Method and system for determining global optimal splitting in longitudinal federal XGBoost training with hidden gradient sum, computer equipment and medium

Abstract

The invention discloses a method and system for determining global optimal splitting in vertical federal XGBoost training with a hidden gradient sum, computer equipment and a medium, and the method comprises the steps: an active party calculates the gradient of each sample in an instance space, carries out the homomorphic encryption, and broadcasts a gradient ciphertext to all passive parties; the passive party determines a split point scheme and an aggregation gradient ciphertext, performs privacy hiding on the aggregation gradient ciphertext by using a random number, constructs a confusion circuit, and sends the split point scheme, a hidden gradient ciphertext and the confusion circuit to the active party; the active party determines a hidden left aggregation gradient sum and a hidden right aggregation gradient sum, and determines a local optimal split point scheme of all the passive parties by using a confusion circuit; and the active party determines a global optimal splitting point scheme based on all local optimal splitting point schemes. According to the method, security multi-party computing technologies such as homomorphic encryption and a confusion circuit are fused, and privacy protection and model performance are considered on the premise that data of all parties are not out of a domain.

Description

Technical Field

[0001] This invention relates to the field of network information security technology, and in particular to a method, system, computer device, and medium for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums. Background Technology

[0002] Building machine learning models while maintaining user privacy is fundamental to intelligent systems. While training such models heavily relies on user data, most organizations and enterprises lack sufficient data to build high-quality machine learning models. Data sharing and collaboration can address this issue, but privacy remains one of the biggest concerns hindering such cooperation. For example, data owners may hesitate to share raw user data due to conflicts of interest, or may be prohibited from sharing due to policies such as GDPR (GDPR). Federated Machine Learning (FML) technology allows businesses and institutions to collaboratively train models without transmitting raw data by integrating their own datasets with user / customer data (such as customer information stored by banks and insurance companies). Its core mechanism is that each participant first trains a local model, then merges the model parameters into a global model, and finally sends the global model's parameters back to all participants. Through this iterative process, a complete global model is eventually generated. It is important to note that the communication required to compute the global model only includes local model parameters, gradient information, and some training results, without transmitting the raw data. Although collaborators do not transmit raw data in FML protocols, this does not automatically guarantee data privacy. The transmitted numerical values ​​themselves (such as model parameters, parts of the model, and gradients) may still leak information that could be used to infer the original data. In fact, a growing body of research indicates that this indirect information leakage can reveal important information that should otherwise be protected.

[0003] Federated machine learning can be divided into horizontal federated learning (different samples, same features), vertical federated learning (same samples, different features), and federated transfer learning (different samples and features). Among them, vertical federated learning has significant application value in scenarios such as financial risk control and joint marketing because it can achieve feature complementarity. The general operating mechanism of most current vertical federated XGBoost protocols is as follows: During the iterative tree node splitting process, all parties jointly determine the optimal splitting scheme that minimizes the loss function. To ensure data sample privacy, homomorphic encryption, secret sharing, and other techniques are often used to hide sample data and model gradients, achieving the following information exchange: confidential information exchange, privacy-preserving entity alignment, gradient information sharing, possible node splitting, and sub-model updates. The passive party calculates all possible splitting schemes for its feature variables in an aggregate manner and returns all results to the active party. The next step is for the active party to compare all splitting results and find the optimal scheme. In this process, the active party obtains the gradient sum under different splitting schemes and uses it to calculate the loss function and make comparisons. This information leakage (i.e., gradient sum) may reveal important information about the passive party's data samples to the active party, namely, the partial ranking of the passive party's samples in each feature dimension. The active party, as an honest but curious participant, can utilize these gradients and derived sample ranking information, as well as some prior information such as data distribution. This information leakage seriously threatens the data privacy of the passive party, violating the core objective of federated learning: "data does not leave the domain, and privacy is not compromised."

[0004] Therefore, existing technologies still need improvement. Summary of the Invention

[0005] To address the aforementioned deficiencies in existing technologies, this invention provides a method, system, computer device, and medium for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums. The technical solution adopted by this invention is as follows: In a first aspect, the present invention provides a method for determining the globally optimal split during longitudinal federated XGBoost training with hidden gradient sums, the method comprising: During the training of vertical federated XGBoost, the active party calculates the gradient of each sample in the instance space of the current iteration, and after homomorphically encrypting the gradient, broadcasts the generated gradient ciphertext to all passive parties. After receiving the gradient ciphertext, the passive party traverses all its own features, determines the split point scheme and aggregated gradient ciphertext for each feature, uses random numbers to perform privacy hiding on the aggregated gradient ciphertext to obtain the hidden aggregated gradient ciphertext, and constructs a confusion circuit for calculating the split score. The passive party then sends the split point scheme, the hidden gradient ciphertext, and the confusion circuit to the active party. The active party summarizes the total gradient sum of the instance space, and based on the hidden aggregate gradient ciphertext and the total gradient sum, determines the hidden left aggregate gradient sum and the hidden right aggregate gradient sum respectively. Then, using the obfuscation circuit, the hidden left aggregate gradient sum and the hidden right aggregate gradient sum, together with the encrypted random number input by the passive party, the split score of all split point schemes is calculated, and the local optimal split point scheme of all passive parties is determined. The active party determines the globally optimal split point scheme based on all locally optimal split point schemes.

[0006] In one implementation, the active party computes the gradient of the instance space for the current iteration, including: The active party determines the instance space for the current iteration; The active party calculates the first and second gradients of the loss function for each sample in the instance space.

[0007] In one implementation, after homomorphically encrypting the sample gradient, the generated gradient ciphertext is broadcast to all passive parties, including: The first and second gradients are encrypted using the Paillier additive homomorphic encryption algorithm to obtain the first gradient ciphertext and the second gradient ciphertext. Broadcast the first gradient ciphertext and the second gradient ciphertext to all passive parties.

[0008] In one implementation, after receiving the gradient ciphertext, the passive party traverses all its own features to determine the splitting point scheme and aggregated gradient ciphertext for each feature, including: The passive party receives the first and second gradient ciphertexts sent by the active party, and iterates through all its own features and all split point schemes for each feature. For each split point scheme, the instance space is divided into a left instance space and a right instance space. Using the Paillier additive homomorphism property, the first aggregated gradient ciphertext of the first gradient ciphertext and the second aggregated gradient ciphertext of the second gradient ciphertext are calculated.

