CN114996772A - Method and device for federated learning, electronic equipment and storage medium - Google Patents

Method and device for federated learning, electronic equipment and storage medium Download PDF

Info

Publication number
CN114996772A
CN114996772A CN202210693204.2A CN202210693204A CN114996772A CN 114996772 A CN114996772 A CN 114996772A CN 202210693204 A CN202210693204 A CN 202210693204A CN 114996772 A CN114996772 A CN 114996772A
Authority
CN
China
Prior art keywords
gradient
value
random value
noise
random
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210693204.2A
Other languages
Chinese (zh)
Inventor
陈立峰
卞阳
李腾飞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shanghai Fudata Technology Co ltd
Original Assignee
Shanghai Fushu Technology Co ltd Guangzhou Branch
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shanghai Fushu Technology Co ltd Guangzhou Branch filed Critical Shanghai Fushu Technology Co ltd Guangzhou Branch
Priority to CN202210693204.2A priority Critical patent/CN114996772A/en
Publication of CN114996772A publication Critical patent/CN114996772A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F21/70—Protecting specific internal or peripheral components, in which the protection of a component leads to protection of the entire computer
    • G06F21/71—Protecting specific internal or peripheral components, in which the protection of a component leads to protection of the entire computer to assure secure computing or processing of information
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F21/60—Protecting data
    • G06F21/602—Providing cryptographic facilities or services
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Hardware Design (AREA)
  • Mathematical Physics (AREA)
  • Computer Security & Cryptography (AREA)
  • Artificial Intelligence (AREA)
  • Bioethics (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本申请提供一种联邦学习方法、装置、电子设备及存储介质，用于改善目前联邦学习过程的安全性较低的问题。该方法包括：接收无标签方设备发送的密态梯度随机值，无标签方设备上存储有样本数据，密态梯度随机值是对样本数据对应的损失值进行同态加密并加随机数获得的；对密态梯度随机值进行同态解密，获得梯度随机值；在梯度随机值中添加噪声值，获得加噪梯度随机值；向无标签方设备发送加噪梯度随机值，以使无标签方设备根据加噪梯度随机值对本地模型进行联邦学习。

Figure 202210693204

The present application provides a federated learning method, device, electronic device and storage medium, which are used to improve the problem of low security in the current federated learning process. The method includes: receiving a dense gradient random value sent by an unlabeled device, storing sample data on the unlabeled device, and obtaining the dense gradient random value by homomorphically encrypting a loss value corresponding to the sample data and adding a random number. ; Homomorphically decrypt the gradient random value of the dense state to obtain a gradient random value; add a noise value to the gradient random value to obtain a noise-added gradient random value; send the noise-added gradient random value to the unlabeled device so that the unlabeled The device performs federated learning on the local model based on random values of the noised gradient.

