchange between participating parties secure and reliable, so that the transmitted content cannot be decrypted even if the network is monitored. During the model training process, each participating party securely sends the intermediate results calculated locally (such as model gradients, loss values, etc.) to the server through an encrypted transmission protocol. To improve the transmission efficiency and privacy protection level, the intermediate results will be encrypted, fragmented, and sent in batches before transmission. This transmission strategy not only reduces communication latency but also effectively avoids potential data leakage risks. Finally, to ensure the integrity and security of data during transmission, the system adopts an end-to-end encryption and message authentication code (MAC) mechanism. Each data fragment is attached with a digital signature when sent, and the receiving end performs integrity verification and identity authentication when the data arrives to prevent data tampering and forgery, ensuring the high security and reliability of the training data.

[0090] When the server receives the encrypted data uploaded by each participating party, it first enters the data reception buffer for preliminary verification. The server checks the arrival order and structural integrity of the data packets according to the predefined transmission batches and sharding rules. If it is found that the data packets are incomplete or out of order, the server immediately sends a request to retransmit the missing or damaged data segments to ensure the consistency and continuity of the batch transmission. Subsequently, the server authenticates the digital signatures of the data packets. Each data packet has been attached with a unique digital signature when sent, and the server verifies the legality of the signature through the public key of the participating party. If the signature verification fails, the server will regard it as a potential security threat, terminate the current data processing process, record the suspicious activity in the log, trigger the security alert mechanism, and notify the administrator for further investigation. After the authentication is passed, the server uses a multiple integrity verification mechanism to deeply verify the data. The hash verification algorithm (such as SHA-256) is used to calculate the hash of the received data segments and compare it with the hash value provided by the sender. This two-way verification mechanism can effectively prevent the data from being maliciously tampered with during the transmission process and ensure the integrity and accuracy of the data. In addition, to improve data privacy and security, the server enables the zero-knowledge proof protocol. Without exposing the data content, the server and the client jointly execute an interactive verification based on zero-knowledge proof to ensure that the uploaded data meets the model training requirements, further improving the data security and compliance of the system and avoiding the leakage of sensitive information. If data anomalies are detected at any stage, including signature verification failure, hash value mismatch, or zero-knowledge proof verification failure, the server will immediately trigger the data retransmission mechanism. According to the recorded transmission log and the model status, the server automatically resumes to the last successful state, re-requests the missing or damaged data shards, and rolls back according to the system policy to ensure the continuity and security of the model training process.

[0091] During the training process of the VFL model, after the server receives the complete samples uploaded by each participant, it will initiate a sample scheduling and priority distribution mechanism to maximize the training efficiency and accuracy of the model. First, the server conducts basic screening and pre-evaluation on the currently received sample data, including data format verification, sample integrity check, and preliminary data quality analysis, to ensure that the sample data meets the training standards. Next, the server performs a sample priority assessment according to the training requirements of the model and the current state of the model. This assessment is based on multiple metrics, including the current loss change rate of the model, the model convergence speed, the sample feature importance score, and the historical performance of the sample in model training. For example, samples with a large gradient contribution or those that play a key role in the model loss change will be assigned a higher priority to accelerate the model convergence process. To achieve dynamic scheduling, the server adopts sample priority distribution algorithms, such as queue-based dynamic scheduling strategies or reinforcement learning algorithms. The server places high-priority samples in a high-speed processing queue to ensure they are processed first in the next round of training iterations. For ordinary samples, the server assigns them to a regular processing queue to participate in subsequent training processes as needed. This scheduling mechanism improves the overall efficiency of sample transmission and processing while ensuring the comprehensiveness of model training. In addition, the server will periodically evaluate the effectiveness of the sample scheduling strategy. By monitoring the real-time training performance of the model (such as the loss convergence curve and the change in model accuracy), the server can dynamically adjust the sample scheduling strategy. If it is found that the priority setting of a certain type of sample does not achieve the expected effect, the system will automatically correct the sample distribution strategy to improve the adaptability and flexibility of model training. Finally, to avoid resource competition and transmission bottlenecks caused by high load, the server will intelligently allocate the sample transmission batches and frequencies in combination with the current load status of the system and the network transmission capacity. This mechanism ensures the stability and scalability of the training process. Even in a large-scale distributed learning environment with multiple participants, sample scheduling and priority distribution can operate efficiently, continuously improving the training performance of the model.