[0009] In one implementation, the passive party uses random numbers to perform privacy hiding on the aggregated gradient ciphertext to obtain hidden aggregated gradient ciphertext, including: Generate a first random number corresponding to the first gradient and a second random number corresponding to the second gradient; Privacy hiding is performed on the first aggregated gradient ciphertext based on the first random number to obtain the first hidden aggregated gradient ciphertext; Privacy hiding is performed on the second aggregated gradient ciphertext based on the second random number to obtain the second hidden aggregated gradient ciphertext.

[0010] In one implementation, the passive party constructs a confusion circuit for calculating the split score, and sends the split point scheme, the hidden gradient ciphertext, and the confusion circuit to the active party, including: A confusion circuit generation algorithm is invoked to construct a confusion circuit for calculating the split score. The confusion circuit includes: circuit information, encoding information, and decoding information. The split point scheme, the first hidden aggregation gradient ciphertext, the second hidden aggregation gradient ciphertext, and the circuit information and encoding information of the obfuscated circuit are sent to the active party.

[0011] In one implementation, the active party determines the hidden left aggregated gradient sum and the hidden right aggregated gradient sum based on the hidden aggregated gradient ciphertext and the total gradient sum, respectively, including: The Paillier private key is used to decrypt the first hidden aggregate gradient ciphertext and the second hidden aggregate gradient ciphertext to obtain the first hidden left aggregate gradient and the second hidden left aggregate gradient. Based on the total gradient sum, the first hidden right aggregated gradient sum and the second hidden right aggregated gradient sum are derived.

[0012] In one implementation, the obfuscation circuit, the hidden left-aggregated gradient sum, and the hidden right-aggregated gradient sum are used to calculate the splitting score of all splitting point schemes using encrypted first and second random numbers, and to determine the locally optimal splitting point scheme for all passive parties, including: The passive party encrypts the first random number and the second random number, and inputs the first encrypted random number and the second encrypted random number into the obfuscation circuit; The active party uses the encoding information to encode the first hidden left aggregate gradient sum, the second hidden left aggregate gradient, the first hidden right aggregate gradient sum, the second hidden right aggregate gradient sum, and the total gradient sum to obtain the encoded gradient summary information, and inputs the encoded gradient summary information into the obfuscation circuit; Based on the aforementioned confusion circuit, the splitting score of all splitting point schemes for each passive party is output. The active party iterates through all splitting point schemes of each passive party and calculates the splitting score. The splitting point scheme with the highest splitting score is taken as the local best splitting point scheme for each passive party.

[0013] In one implementation, the active party determines the globally optimal split point scheme based on all locally optimal split point schemes, including: The active player collects all the local best splitting schemes and corresponding splitting scores from the passive players. The locally optimal splitting scheme with the highest splitting score is taken as the globally optimal splitting scheme.

[0014] In one implementation, the method further includes: The active party determines the target passive party, target features, and optimal split point corresponding to the globally optimal splitting scheme; The target features and the optimal split point are sent to the target passive party so that the target passive party can determine the association information of the globally optimal splitting scheme.

[0015] In one implementation, when determining the association information of the globally optimal splitting scheme, the target passive party includes: The target passive party divides the current instance space into a left subspace and a right subspace based on the target features and the optimal split point; Create a local lookup table to record the node ID and subspace index corresponding to the optimal splitting scheme, and obtain the association information; The associated information is returned to the initiating party.

[0016] In one implementation, the method further includes: Based on the aforementioned association information, the active party determines the subspace index, splits the current tree node, and associates the split current tree node with the target passive party and the local lookup table; If the current tree depth has not reached the maximum depth, the left and right subspaces are used as new instance spaces, and the next iteration is performed. If the current tree depth has reached the maximum depth, then the training of the current tree is complete.

[0017] In a second aspect, embodiments of the present invention also provide a system for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums, characterized in that the system is used to implement the steps of the method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums as described above, and the system includes: an active side and a passive side; The active party includes: The gradient sharing module is used to calculate the gradient of each sample in the instance space of the current iteration during the training of longitudinal federated XGBoost, and broadcast the generated gradient ciphertext to all passive parties after homomorphically encrypting the gradient. The first local optimal split point scheme determination module is used to summarize the total gradient sum of the instance space, determine the hidden left aggregate gradient sum and the hidden right aggregate gradient sum based on the hidden aggregate gradient ciphertext and the total gradient sum, and use the obfuscation circuit, the hidden left aggregate gradient sum and the hidden right aggregate gradient sum, together with the encrypted random number input by the passive party, to calculate the split score of all split point schemes and determine the local optimal split point scheme of all passive parties. The global optimal split point scheme determination module is used by the active party to determine the global optimal split point scheme based on all local optimal split point schemes. The passive party includes: The gradient privacy hiding module is used to traverse all its own features after receiving the gradient ciphertext, determine the split point scheme and aggregated gradient ciphertext for each feature, use random numbers to perform privacy hiding on the aggregated gradient ciphertext to obtain the hidden aggregated gradient ciphertext, and construct a confusion circuit for calculating the split score. The split point scheme, the hidden gradient ciphertext and the confusion circuit are sent to the active party.

[0018] The second local optimal split point scheme determination module encrypts the first and second random numbers and inputs them into the confusion circuit to calculate the split score of all split point schemes together with the active party.

[0019] In one implementation, the gradient sharing module includes: The gradient calculation unit is used to determine the instance space of the current iteration and calculate the first and second gradients of the loss function for each sample in the instance space. The gradient encryption unit is used to encrypt the first gradient and the second gradient using the Paillier additive homomorphic encryption algorithm to obtain the first gradient ciphertext and the second gradient ciphertext. The gradient ciphertext broadcasting unit is used to broadcast the first gradient ciphertext and the second gradient ciphertext to all passive parties.

[0020] In one implementation, the gradient privacy hiding module includes: The split point scheme determination unit is used to receive the first gradient ciphertext and the second gradient ciphertext sent by the initiator, and traverse all its own features and all split point schemes for each feature. The gradient ciphertext aggregation unit is used to divide the instance space into a left instance space and a right instance space for each split point scheme, and to calculate the first aggregated gradient ciphertext of the first gradient ciphertext and the second aggregated gradient ciphertext of the second gradient ciphertext using the Paillier additive homomorphism property. A random number generation unit is used to generate a first random number corresponding to the first gradient and a second random number corresponding to the second gradient. The privacy hiding unit is used to perform privacy hiding on the first aggregated gradient ciphertext based on the first random number to obtain a first hidden aggregated gradient ciphertext, and to perform privacy hiding on the second aggregated gradient ciphertext based on the second random number to obtain a second hidden aggregated gradient ciphertext.