Figure 202210693204

Description

Method and device for federated learning, electronic equipment and storage medium
Technical Field
The present application relates to the technical field of machine learning and federal learning, and in particular, to a method and an apparatus for federal learning, an electronic device, and a storage medium.
Background
Federal Machine Learning (FML), also known as federal Learning, joint Learning, or league Learning, is a Machine Learning technique, specifically, people train algorithms on multiple distributed edge devices or servers that own local data samples; this approach is significantly different from traditional centralized machine learning techniques that upload all local data sets onto one server, whereas more classical decentralized approaches typically assume that the local data samples are all equally distributed. Federal machine learning is a machine learning framework, and can effectively help a plurality of organizations to perform data use and machine learning modeling under the condition of meeting the requirements of user privacy protection, data safety and laws and regulations.
In the current federal learning process, equipment on a labeled side and equipment on an unlabeled side can communicate with each other; the non-labeled side equipment is equipment for storing sample data used for training the model, and the labeled side equipment is equipment for storing a sample label corresponding to the sample data. The equipment on the part with the label calculates the loss values of the prediction label and the sample label corresponding to the sample data under the condition that the sample data stored in the plaintext on the equipment without the label is not needed to be obtained, and sends the loss values to the equipment on the part without the label, so that the equipment on the part without the label calculates the gradient value according to the loss values, and the local model is trained by using the gradient value. In specific practice, the current federal learning process is found to be less safe.
Disclosure of Invention
An object of the embodiment of the present application is to provide a federated learning method, an apparatus, an electronic device, and a storage medium, which are used to solve the problem of low security in the current federated learning process.
The embodiment of the application provides a federated learning method, which comprises the following steps: receiving a secret state gradient random value sent by equipment on a non-tag side, wherein sample data is stored in the equipment on the non-tag side, and the secret state gradient random value is obtained by carrying out homomorphic encryption on a loss value corresponding to the sample data and adding a random number; homomorphic decryption is carried out on the secret state gradient random value to obtain a gradient random value; adding a noise value to the gradient random value to obtain a noise-added gradient random value; and sending the noise-added gradient random value to the equipment on the non-tag side so that the equipment on the non-tag side performs federal learning on the local model according to the noise-added gradient random value. In the implementation process, a noise-added gradient random value is obtained by adding a noise value to the obtained gradient random value; and sending the noise-added gradient random value to the equipment on the non-tag side, so that the equipment on the non-tag side is difficult to use the gradient value, the prediction tag and the sample data to break the sample tag stored on the equipment on the tag, and the security of federal learning is effectively improved.
Optionally, in this embodiment of the present application, adding a noise value to the gradient random value includes: a noise value is generated using a local differential privacy algorithm and added to the gradient random value. In the implementation process, the noise value is generated by using the local differential privacy algorithm, and the local differential privacy algorithm only considers the statistical characteristics of the local data set, so that the risk that the local data is leaked to other equipment is better avoided in the process of multi-party federal learning or multi-party safety calculation, and the safety of federal learning is effectively improved.
Optionally, in this embodiment of the application, before receiving the secret gradient random value sent by the non-tag side device, the method further includes: obtaining a dense state loss value, wherein the dense state loss value is calculated by a labeled side device, and the labeled side device is a device for storing a sample label corresponding to sample data; and sending the secret state loss value to the non-tag side equipment so that the non-tag side equipment calculates a secret state gradient value according to the secret state loss value, and adding the generated random number to the secret state gradient value to obtain and return a secret state gradient random value. In the implementation process, the secret state loss value is sent to the non-tag side equipment, so that the non-tag side equipment calculates the secret state gradient value according to the secret state loss value, the generated random number is added to the secret state gradient value, and the secret state gradient random value is obtained and returned, so that the problem that the gradient value is leaked to the non-tag side equipment in a plaintext state is solved, and the security of federal learning is effectively improved.
Optionally, in this embodiment of the present application, obtaining the secret state loss value includes: sending a public key of the central server to the equipment with the label so that the equipment with the label can use the public key to perform homomorphic encryption on the obtained loss value, and obtaining and returning a secret loss value; and receiving the secret state loss value sent by the equipment on the label side.
The embodiment of the application further provides a federated learning method, which is applied to equipment without a tag, and comprises the following steps: obtaining a secret state loss value, and calculating a secret state gradient value according to the secret state loss value and sample data stored on the non-tag side equipment, wherein the secret state loss value is obtained by homomorphic encryption; adding the generated random number to the secret gradient value to obtain a secret gradient random value; sending the secret state gradient random value to a central server so that the central server decrypts the gradient random value from the secret state gradient random value, adding a noise value into the gradient random value, and obtaining and returning the noise-added gradient random value; and receiving a noise-added gradient random value sent by the central server, and subtracting a random number from the noise-added gradient random value to obtain a noise-added gradient value, wherein the noise-added gradient value is used for carrying out federal learning on the local model. In the implementation process, a noise-added gradient random value is obtained by adding a noise value to the obtained gradient random value; and sending the noise-added gradient random value to the equipment on the non-tag side, so that the equipment on the non-tag side is difficult to use the gradient value, the prediction tag and the sample data to break the sample tag stored on the equipment on the tag, and the security of federal learning is effectively improved.
Optionally, in this embodiment of the present application, obtaining the secret state loss value includes: obtaining model parameters and sample data of a local model, and calculating an inner product result between the model parameters and the sample data; sending the inner product result to the equipment of the tagged party to enable the equipment of the tagged party to calculate a prediction label corresponding to the inner product result, calculating a loss value between the prediction label and a sample label corresponding to the inner product result, and then carrying out homomorphic encryption on the loss value by using a public key of the central server to obtain and return a secret state loss value; and receiving the secret state loss value sent by the equipment on the label side. In the implementation process, the inner product result between the model parameter and the sample data is calculated by acquiring the model parameter and the sample data of the local model; and the inner product result is sent to the equipment with the label, so that the problem that the sample data plaintext is directly leaked to the equipment with the label is avoided, and only the inner product result between the model parameter and the sample data is leaked to the equipment with the label, so that the security of federal learning is effectively improved.