[0092] In each round of the global model update phase, the server performs global model updates based on the current training round and the received sample data. The server calculates the new model weights through weighted averaging or other aggregation algorithms, thereby improving the overall performance and generalization ability of the model. This process involves the model parameters and gradients uploaded by all participating parties, ensuring the effectiveness of data diversity and algorithm optimization. While the model is being updated, the model identification management module activates the identification generation mechanism. Based on the current training round, timestamp, and hash value of the model parameters, this module uses a cryptographic hash algorithm (such as SHA-256) to generate a unique global model identification. This identification serves as the sole marker for the model version, used for model traceability and iterative management, preventing model version conflicts and confusion of historical versions. To ensure the security and integrity of the model version, the server stores the model identification of each participating party together with the model parameters in a distributed database through an end-to-end encryption algorithm. The model identification is protected by digital signatures and encryption certificate mechanisms to prevent malicious tampering and data leakage. This security mechanism ensures data compliance and transparency during the model training process. Additionally, the server also maintains a model version control log, recording the detailed information of each model update, including the timestamp, training round, model identification, and performance metrics. Through the log traceability function, system administrators and developers can easily check the historical changes of the model, ensuring the transparency and controllability of the model development process. Finally, if any anomalies occur during the model update process, such as data inconsistency, training interruption, or performance degradation, the server will trigger an automatic rollback mechanism. By comparing the latest model identification with historical versions, the system can quickly revert to the model state of the most recent successful training, ensuring the continuity and stability of the model training process.

[0093] After the model training is completed, the server broadcasts the updated model parameters and model identification to all participating parties to ensure the consistency of the model version. The server confirms the reception status of each participating party through timestamp alignment and multi-party digital signature mechanisms, avoiding duplication or omission of model versions.

[0094] To prevent model conflicts, the server performs two-way verification on model parameters after receiving confirmations from participants. If parameter mismatches are detected, the system immediately triggers a conflict resolution strategy, recovering the correct parameters through historical version comparison and an automatic repair algorithm to avoid incorrect updates overwriting the training results. In the case of an unstable network environment, the server automatically switches to an asynchronous transmission mode. This mode supports breakpoint resumption and retransmission of models and samples, ensuring that all participants can automatically synchronize to the latest model state after the network resumes, reducing data loss during the model parameter update process. During the synchronization process, the server maintains a real-time model version log, recording the time of each synchronization, model identifier, and parameter change history, ensuring transparency and traceability in version management. This log mechanism provides an important historical basis for anomaly recovery and problem diagnosis. After synchronization is complete, the server broadcasts the performance evaluation results of the latest model to all participants. If a decrease in model performance or version mismatch is detected, the system will trace back the model state based on the log and perform necessary version rollbacks and retraining to ensure stable and reliable model performance.

[0095] The trained models and their corresponding identifiers are securely stored in a distributed model management system. This system ensures the security and persistence of model data at both the physical and logical levels through a multi-layer data redundancy and encrypted storage mechanism. A mapping relationship between each model identifier and the storage location of the model file is established in the database to support fast retrieval and reading. According to the model lifecycle management strategy, the system classifies and manages different states of model identifiers, including activation, expiration, and destruction of identifiers. An activated model identifier supports the normal use and invocation of the model, while an expired identifier will enter a restricted storage area, restricting external access to prevent misuse or leakage. For model identifiers that have exceeded the storage period or are no longer in use, the system will automatically execute a secure destruction process. This process includes the secure erasure of physical storage media and the complete deletion of database records to ensure that sensitive model data cannot be recovered or stolen. The entire destruction process is logged using encryption to support auditing and compliance checks. To avoid multi-version conflicts and redundant storage issues in model management, the system periodically performs model version checks and automatic archiving. The system decides which versions need to be retained long-term and which versions should be archived or deleted based on the usage frequency, update history, and performance of the models, thus achieving efficient management and utilization of storage resources. In addition, the system also supports a lifecycle event notification function for model versions. When the state of a model changes, such as identifier activation, expiration, or destruction, the system will send notifications to relevant administrators to ensure that the management team can promptly understand the lifecycle dynamics of the models, facilitating subsequent decision-making and operations.