[0021] In one implementation, the gradient privacy hiding module further includes: The obfuscation circuit construction unit is used to call the obfuscation circuit generation algorithm to construct an obfuscation circuit for calculating the split score. The obfuscation circuit includes: circuit information, encoding information and decoding information. The circuit information sending unit is used to send the split point scheme, the first hidden aggregation gradient ciphertext, the second hidden aggregation gradient ciphertext, the circuit information and encoding information of the obfuscated circuit to the active party.

[0022] In one implementation, the first locally optimal split point scheme determination module includes: The hidden aggregate gradient ciphertext decryption unit is used to decode the first hidden aggregate gradient ciphertext and the second hidden aggregate gradient ciphertext using the Paillier private key to obtain the first hidden left aggregate gradient and the second hidden left aggregate gradient. A hidden aggregate gradient and derivation unit is provided for deriving a first hidden right aggregate gradient and a second hidden right aggregate gradient based on the total gradient sum; The gradient encoding summation unit is used to encode the first hidden left aggregated gradient sum, the second hidden left aggregated gradient, the first hidden right aggregated gradient sum, the second hidden right aggregated gradient sum, and the total gradient sum using the encoding information to obtain encoded gradient summation information, and input the encoded gradient summation information to the obfuscation circuit. The split score calculation unit is used to output the split score of all split point schemes for each passive party based on the confusion circuit. The local optimal split determination unit is used to traverse the split score of all split point schemes for each passive party, and the split point scheme with the highest split score is taken as the local optimal split point scheme for each passive party. The second local optimal split point scheme determination module includes: a random number encryption unit, used to encrypt the first random number and the second random number, and input the first encrypted random number and the second encrypted random number into the obfuscation circuit, and calculate the split score of all split point schemes together with the active party.

[0023] In one implementation, the global optimal split point scheme determination module includes: The split score collection unit is used to collect the local best split schemes and corresponding split scores of all passive players. The global optimal split determination unit is used to determine the locally optimal split scheme with the highest split score as the global optimal split scheme.

[0024] Thirdly, embodiments of the present invention also provide a computer device, wherein the computer device includes a memory, a processor, and a program stored in the memory and executable on the processor for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums. When the processor executes the program for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums, it implements the steps of the method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums in any of the above schemes.

[0025] Fourthly, embodiments of the present invention also provide a computer-readable storage medium, wherein the computer-readable storage medium stores a program for determining the globally optimal split in vertical federated XGBoost training with hidden gradient sums, the program for determining the globally optimal split in vertical federated XGBoost training with hidden gradient sums implementing the steps of the method for determining the globally optimal split in vertical federated XGBoost training with hidden gradient sums as described in any of the above schemes on the computer-readable storage medium.

[0026] Beneficial Effects: Compared with existing technologies, this invention provides a method for determining the globally optimal split in vertical federated XGBoost training with hidden gradient sums. First, during the vertical federated XGBoost training process, the active party calculates the gradient of each sample in the instance space of the current iteration. After homomorphically encrypting the gradient, the generated gradient ciphertext is broadcast to all passive parties. Next, after receiving the gradient ciphertext, the passive party traverses all its features, determines the split point scheme and aggregated gradient ciphertext for each feature, uses random numbers to privacy-hide the aggregated gradient ciphertext, obtains a hidden aggregated gradient ciphertext, and constructs a confusion circuit for calculating the split score. The split point scheme, hidden gradient ciphertext, and confusion circuit are then sent to the active party. Then, the active party summarizes the total gradient sum of the instance space, determines the hidden left aggregated gradient sum and the hidden right aggregated gradient sum based on the hidden aggregated gradient ciphertext and the total gradient sum, and uses the confusion circuit, the hidden left aggregated gradient sum, and the hidden right aggregated gradient sum, along with the encrypted random numbers input by the passive parties, to calculate the split score of all split point schemes and determine the locally optimal split point scheme for all passive parties. Finally, the active party determines the globally optimal split point scheme based on all locally optimal split point schemes. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating a preferred embodiment of a method for determining the globally optimal split during longitudinal federated XGBoost training of hidden gradients, as provided in this invention.

[0028] Figure 2 This is a schematic diagram of a vertical federated learning architecture.

[0029] Figure 3 This is a schematic diagram illustrating the technical principle of the method for determining the globally optimal split in the longitudinal federated XGBoost training of hidden gradient sums, as provided in an embodiment of the present invention.

[0030] Figure 4 The system block diagram for determining the globally optimal split in the longitudinal federated XGBoost training of hidden gradient sums provided in the embodiments of the present invention is shown.

[0031] Figure 5 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, 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 of the invention and are not intended to limit the invention.

[0033] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content, operations, or steps, nor does it require execution in the described order. For example, some operations or steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0034] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0035] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. For example, the first control information and the second control information are only used to distinguish different control information and do not limit their order.

[0036] Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or the order of execution, and that the words "first" and "second" do not necessarily imply that they are different.

[0037] It should also be understood that the terms "and / or" as used in this specification and the appended claims refer to any combination of one or more of the associated listed items and all possible combinations, and include such combinations.

[0038] To address the problems of existing technologies, this invention provides a method for determining the globally optimal split during longitudinal federated XGBoost training using hidden gradient sums. This method can also be applied to terminals, such as computers and other intelligent product terminals. Specifically, as follows... Figure 1 As shown, the method includes the following steps: Step S100: During the training of the vertical federated XGBoost, the active party calculates the gradient of each sample in the instance space of the current iteration, and after homomorphically encrypting the gradient, broadcasts the generated gradient ciphertext to all passive parties.

[0039] Combination Figure 2 As shown, Figure 2 The typical architecture of vertical federated learning is shown, which includes 1 active party and m passive parties (m≥1): the active party holds sample IDs, feature data (such as "fund flow" of a bank), and label data (such as "overdue default label"); each passive party holds sample IDs and independent feature data (such as "purchase frequency" of passive party 1 and "product preference" of passive party m); the parties perform privacy-preserving entity alignment through sample IDs to find common samples (ID1~IDn), and perform collaborative modeling based on the joint features and labels of common samples, with all data remaining locally.

[0040] In the architecture of vertical federated learning, participants possess data in the same sample space but different feature spaces. They securely collaborate on modeling using shared sample data, finding applications in fields such as finance and advertising. Participants in vertical federated learning need to collaborate on data intersection, joint model training, and joint model inference. Furthermore, the more participants, the higher the complexity of the vertical federated learning system. There are two main roles among the participants: the active party has feature data and labeled data, enabling independent modeling; the passive party has feature data but lacks labeled data, thus unable to model independently. Due to privacy regulations and industry standards, data cannot be directly shared between different participants. In such cases, a vertical federated learning solution can be adopted for collaboration, ensuring data remains local and using shared sample data for joint modeling and training. Ultimately, both parties obtain a more powerful model.

