Alipay (hangzhou) Information Technology Co., Ltd.

Advanced New Technologies Co., Ltd.

Alibaba Group Holding Limited

Ant Blockchain Technology (shanghai) Co., Ltd.

Advantageous New Technologies Co., Ltd.

Chaochao Chen

Li Wang

Jun Zhou

Longfei Zheng

Peng Li

Yingting Liu

Xiaolong Li

Jinming Cui

Ziqi Liu

Yinggui Wang

Yilun Lin

Hongjun Yin

One or more embodiments of the present specification provide privacy protection-based multicollinearity detection methods, apparatuses, and systems. Data alignment is performed by a member device on respective local feature data with other member devices to construct a joint feature matrix. Privacy protection-based multi-party matrix multiplication computation is performed to compute a product matrix of a transposed matrix of the joint feature matrix and the joint feature matrix. An inverse matrix of the product matrix is determined based on respective submatrices of the product matrix. A variance inflation factor of each attribute feature is determined by the member device with the other member devices using respective submatrices of the inverse matrix and the respective local feature data. Multicollinearity is determined by the member device with the other member devices based on fragment data of the variance inflation factor of each attribute feature.

Privacy protection-based multicollinearity detection methods, apparatuses, and systems

A client device determines a local user gradient value based on a current user preference vector and a local item gradient value based on a current item feature vector. The client device updates a user preference vector by using the local user gradient value and updates an item feature vector by using the local item gradient value. The client device determines a neighboring client device based on a predetermined adjacency relationship. The local item gradient value is sent by the client device to the neighboring client device. The client device receives a neighboring item gradient value sent by the neighboring client device. The client device updates the item feature vector by using the neighboring item gradient value. In response to the client device determining that a predetermined iteration stop condition is satisfied, the client device outputs the user preference vector and the item feature vector.

Recommendation system construction method and apparatus

A computer-implemented method, medium, and system are disclosed. One example method includes determining multiple model bases by multiple service parties. A respective local service model is constructed by each service party. Respective local training samples are processed by each service party using the respective local service model to determine respective gradient data corresponding to each model basis. The respective gradient data is sent to a server. In response to determining that the first model basis satisfies a gradient update condition, corresponding gradient data of the first model basis received from each service party are combined to obtain global gradient data corresponding to the first model basis. The global gradient data is sent to each service party. Reference parameters in local model basis corresponding to the first model basis are updated by each service party using the global gradient data to train the respective local service model.

Methods, apparatuses, and systems for updating service model based on privacy protection

A computer-implemented method, medium, and system are disclosed. One example computer-implemented method performed by a server includes obtaining training task information from a task party. The training task information includes information about a to-be-pretrained model and information about a to-be-trained target model. A respective task acceptance indication from each of at least one of a plurality of data parties is received to obtain a candidate data party set. The information about the to-be-pretrained model is sent to each data party in the candidate data party set. A respective pre-trained model of each data party is received. A respective performance parameter of the respective pre-trained model of each data party is obtained. One or more target data parties from the candidate data party set is determined. The information about the to-be-trained target model is sent to the one or more target data parties to obtain a target model.

Target data party selection methods and systems for distributed model training

Disclosed herein are methods, systems, and apparatus, including computer programs encoded on computer storage media for secure collaborative computation of a matrix product of a first matrix including private data of a first party and a second matrix including private data of the second party by secret sharing without a trusted initializer. One method includes obtaining a first matrix including private data of the first party; generating a first random matrix; identifying a first sub-matrix and a second sub-matrix of the first random matrix; computing first scrambled private data of the first party based on the first matrix, the first random matrix, the first sub-matrix, and the second sub-matrix; receiving second scrambled private data of the second party; computing a first addend of the matrix product; receiving a second addend of the matrix product; and computing the matrix product by summing the first addend and the second addend.

Secure multi-party computation with no trusted initializer

A method for training a neural network model includes: at each first member device, obtaining predicted label data according to a first neural network submodel of the first neural network submodel by using private data for model training, and determining model update information of the first neural network submodel according to the predicted label data and real label data; providing, by each first member device, the model update information of the first neural network submodel and local sample distribution information to a second member device; at the second member device, performing neural network model reconstruction, determining an overall sample probability distribution, and allocating a reconstructed neural network model and the overall sample probability distribution to each first member device; and updating the first neural network submodel at each first member device according to a local sample probability distribution, the reconstructed neural network model, and the overall sample probability distribution.

Method, apparatus, and system for training neural network model

Embodiments of the present specification provide a method and an apparatus for clustering privacy data of a plurality of parties. The method includes: a first party determining a first data part of central data currently corresponding to a cluster of K clusters, where the first data part corresponds to a first dimension set; a second party has a second data part of the central data corresponding to a second dimension set; separately using central data of a cluster of the K clusters as target central data, and obtaining a first segment of a target distance between a sample of the N samples and the target central data through local calculation based on a first feature part of any one of N samples and the first data part of the target central data; determining a shortest target distance among target distances by performing joint comparison using the first segment of each target distance and a second segment of each target distance of the second party through secret sharing; and determining a cluster corresponding to the shortest target distance as a cluster that any sample currently belongs to. As such, leakage of privacy data can be prevented.