[0096] During the training process of the VFL model, if an abnormal event occurs (such as a participant interruption, network instability, or server failure), the system will immediately activate the fault recovery mechanism. The server monitors the running status logs during the training process, quickly detects the fault point, and records relevant abnormal information to ensure that problems are identified and responded to in a timely manner. Once the fault is confirmed, the server will start the automatic rollback mechanism to restore the model to the version of the last successful training. The system accurately locates the state of the most recent stable model by comparing the latest model identifier with historical model versions, avoiding model data loss and damage to training results, and ensuring the continuity of the model training process. While performing the rollback operation, the server will notify all participants, prompting them to rejoin the model training process. The notification content includes the specific version of the rollback, details of the fault, and recovery steps, helping the participants to resume normal training as soon as possible and minimizing the impact of the interruption on the overall training progress. To improve the accuracy of recovery, the server stores the fault diagnosis and recovery records in a centralized log management system. This system supports in-depth analysis and fault mode prediction, providing effective reference for future abnormal events and enhancing the robustness and adaptability of the system. Finally, after the fault recovery is completed, the system will generate a detailed recovery report, including the fault type, rollback version, repair steps, and model performance verification results. This report will be available for system administrators to review and archive, providing key support for continuous improvement and performance optimization.

[0097] The secure data management and model training system based on vertical federated learning provided by the embodiments of the present invention uses end-to-end encryption, distributed key exchange, and zero-knowledge proof protocols to ensure the privacy and security of data during transmission and storage, and prevent the leakage of sensitive information. Based on the shard transmission and hash verification mechanism, it ensures the integrity of data transmission across clients and servers, and detects and repairs data loss or damage. According to the model loss change rate, gradient importance, and training history records, it dynamically adjusts the sample priority to optimize the model training efficiency and convergence speed. It generates a unique model identifier through timestamps and hash algorithms to manage the life cycle of model training, supports model version tracing, anomaly detection, and automatic rollback. Based on the multi-party signature and version comparison mechanism, it ensures the synchronization consistency of models among all parties, detects and repairs parameter conflicts, and prevents version overwriting and data inconsistency. When a server or client fails, it automatically restores to the version of the last successful model, avoiding data loss and model training interruption. It generates log records for all operations of data transmission, model update, and fault recovery, supporting system performance auditing and compliance checks. It manages the entire life cycle of model parameters and identifiers, including activation, expiration, archiving, and destruction, ensuring the secure storage and automatic cleaning of historical data. By integrating end-to-end encryption, model identifier management, automatic rollback, and multi-party conflict repair mechanisms, it achieves more comprehensive data security and model management, significantly improving the stability and scalability of the vertical federated learning system, and is applicable to complex distributed environments and cross-institutional data collaboration scenarios.

[0098] In this embodiment, a model training device based on vertical federated learning is further provided. This device is used to implement the above-mentioned embodiments and preferred implementation manners, and those that have been described will not be repeated here. As used hereinafter, the term "module" may be a combination of software and / or hardware that can achieve a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0099] This embodiment provides a model training device based on vertical federated learning, as Figure 4 shown, which is applied to a distributed network including a coordination node and multiple participating nodes. This device is executed by the coordination node, and the device includes:

[0100] An acquisition module 401, configured to acquire encrypted data of each participating node; the encrypted data includes intermediate results; the intermediate results are generated by each participating node based on locally preprocessed data and current global model parameters;

[0101] An update module 402, configured to aggregate the encrypted data of each participating node, update the global model, and generate updated global model parameters;

[0102] A distribution module 403, configured to transmit the updated global model parameters to each participating node;

[0103] A termination module 404, configured to determine whether the termination condition is satisfied at this time. If it is satisfied, the training of the global model is completed.

[0104] The further function descriptions of the above-mentioned modules and units are the same as those in the corresponding above-mentioned embodiments, and will not be repeated here.

[0105] The model training device based on vertical federated learning in this embodiment is presented in the form of functional units. Here, the unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and a memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0106] This embodiment of the present invention further provides a computer device having the above-mentioned Figure 4 model training device based on vertical federated learning.

[0107] Please refer to Figure 5 , Figure 5 which is a schematic structural diagram of a computer device provided by an alternative embodiment of the present invention. As Figure 5As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting the components, including a high-speed interface and a low-speed interface. Each component communicates with each other using different buses and can be installed on a common motherboard or installed in other ways as needed. The processor can process instructions executed within the computer device, including instructions stored in the memory or on the memory to display graphical information of the GUI on an external input / output device (such as a display device coupled to the interface). In some alternative embodiments, if necessary, multiple processors and / or multiple buses can be used with multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a set of blade servers, or a multi-processor system). Figure 5 Taking one processor 10 as an example in

[0108] The processor 10 can be a central processing unit, a network processor, or a combination thereof. Among them, the processor 10 can further include a hardware chip. The above hardware chip can be an application-specific integrated circuit, a programmable logic device, or a combination thereof. The above programmable logic device can be a complex programmable logic device, a field programmable gate array, a generic array logic, or any combination thereof.

[0109] Among them, the memory 20 stores instructions executable by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiments.