[0041] XGBoost stands for eXtreme Gradient Boosting. The basic components of XGBoost are decision trees, which are weak classifiers. Together, they form a strong classifier, XGBoost. The decision trees that make up XGBoost are arranged in a specific order: the generation of the next decision tree takes into account the bias of the previous decision tree, so that the training samples that were misclassified by the previous decision tree receive more attention in the future. Then, the next decision tree is trained based on the adjusted sample distribution. The first step of the federated training process is to use a privacy-preserving entity alignment technique to find common samples among the participants. Then, through the interaction between the active party with sensitive classification labels and the passive party with feature sample values, the XGBoost tree ensemble model is iteratively constructed. When constructing such a regression tree, it starts from a tree with a depth of 0 and continuously splits the leaf nodes until the maximum depth is reached. Specifically, the model uses the following equation to determine the optimal split point of each leaf node (the split point that maximizes equation (1)).

[0042]

[0043] in, and These are the left instance space and the right instance space after the split. and These are the first and second gradients of the model's loss function, respectively. It is the minimum reduction required for division. This is a regularization term. In each iteration, the active party calculates the nodes belonging to the current tree node (...). and The sample gradients of the active party need to be sent to all passive parties in encrypted form (to prevent passive parties from receiving gradient values, which could potentially leak label information). Each passive party iterates through its own features and sample values, uses privacy-preserving computation methods (secret sharing, homomorphic encryption, obfuscated circuits, etc.) to calculate the sum of all possible local split gradients using the encrypted gradient values, and returns all results to the active party. In the next step, the active party calculates the gradients of all received splitting schemes and their corresponding split scores, and finds the optimal global split. It then sends the feature ID and its threshold (split point) to the corresponding passive party. The target passive party partitions the current instance space based on the selected feature value and threshold, creates a local lookup table and adds a record, and then returns the index of the left node after splitting and the instance space to the active party. Finally, the active party splits the current node based on the received instance space and associates the current node with the target party's index and its record ID.

[0044] Based on this, in the process of vertical federated XGBoost training, the active party calculates the gradient of each sample in the instance space of the current iteration, and after homomorphically encrypting the gradient, broadcasts the generated gradient ciphertext to all passive parties.

[0045] In practical applications, combined with Figure 3 As shown, during the initial iteration, the active party determines the instance space for the current iteration; the current tree node is the root node (depth 0), and the current instance space I = {ID1, ID2, ..., ID1000}, which includes all common samples. The active party bases its decisions on local labels. and initial predicted value =0, calculate each sample in the instance space. The first and second gradients of the loss function are calculated for each sample. First gradient With the second gradient First gradient With the second gradient The calculation formula is:

[0046]

[0047] In practical applications, It is 0.5.

[0048] Next, this embodiment uses the Paillier additive homomorphic encryption algorithm to process the first gradient. Second gradient Encryption is performed to obtain the first-gradient ciphertext and the second-gradient ciphertext. Specifically, homomorphic encryption algorithms are cryptographic techniques based on the theory of computational complexity in mathematical problems. They allow specific computations to be performed on ciphertext without first decrypting it, and the result of decrypting the computed ciphertext is consistent with the result of performing the same computation on the plaintext. Its core mathematical definition is as follows: Let plaintext space be The ciphertext space is , and These are public and private key pairs, and the encryption function is... , For any plaintext space, the decryption function is: .

[0049] For plaintext operations If ciphertext operations exist , Let be the number of ciphertext spaces, such that for any The following equation holds true: Then the encryption scheme is said to be effective for computation. It exhibits homomorphic properties. This embodiment employs the Paillier additive homomorphic algorithm, i.e., the above operation... For addition operations, the passive party can calculate the sum of the input gradient values ​​with encrypted input, thus preventing the gradient values ​​from being leaked to the passive party. This represents the ciphertext after homomorphic encryption, for example For input The homomorphic encrypted ciphertext.

[0050] After encrypting the gradient, this embodiment broadcasts the first gradient ciphertext and the second gradient ciphertext to all passive parties. For example, the active party broadcasts the first gradient ciphertext and the second gradient ciphertext to passive parties B1 and B2 via a secure communication channel (such as a TLS 1.3 encrypted channel). Due to the semantic security of Paillier encryption, passive parties B1 and B2 can only obtain the ciphertext and cannot deduce the plaintext first gradient from the ciphertext. With the second gradient This protects the privacy of the active party's tags.

[0051] Step S200: After receiving the gradient ciphertext, the passive party traverses all its own features, determines the split point scheme and aggregated gradient ciphertext for each feature, uses random numbers to perform privacy hiding on the aggregated gradient ciphertext to obtain the hidden aggregated gradient ciphertext, and constructs a confusion circuit for calculating the split score. The split point scheme, the hidden gradient ciphertext, and the confusion circuit are then sent to the active party.

[0052] Specifically, the passive party receives the first and second gradient ciphertexts sent by the active party, and iterates through all its own features and all split point schemes for each feature. For example, the passive party's features are... (Values ​​range from 1 to 50). The percentile method can be used to determine the split points. All features are sorted, and the values ​​corresponding to the 10th, 20th, ..., 90th percentiles are taken as the split points, for a total of 9 split points. =5、 =10、 =15、 =20、 =25、 =30、 =35、 =40、 =45.

[0053] For each split point scheme, the instance space is divided into a left instance space and a right instance space. Utilizing the Paillier additive homomorphism property, the first aggregated gradient ciphertext of the first gradient ciphertext and the second aggregated gradient ciphertext of the second gradient ciphertext are calculated. Specifically, in this embodiment, each split point... For each splitting scheme, the instance space is divided into a left instance space and a right instance space. Then, based on the Paillier additive homomorphism property, the passive side computes the left instance space aggregated gradient ciphertext for each splitting scheme in the ciphertext state, i.e., the first aggregated gradient ciphertext. Second aggregation gradient ciphertext .