Method and apparatus for clustering privacy data of plurality of parties

One or more embodiments of the present specification relate to a method and system for obtaining a jointly trained model based on privacy protection. An example method includes jointly training a first model, by a first device and with a second device, the first device and the second device each holding respective training data that includes first training data with a sample label, and second training data without the sample label. Jointly training the first model includes privately obtaining the first training data, and performing joint model training using the first training data. The second training data is input to the trained first model to obtain a predicted label for the second training data. Jointly training a second model includes privately obtaining labeled training data, and performing joint model training using the first training data, the second training data, and the labeled training data.

Obtaining jointly trained model based on privacy protection

Methods, systems, and apparatus for facial recognition. An example method includes storing encrypted facial features of a target user that are encrypted by using a first public key; receiving a recognition request from an end-user device of the target user, wherein the recognition request comprises an encrypted facial image and a second public key; performing homomorphic feature processing on the encrypted facial image; obtain obfuscated and encrypted output features; obtain obfuscated and encrypted facial features; transmitting the obfuscated and encrypted output features and the obfuscated and encrypted facial features to the end-user device; receiving from the end-user device a difference between the second intermediate value and the first intermediate value; and determining whether the to-be-recognized facial image corresponds to a facial image of the target user, comprising removing an impact of the first obfuscation and the second obfuscation on the difference.

Facial recognition methods and apparatuses

A first training participant performs an iterative process until a predetermined condition is satisfied, where the iterative process includes: obtaining, using secret sharing matrix addition and based on the current sub-model of each training participant and a corresponding feature sample subset of each training participant, a current prediction value of the regression model for a feature sample set, where the corresponding feature sample subset of each training participant is obtained by performing vertical segmentation on the feature sample set; determining a prediction difference between the current prediction value and a label corresponding to the current prediction value; sending the prediction difference to each second training participant; and updating a current sub-model of the first training participant based on the current sub-model of the first training participant and a product of a corresponding feature sample subset of the first training participant and the prediction difference.

Model training methods, apparatuses, and systems

Disclosed are a model training method and apparatus based on gradient boosting decision tree (GBDT). A GBDT algorithm flow is divided into two stages. In the first stage, labeled samples are obtained from a data domain of a service scenario similar to a target service scenario to sequentially train several decision trees, and training residual generated after the training in the first stage is determined; in the second stage, labeled samples are obtained from a data domain of the target service scenario, and several decision trees continue to be trained based on the training residual. Finally, a model applied to the target service scenario is actually obtained by integrating the decision trees trained in the first stage with the decision trees trained in the second stage.

Model training method and apparatus based on gradient boosting decision tree

Implementations of the present specification disclose graph structure model training and junk account identification methods, apparatuses, and devices. The solution includes: obtaining an account medium network graph, a node in the account medium network graph representing an account and a medium, and at least some edges indicating that a login behavior relationship exists between nodes connected by the edges; obtaining feature data and risk labeling data of the node, the feature data reflecting a login behavior of the corresponding node in a time sequence; and training, based on the account medium network graph, the feature data, and the risk labeling data, a predefined graph structure model for identifying a junk account.

Graph structure model training and junk account identification

An item rating and recommendation platform identifies rating data including respective ratings of multiple items with respect to multiple users; identifies user-feature data including user features contributing to the respective ratings of the multiple items with respect to the multiple users; and receives, from a social network platform via a secret sharing scheme without a trusted initializer, manipulated social network data computed based on social network data and a first number of random variables. The social network data indicate social relationships between any two of the number of users. In the secret sharing scheme without the trust initializer, the social network platform shares with the item rating and recommendation platform manipulated social network data without disclosing the social network data. The item rating and recommendation platform updates the user-feature data based on the rating data and the manipulated social network data.

Secret sharing with no trusted initializer

An item rating and recommendation platform identifies rating data comprising respective ratings of multiple items with respect to multiple users, identifies user-feature data comprising multiple user features contributing to the respective ratings of the multiple items with respect to the multiple users, and receives, from a social network platform via a secret sharing scheme with a trusted initializer, manipulated social network data computed based on social network data and first input data from the trusted initializer. The social network data indicate social relationships between any two of the multiple users. In the secret sharing scheme with the trusted initializer, the social network platform shares with the item rating and recommendation platform the manipulated social network data without disclosing the social network data. The item rating and recommendation platform updates the user-feature data based on the rating data and the manipulated social network data.

Growing community of inventors

Hangzhou, China

Chaochao Chen