[0110] The memory 20 can include a program storage area and a data storage area. Among them, the program storage area can store an operating system and application programs required for at least one function; the data storage area can store data created according to the use of the computer device. In addition, the memory 20 can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some alternative embodiments, the memory 20 can optionally include a memory remotely set relative to the processor 10, and these remote memories can be connected to the computer device through a network. Examples of the above network include but are not limited to the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0111] The memory 20 can include a volatile memory, such as a random access memory; the memory can also include a non-volatile memory, such as a flash memory, a hard disk, or a solid-state drive; the memory 20 can also include a combination of the above types of memories.

[0112] The computer device further includes a communication interface 30 for the computer device to communicate with other devices or communication networks.

[0113] Embodiments of the present invention also provide a computer-readable storage medium. The method according to the embodiments of the present invention can be implemented in hardware, firmware, or be implemented as computer code that can be recorded on a storage medium, or be implemented as computer code that is originally stored in a remote storage medium or a non-transitory machine-readable storage medium and downloaded through a network and will be stored in a local storage medium, so that the method described herein can be stored as such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk, or a solid-state drive, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code, and when the software or computer code is accessed and executed by the computer, the processor, or the hardware, the method shown in the above embodiments is implemented.

[0114] A part of the present invention can be applied as a computer program product, for example, computer program instructions, which when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should be able to understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to the computer.

[0115] Although the embodiments of the present invention are described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A model training method based on vertical federated learning, characterized in that Applied to a distributed network including a coordinating node and multiple participating nodes, the method is executed by the coordinating node, and the method includes: Obtaining the encrypted data of each participating node; the encrypted data includes intermediate results; the intermediate results are generated by each participating node based on the locally preprocessed data and the current global model parameters; Aggregating the encrypted data of each participating node, updating the global model, and generating updated global model parameters; Transmitting the updated global model parameters to each participating node; Judging whether the termination condition is satisfied at this time. If it is satisfied, the training of the global model is completed.

2. The method according to claim 1, characterized in that The intermediate results include at least one of the parameter gradient of the global model, the loss value of the global model, and the intermediate activation value of the global model.

3. The method according to claim 2, wherein The preprocessing includes data standardization, feature dimensionality reduction, feature selection based on feature importance scoring, and sample screening.

4. The method according to claim 3, wherein The encrypted data is generated by the participating node through symmetric encryption or asymmetric encryption of the intermediate results.

5. The method according to claim 4, characterized in that, Before the step of aggregating the encrypted data of each participating node, it further includes: Performing shard integrity verification, digital signature authentication, and zero-knowledge proof verification on each encrypted data; If the verification is successful, determine the priority weights of each participating node based on the current loss change rate of the global model, the sample gradient contribution degree, and the feature importance scoring; allocate the aggregation weights of each participating node according to the priority weights of each participating node; and aggregate the encrypted data of each participating node according to the aggregation weights; If the verification fails, trigger data retransmission.

6. The method according to claim 5, characterized in that, The performing shard integrity verification, digital signature authentication, and zero-knowledge proof verification on each encrypted data includes: Comparing the hash value of each encrypted data with the hash value provided by the corresponding participating node to perform shard integrity verification; Matching the decrypted signature of the public key of each participating node with the pre-stored certificate of the coordinating node to perform digital signature authentication; Confirming whether the encrypted data meets the preset format and privacy compliance requirements to perform zero-knowledge proof verification.

7. The method according to claim 6, characterized in that, The method further includes: Generating a corresponding identification code for the updated global model parameters each time; the identification code includes the current training round, the time stamp, and the hash value of the global model parameters; After receiving the global model parameters, the participating node compares the identification codes to confirm the version consistency. If they are inconsistent, a parameter synchronization request is triggered.

8. The method according to claim 7, wherein The method further includes: Dividing the global model parameters into multiple shards; Each participating node only downloads the parameter shards related to its own features for local training.

9. The method according to any one of claims 1 to 8, characterized in that The termination condition includes global model convergence or reaching the maximum number of iterations.

10. A model training device based on vertical federated learning, characterized in that, Applied to a distributed network including a coordinating node and multiple participating nodes, the device is executed by the coordinating node, and the device includes: An obtaining module, configured to obtain the encrypted data of each participating node; the encrypted data includes intermediate results; the intermediate results are generated by each participating node based on the locally preprocessed data and the current global model parameters; An updating module, configured to aggregate the encrypted data of each participating node, update the global model, and generate updated global model parameters; A sending module, configured to transmit the updated global model parameters to each participating node; A termination module is used to determine whether the termination condition is met at this time. If it is met, the training of the global model is completed.

Images