[0054] Furthermore, in this embodiment, a first random number corresponding to the first gradient is generated. The second random number corresponding to the second gradient For example, the passive party generates multiple pairs of random numbers using a CSPRNG (Cryptographically Secure Pseudo-Random Number Generator). Each pair of random numbers corresponds to a split point scheme. The role of the random numbers is to hide the true aggregate gradient sum. The passive side obtains the hidden aggregate gradient ciphertext by adding the ciphertext of the aggregate gradient of each split point scheme to the ciphertext of the random number (homomorphic addition). Therefore, in this embodiment, the first hidden aggregate gradient ciphertext can be obtained by privacy hiding the first aggregate gradient ciphertext based on the first random number. Privacy hiding is performed on the second aggregated gradient ciphertext based on the second random number to obtain the second hidden aggregated gradient ciphertext. .

[0055] Furthermore, in this embodiment, the passive party invokes a garbled circuit generation algorithm to construct a garbled circuit for calculating the split score. Yao's Garbled Circuit (GC) is a classic method for secure two-party computation. This method allows two entities, the garbled party and the evaluator, to jointly compute a function based on shared private data. From a macroscopic perspective, the garbled circuit is an encrypted version of a Boolean circuit, its function being to compute a specific function by encrypting and replacing the truth table of each logic gate with a symmetric key. In specific implementation, the garbled party selects a key and computes the garbled circuit, then sends the circuit along with the input data (represented in binary form and specially encrypted using the same key) to the evaluator. To compute the output of the garbled circuit, the evaluator needs to encrypt the input data itself, but it does not know the garbled party's key. This process is achieved through a blind transmission protocol with the garbled party. A Garbled Scheme Four polynomial-time algorithms composition: Obfuscation algorithms, with security parameters and circuit Input is the confused circuit information, output is the confused circuit information. Encoding information and decoding information .

[0056] Encoding algorithm, based on input encoded information and input... The parameter is used to output the obfuscated information. .

[0057] Evaluation algorithm based on the input obfuscated circuit information and obfuscated input The parameter is used to output the obfuscated information. .

[0058] Decoding algorithm, based on the input decoding information and obfuscated output The parameter is used to output plaintext information.

[0059] Next, in this embodiment, the split point scheme, the first hidden aggregation gradient ciphertext, the second hidden aggregation gradient ciphertext, the circuit information and encoding information of the obfuscation circuit are sent to the active party.

[0060] Step S400: The active party summarizes the total gradient sum of the instance space, and determines the hidden left aggregate gradient sum and the hidden right aggregate gradient sum based on the hidden aggregate gradient ciphertext and the total gradient sum. Then, using the obfuscation circuit, the hidden left aggregate gradient sum and the hidden right aggregate gradient sum, together with the encrypted random number input by the passive party, the split score of all split point schemes is calculated, and the local optimal split point scheme of all passive parties is determined.

[0061] In this embodiment, the active party can first calculate the current instance space. Total gradient of all samples , Next, decrypt using the Paillier private key (the decryption function is...). Decoding the first hidden aggregated gradient ciphertext and the second hidden aggregated gradient ciphertext yields the first hidden left aggregated gradient (represented as: And the second hidden left aggregate gradient (represented as: Next, based on the total gradient sum, the first hidden right aggregated gradient sum is derived (denoted as: ) and the second hidden right-aggregating gradient (represented as: ).

[0062] Further, the passive party encrypts the first and second random numbers and inputs them into the obfuscation circuit. Next, the active party uses the encoded information to encode the first hidden left aggregate gradient sum, the second hidden left aggregate gradient, the first hidden right aggregate gradient sum, the second hidden right aggregate gradient sum, and the total gradient sum, obtaining encoded gradient summary information, which is then input into the obfuscation circuit. Finally, based on the obfuscation circuit, the splitting score of all splitting point schemes for each passive party is output. Finally, the active party iterates through the splitting scores of all splitting point schemes for each passive party and selects the splitting point scheme with the highest splitting score as the locally optimal splitting point scheme for each passive party.

[0063] The formula for calculating the split score using the confusion circuit in this embodiment is expressed as follows: .

[0064] Based on the above formula, the current feature can be calculated. Below the split point Split scores. Finally, in this embodiment, the active party collects the locally optimal splitting schemes and their corresponding splitting scores from all passive parties. Then, the locally optimal splitting scheme with the highest splitting score is taken as the globally optimal splitting scheme, such as (current instance space). In the passive party Characteristics of holding From the split point (To split).

[0065] Furthermore, in other implementations, in this embodiment, after obtaining the globally optimal splitting scheme, the active party can also determine the target passive party, target features, and optimal split point corresponding to the globally optimal splitting scheme. Then, the target features and the optimal split point are sent to the target passive party so that the target passive party can determine the association information of the globally optimal splitting scheme. When determining the association information of the globally optimal splitting scheme, the target passive party divides the current instance space into a left subspace and a right subspace based on the target features and the optimal split point. Then, a local lookup table is created to record the node ID and subspace index corresponding to the optimal splitting scheme, thus obtaining the association information. Finally, the association information is returned to the active party. In addition, the active party can also determine the subspace index, split the current tree node, and associate the split current tree node with the target passive party and the local lookup table based on the association information. If the current tree depth has not reached the maximum depth, the left and right subspaces are used as new instance spaces, and the next iteration is performed. If the current tree depth has reached the maximum depth, the current tree training is complete.

[0066] In a semi-honest scenario, the homomorphic encryption combined with obfuscation circuitry in this embodiment can securely receive gradient information from the active party and obtain samples from the passive party, calculate the optimal splitting scheme, and feed it back to the passive party. Simultaneously, it ensures that the gradient values ​​of the active party are not known to the passive party, and that the gradients obtained from different splitting schemes for the passive party's samples are not known to the active party. This stems from the following three points: (1) Employing a semantic security scheme for the transmitted gradient values and Homomorphic encryption is used, so the passive party cannot know the plaintext gradient value; (2) The padded values ​​are statistically indistinguishable from random values. Therefore, after decryption, the statistical characteristics of the padded aggregated gradient are completely indistinguishable from random values. The active party cannot know the specific aggregated gradients and values ​​corresponding to all possible splitting schemes of the passive party; (3) Use a secure obfuscation circuit scheme to calculate the score when neither party knows the other's input, and realize the function of determining the best split scheme.

[0067] In summary, this embodiment addresses the core pain point of vertical federated XGBoost training (gradient and information leakage) by embedding a privacy protection mechanism into its original training process, ultimately achieving the goal of "not changing the collaborative modeling logic of vertical federated XGBoost, not leaking privacy, and not degrading model performance." This is mainly achieved by hiding the aggregated gradient sum using random numbers, blocking the inference path of the active party to the passive party's sample information; and by using obfuscation circuits to achieve privacy-preserving split score calculation, balancing privacy and efficiency. The entire solution has a clear process and complete technical details, making it directly applicable and providing a reliable solution for privacy-preserving training of vertical federated XGBoost. It achieves dual privacy protection for gradient values ​​and aggregated gradient sums. This solution completely solves the privacy risks caused by aggregated gradient sum leakage in existing solutions without changing the original splitting logic of XGBoost or sacrificing model performance, while also considering training efficiency and compatibility. It can be widely applied in fields such as finance, advertising, and healthcare that require multi-party collaborative modeling and are privacy-sensitive.

[0068] Based on the above embodiments, the present invention also provides a system for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums. This system is used to implement the steps in the above method embodiments. Specifically, the system of this embodiment can be referred to... Figure 4As shown, the system may include an active party and multiple passive parties. The active party includes a gradient sharing module, a first locally optimal split point scheme determination module, and a globally optimal split point scheme determination module. The gradient sharing module calculates the gradient of each sample in the instance space of the current iteration during longitudinal federated XGBoost training, and broadcasts the generated gradient ciphertext to all passive parties after homomorphically encrypting the gradient. The first locally optimal split point scheme determination module summarizes the total gradient sum of the instance space, determines the hidden left aggregated gradient sum and the hidden right aggregated gradient sum based on the hidden aggregated gradient ciphertext and the total gradient sum, and uses the obfuscation circuit, the hidden left aggregated gradient sum, and the hidden right aggregated gradient sum, along with encrypted random numbers input by the passive parties, to calculate the split score of all split point schemes and determine the locally optimal split point scheme for all passive parties. The globally optimal split point scheme determination module is used by the active party to determine the globally optimal split point scheme based on all locally optimal split point schemes.

[0069] The passive party includes a gradient privacy hiding module and a second locally optimal split point scheme determination module. The gradient privacy hiding module, upon receiving the gradient ciphertext, iterates through all its own features, determines the split point scheme and aggregated gradient ciphertext for each feature, performs privacy hiding on the aggregated gradient ciphertext using random numbers to obtain hidden aggregated gradient ciphertext, and constructs an obfuscation circuit for calculating the split score. It then sends the split point scheme, hidden gradient ciphertext, and obfuscation circuit to the active party. The second locally optimal split point scheme determination module encrypts the first and second random numbers and inputs them into the obfuscation circuit to calculate the split score for all split point schemes together with the active party.

[0070] In one implementation, the gradient sharing module includes: The gradient calculation unit is used to determine the instance space of the current iteration and calculate the first and second gradients of the loss function for each sample in the instance space. The gradient encryption unit is used to encrypt the first gradient and the second gradient using the Paillier additive homomorphic encryption algorithm to obtain the first gradient ciphertext and the second gradient ciphertext. The gradient ciphertext broadcasting unit is used to broadcast the first gradient ciphertext and the second gradient ciphertext to all passive parties.

[0071] In one implementation, the gradient privacy hiding module includes: The split point scheme determination unit is used to receive the first gradient ciphertext and the second gradient ciphertext sent by the initiator, and traverse all its own features and all split point schemes for each feature. The gradient ciphertext aggregation unit is used to divide the instance space into a left instance space and a right instance space for each split point scheme, and to calculate the first aggregated gradient ciphertext of the first gradient ciphertext and the second aggregated gradient ciphertext of the second gradient ciphertext using the Paillier additive homomorphism property. A random number generation unit is used to generate a first random number corresponding to the first gradient and a second random number corresponding to the second gradient. The privacy hiding unit is used to perform privacy hiding on the first aggregated gradient ciphertext based on the first random number to obtain a first hidden aggregated gradient ciphertext, and to perform privacy hiding on the second aggregated gradient ciphertext based on the second random number to obtain a second hidden aggregated gradient ciphertext.

[0072] In one implementation, the gradient privacy hiding module further includes: The obfuscation circuit construction unit is used to call the obfuscation circuit generation algorithm to construct an obfuscation circuit for calculating the split score. The obfuscation circuit includes: circuit information, encoding information and decoding information. The circuit information sending unit is used to send the split point scheme, the first hidden aggregation gradient ciphertext, the second hidden aggregation gradient ciphertext, the circuit information and encoding information of the obfuscated circuit to the active party.

[0073] In one implementation, the first locally optimal split point scheme determination module includes: The hidden aggregate gradient ciphertext decryption unit is used to decode the first hidden aggregate gradient ciphertext and the second hidden aggregate gradient ciphertext using the Paillier private key to obtain the first hidden left aggregate gradient and the second hidden left aggregate gradient. A hidden aggregate gradient and derivation unit is provided for deriving a first hidden right aggregate gradient and a second hidden right aggregate gradient based on the total gradient sum; The gradient encoding summation unit is used to encode the first hidden left aggregated gradient sum, the second hidden left aggregated gradient, the first hidden right aggregated gradient sum, the second hidden right aggregated gradient sum, and the total gradient sum using the encoding information to obtain encoded gradient summation information, and input the encoded gradient summation information to the obfuscation circuit. The split score calculation unit is used to output the split score of all split point schemes for each passive party based on the confusion circuit. The local optimal split determination unit is used to traverse the split score of all split point schemes for each passive party, and the split point scheme with the highest split score is taken as the local optimal split point scheme for each passive party. The second local optimal split point scheme determination module includes: a random number encryption unit, used to encrypt the first random number and the second random number, and input the first encrypted random number and the second encrypted random number into the obfuscation circuit, and calculate the split score of all split point schemes together with the active party.

[0074] In one implementation, the global optimal split point scheme determination module includes: The split score collection unit is used to collect the local best split schemes and corresponding split scores of all passive players. The global optimal split determination unit is used to determine the locally optimal split scheme with the highest split score as the global optimal split scheme.

[0075] The functions and principles of each module of the system of the present invention are the same as those of each step in the above method embodiments, and will not be repeated here.

[0076] In some embodiments, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums. The display unit of the computer device is used to form a visually visible image. It can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0077] Those skilled in the art will understand that Figure 5The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0078] In some embodiments, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: During the training of vertical federated XGBoost, the active party calculates the gradient of each sample in the instance space of the current iteration, and after homomorphically encrypting the gradient, broadcasts the generated gradient ciphertext to all passive parties. After receiving the gradient ciphertext, the passive party traverses all its own features, determines the split point scheme and aggregated gradient ciphertext for each feature, uses random numbers to perform privacy hiding on the aggregated gradient ciphertext to obtain the hidden aggregated gradient ciphertext, and constructs a confusion circuit for calculating the split score. The passive party then sends the split point scheme, the hidden gradient ciphertext, and the confusion circuit to the active party. The active party summarizes the total gradient sum of the instance space, and based on the hidden aggregate gradient ciphertext and the total gradient sum, determines the hidden left aggregate gradient sum and the hidden right aggregate gradient sum respectively. Then, using the obfuscation circuit, the hidden left aggregate gradient sum and the hidden right aggregate gradient sum, together with the encrypted random number input by the passive party, the split score of all split point schemes is calculated, and the local optimal split point scheme of all passive parties is determined. The active party determines the globally optimal split point scheme based on all locally optimal split point schemes.

[0079] It should be noted that the user information (including but not limited to user device function information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0080] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0081] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0082] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for determining the globally optimal split during longitudinal federated XGBoost training with hidden gradient sums, characterized in that, The method includes: During the training of vertical federated XGBoost, the active party calculates the gradient of each sample in the instance space of the current iteration, and after homomorphically encrypting the gradient, broadcasts the generated gradient ciphertext to all passive parties. After receiving the gradient ciphertext, the passive party traverses all its own features, determines the split point scheme and aggregated gradient ciphertext for each feature, uses random numbers to perform privacy hiding on the aggregated gradient ciphertext to obtain the hidden aggregated gradient ciphertext, and constructs a confusion circuit for calculating the split score. The passive party then sends the split point scheme, the hidden gradient ciphertext, and the confusion circuit to the active party. The active party summarizes the total gradient sum of the instance space, and based on the hidden aggregate gradient ciphertext and the total gradient sum, determines the hidden left aggregate gradient sum and the hidden right aggregate gradient sum respectively. Then, using the obfuscation circuit, the hidden left aggregate gradient sum and the hidden right aggregate gradient sum, together with the encrypted random number input by the passive party, the split score of all split point schemes is calculated, and the local optimal split point scheme of all passive parties is determined. The active party determines the globally optimal split point scheme based on all locally optimal split point schemes.

2. The method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums as described in claim 1, characterized in that, The active side calculates the gradient of the instance space in the current iteration, including: The active party determines the instance space for the current iteration; The active party calculates the first and second gradients of the loss function for each sample in the instance space.

3. The method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums according to claim 2, characterized in that, After homomorphically encrypting the gradient, the generated gradient ciphertext is broadcast to all passive parties, including: The first gradient and the second gradient are encrypted using the Paillier additive homomorphic encryption algorithm to obtain the first gradient ciphertext and the second gradient ciphertext. Broadcast the first gradient ciphertext and the second gradient ciphertext to all passive parties.

4. The method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums as described in claim 3, characterized in that, After receiving the gradient ciphertext, the passive party traverses all its own features to determine the splitting point scheme and aggregated gradient ciphertext for each feature, including: The passive party receives the first and second gradient ciphertexts sent by the active party, and iterates through all its own features and all split point schemes for each feature. For each split point scheme, the instance space is divided into a left instance space and a right instance space. Using the Paillier additive homomorphism property, the first aggregated gradient ciphertext of the first gradient ciphertext and the second aggregated gradient ciphertext of the second gradient ciphertext are calculated.

5. The method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums as described in claim 4, characterized in that, The passive party uses random numbers to perform privacy hiding on the aggregated gradient ciphertext, obtaining hidden aggregated gradient ciphertext, including: Generate a first random number corresponding to the first gradient and a second random number corresponding to the second gradient; Privacy hiding is performed on the first aggregated gradient ciphertext based on the first random number to obtain the first hidden aggregated gradient ciphertext; Privacy hiding is performed on the second aggregated gradient ciphertext based on the second random number to obtain the second hidden aggregated gradient ciphertext.

6. The method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums as described in claim 5, characterized in that, The passive party constructs a confusion circuit for calculating the split score, and sends the split point scheme, hidden gradient ciphertext, and the confusion circuit to the active party, including: A confusion circuit generation algorithm is invoked to construct a confusion circuit for calculating the split score. The confusion circuit includes: circuit information, encoding information, and decoding information. The split point scheme, the first hidden aggregation gradient ciphertext, the second hidden aggregation gradient ciphertext, and the circuit information and encoding information of the obfuscated circuit are sent to the active party.

7. The method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums as described in claim 6, characterized in that, Based on the hidden aggregated gradient ciphertext and the total gradient sum, the active party determines the hidden left aggregated gradient sum and the hidden right aggregated gradient sum, respectively, including: The Paillier private key is used to decrypt the first hidden aggregate gradient ciphertext and the second hidden aggregate gradient ciphertext to obtain the first hidden left aggregate gradient and the second hidden left aggregate gradient. Based on the total gradient sum, the first hidden right aggregated gradient sum and the second hidden right aggregated gradient sum are derived.

8. The method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums according to claim 7, characterized in that, Using the aforementioned obfuscation circuit, hidden left-aggregated gradient sum, and hidden right-aggregated gradient sum, along with the encrypted random numbers input by the passive party, the splitting scores of all splitting point schemes are calculated, and the locally optimal splitting point schemes for all passive parties are determined, including: The passive party encrypts the first random number and the second random number, and inputs the first encrypted random number and the second encrypted random number into the obfuscation circuit; The active party uses the encoding information to encode the first hidden left aggregate gradient sum, the second hidden left aggregate gradient, the first hidden right aggregate gradient sum, the second hidden right aggregate gradient sum, and the total gradient sum to obtain the encoded gradient summary information, and inputs the encoded gradient summary information into the obfuscation circuit; Based on the aforementioned confusion circuit, the splitting score of all splitting point schemes for each passive party is output. The active party iterates through all splitting point schemes of each passive party and calculates the splitting score. The splitting point scheme with the highest splitting score is taken as the local best splitting point scheme for each passive party.

9. The method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums as described in claim 8, characterized in that, The active party determines the globally optimal split point scheme based on all locally optimal split point schemes, including: The active player collects all the local best splitting schemes and corresponding splitting scores from the passive players. The locally optimal splitting scheme with the highest splitting score is taken as the globally optimal splitting scheme.

10. The method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums according to claim 9, characterized in that, The method further includes: The active party determines the target passive party, target features, and optimal split point corresponding to the globally optimal splitting scheme; The target features and the optimal split point are sent to the target passive party so that the target passive party can determine the association information of the globally optimal splitting scheme.

11. The method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums according to claim 10, characterized in that, When determining the association information of the globally optimal splitting scheme, the target passive party includes: The target passive party divides the current instance space into a left subspace and a right subspace based on the target features and the optimal split point; Create a local lookup table to record the node ID and subspace index corresponding to the optimal splitting scheme, and obtain the association information; The associated information is returned to the initiating party.

12. The method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums according to claim 11, characterized in that, The method further includes: Based on the aforementioned association information, the active party determines the subspace index, splits the current tree node, and associates the split current tree node with the target passive party and the local lookup table; If the current tree depth has not reached the maximum depth, the left and right subspaces are used as new instance spaces, and the next iteration is performed. If the current tree depth has reached the maximum depth, then the training of the current tree is complete.

13. A system for determining the globally optimal split during longitudinal federated XGBoost training with hidden gradient sums, characterized in that, The system is used to implement the step of determining the globally optimal split in longitudinal federated XGBoost training of hidden gradient sums as described in any one of claims 1-12, the system comprising: an active party and a passive party; The active party includes: The gradient sharing module is used to calculate the gradient of each sample in the instance space of the current iteration during the training of longitudinal federated XGBoost, and broadcast the generated gradient ciphertext to all passive parties after homomorphically encrypting the gradient. The first local optimal split point scheme determination module is used to summarize the total gradient sum of the instance space, determine the hidden left aggregate gradient sum and the hidden right aggregate gradient sum based on the hidden aggregate gradient ciphertext and the total gradient sum, and use the obfuscation circuit, the hidden left aggregate gradient sum and the hidden right aggregate gradient sum, together with the encrypted random number input by the passive party, to calculate the split score of all split point schemes and determine the local optimal split point scheme of all passive parties. The global optimal split point scheme determination module is used by the active party to determine the global optimal split point scheme based on all local optimal split point schemes. The passive party includes: The gradient privacy hiding module is used to traverse all its own features after receiving the gradient ciphertext, determine the split point scheme and aggregated gradient ciphertext for each feature, use random numbers to perform privacy hiding on the aggregated gradient ciphertext to obtain the hidden aggregated gradient ciphertext, and construct a confusion circuit for calculating the split score. The split point scheme, the hidden gradient ciphertext and the confusion circuit are sent to the active party. The second local optimal split point scheme determination module encrypts the first and second random numbers and inputs them into the confusion circuit to calculate the split score of all split point schemes together with the active party.

14. The system for determining the globally optimal split during longitudinal federated XGBoost training of hidden gradient sums according to claim 13, characterized in that, The gradient sharing module includes: The gradient calculation unit is used to determine the instance space of the current iteration and calculate the first and second gradients of the loss function for each sample in the instance space. The gradient encryption unit is used to encrypt the first gradient and the second gradient using the Paillier additive homomorphic encryption algorithm to obtain the first gradient ciphertext and the second gradient ciphertext. The gradient ciphertext broadcasting unit is used to broadcast the first gradient ciphertext and the second gradient ciphertext to all passive parties.

15. The system for determining the globally optimal split during longitudinal federated XGBoost training of hidden gradient sums according to claim 14, characterized in that, The gradient privacy hiding module includes: The split point scheme determination unit is used to receive the first gradient ciphertext and the second gradient ciphertext sent by the initiator, and traverse all its own features and all split point schemes for each feature. The gradient ciphertext aggregation unit is used to divide the instance space into a left instance space and a right instance space for each split point scheme, and to calculate the first aggregated gradient ciphertext of the first gradient ciphertext and the second aggregated gradient ciphertext of the second gradient ciphertext using the Paillier additive homomorphism property. A random number generation unit is used to generate a first random number corresponding to the first gradient and a second random number corresponding to the second gradient. The privacy hiding unit is used to perform privacy hiding on the first aggregated gradient ciphertext based on the first random number to obtain a first hidden aggregated gradient ciphertext, and to perform privacy hiding on the second aggregated gradient ciphertext based on the second random number to obtain a second hidden aggregated gradient ciphertext.

16. The system for determining the globally optimal split during longitudinal federated XGBoost training of hidden gradient sums according to claim 15, characterized in that, The gradient privacy hiding module also includes: The obfuscation circuit construction unit is used to call the obfuscation circuit generation algorithm to construct an obfuscation circuit for calculating the split score. The obfuscation circuit includes: circuit information, encoding information and decoding information. The circuit information sending unit is used to send the split point scheme, the first hidden aggregation gradient ciphertext, the second hidden aggregation gradient ciphertext, the circuit information and encoding information of the obfuscated circuit to the active party.

17. The system for determining the globally optimal split during longitudinal federated XGBoost training of hidden gradient sums according to claim 16, characterized in that, The first local optimal split point scheme determination module includes: The hidden aggregate gradient ciphertext decryption unit is used to decode the first hidden aggregate gradient ciphertext and the second hidden aggregate gradient ciphertext using the Paillier private key to obtain the first hidden left aggregate gradient and the second hidden left aggregate gradient. A hidden aggregate gradient and derivation unit is provided for deriving a first hidden right aggregate gradient and a second hidden right aggregate gradient based on the total gradient sum; The gradient encoding summation unit is used to encode the first hidden left aggregated gradient sum, the second hidden left aggregated gradient, the first hidden right aggregated gradient sum, the second hidden right aggregated gradient sum, and the total gradient sum using the encoding information to obtain encoded gradient summation information, and input the encoded gradient summation information to the obfuscation circuit. The split score calculation unit is used to output the split score of all split point schemes for each passive party based on the confusion circuit. The local optimal split determination unit is used to traverse the split score of all split point schemes for each passive party, and the split point scheme with the highest split score is taken as the local optimal split point scheme for each passive party. The second local optimal split point scheme determination module includes: a random number encryption unit, used to encrypt the first random number and the second random number, and input the first encrypted random number and the second encrypted random number into the obfuscation circuit, and calculate the split score of all split point schemes together with the active party.

18. The system for determining the globally optimal split in longitudinal federated XGBoost training of hidden gradient sums according to claim 17, characterized in that, The global optimal split point scheme determination module includes: The split score collection unit is used to collect the local best split schemes and corresponding split scores of all passive players. The global optimal split determination unit is used to determine the locally optimal split scheme with the highest split score as the global optimal split scheme.

19. A computer device, characterized in that, The computer device includes a memory, a processor, and a program stored in the memory and executable on the processor for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums. When the processor executes the program for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums, it implements the steps of the method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums as described in any one of claims 1-12.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums, wherein the program for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums implements the steps of the method for determining the globally optimal split in longitudinal federated XGBoost training with hidden gradient sums as described in any one of claims 1-12 on the computer-readable storage medium.

Images