The embodiment of the application further provides a bang learning device, which is applied to a central server and comprises: the system comprises a secret state gradient receiving module, a secret state gradient processing module and a random number generating module, wherein the secret state gradient receiving module is used for receiving a secret state gradient random value sent by equipment on a non-label side, sample data is stored in the equipment on the non-label side, and the secret state gradient random value is obtained by homomorphically encrypting a loss value corresponding to the sample data and adding the random number; the secret state gradient decryption module is used for homomorphic decryption of the secret state gradient random value to obtain a gradient random value; the gradient random noise adding module is used for adding a noise value in the gradient random value to obtain a noise-added gradient random value; and the noise gradient sending module is used for sending the noise gradient random value to the equipment on the non-tag side so as to enable the equipment on the non-tag side to carry out federal learning on the local model according to the noise gradient random value.
Optionally, in an embodiment of the present application, the gradient random noise adding module includes: and the differential privacy generation module is used for generating a noise value by using a local differential privacy algorithm and adding the noise value to the gradient random value.
Optionally, in an embodiment of the present application, the federal learning device further includes: the system comprises a dense state loss obtaining module, a dense state loss obtaining module and a data processing module, wherein the dense state loss obtaining module is used for obtaining a dense state loss value, the dense state loss value is calculated by a device on a labeled side, and the device on the labeled side is a device for storing a sample label corresponding to sample data; and the dense-state loss sending module is used for sending a dense-state loss value to the non-tag side equipment so that the non-tag side equipment calculates a dense-state gradient value according to the dense-state loss value, and adds the generated random number to the dense-state gradient value to obtain and return a dense-state gradient random value.
Optionally, in an embodiment of the present application, the secret state loss obtaining module includes: the public key homomorphic encryption module is used for sending the public key of the central server to the equipment with the label so that the equipment with the label uses the public key to homomorphically encrypt the obtained loss value, and the obtained loss value is returned; and the secret state loss receiving module is used for receiving the secret state loss value sent by the equipment with the label.
The embodiment of the application further provides a bang learning device, is applied to no label side equipment, includes: the secret state gradient calculation module is used for acquiring a secret state loss value and calculating a secret state gradient value according to the secret state loss value and sample data stored on the non-tag side equipment, wherein the secret state loss value is obtained by homomorphic encryption; the dense gradient random module is used for adding the generated random number to the dense gradient value to obtain a dense gradient random value; the secret state gradient sending module is used for sending the secret state gradient random value to the central server so that the central server decrypts the gradient random value from the secret state gradient random value, adds a noise value in the gradient random value, and obtains and returns the noise-added gradient random value; and the noise gradient receiving module is used for receiving the noise gradient random value sent by the central server, subtracting the random number from the noise gradient random value to obtain a noise gradient value, and the noise gradient value is used for carrying out federal learning on the local model.
Optionally, in an embodiment of the present application, the dense gradient calculating module includes: the inner product result calculation module is used for acquiring the model parameters and the sample data of the local model and calculating the inner product result between the model parameters and the sample data; the inner product result sending module is used for sending the inner product result to the equipment with the label, so that the equipment with the label calculates a prediction label corresponding to the inner product result, calculates a loss value between the prediction label and a sample label corresponding to the inner product result, and then uses a public key of the central server to homomorphically encrypt the loss value to obtain and return a secret loss value; and the first loss receiving module is used for receiving the secret state loss value sent by the equipment with the label.
An embodiment of the present application further provides an electronic device, including: a processor and a memory, the memory storing processor-executable machine-readable instructions, the machine-readable instructions when executed by the processor performing the method as described above.
Embodiments of the present application further provide a computer-readable storage medium, on which a computer program is stored, and the computer program is executed by a processor to perform the method as described above.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments of the present application will be briefly described below, it should be understood that the following drawings only illustrate some embodiments of the present application and therefore should not be considered as limiting the scope, and for those skilled in the art, other related drawings can be obtained according to the drawings without creative efforts.
Fig. 1 is a schematic flow chart of a federal learning method provided in an embodiment of the present application;
FIG. 2 illustrates an interaction sequence diagram for a three-party execution federated learning approach provided by an embodiment of the present application;
FIG. 3 illustrates an interaction sequence diagram of two parties performing the federated learning method provided by an embodiment of the present application;
fig. 4 is a schematic flow chart of a federal learning method executed by a non-tag device according to an embodiment of the present application;
FIG. 5 is a schematic structural diagram of a federal learning device provided in an embodiment of the present application;
fig. 6 shows a schematic structural diagram of an electronic device provided in an embodiment of the present application.
Detailed Description
The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only some of the embodiments of the present application, and not all of the embodiments. The components of the embodiments of the present application, as generally described and illustrated in the figures herein, could be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the present application, as presented in the figures, is not intended to limit the scope of the claimed embodiments of the application, but is merely representative of selected embodiments of the application. All other embodiments obtained by a person skilled in the art based on the embodiments of the present application without making any creative effort belong to the protection scope of the embodiments of the present application.
Before introducing the federal learning method provided in the embodiments of the present application, some concepts involved in the embodiments of the present application are introduced:
machine Learning (ML), a branch in the field of artificial intelligence to study human Learning behavior; by referring to the scientific or theoretical viewpoints of cognition science, biology, philosophy, statistics, information theory, control theory, computational complexity and the like, the method explores the human cognition rules and the learning process through the basic methods of induction, generalization, specialization, analogy and the like, and establishes various algorithms capable of being automatically improved through experience, so that the computer system has the capability of automatically learning specific knowledge and skills.
Homomorphic encryption (Homomorphic encryption) is a form of encryption that allows one to perform a particular form of algebraic operation on ciphertext to obtain a result that is still encrypted, and to decrypt it to obtain the same result as performing the same operation on plaintext. In other words, this technique allows one to perform operations such as retrieval, comparison, etc. in the encrypted data to yield the correct result without the need to decrypt the data throughout the process.
Differential Privacy (Differential Privacy) is a means of data sharing that can be implemented to share only some of the statistical features that describe a database, without disclosing specific to-individual information. The intuitive idea behind differential privacy is: if the influence caused by randomly modifying one record in the database is small enough, the obtained statistical characteristics can not be used for reversely deducing the content of a single record; this feature may be used to protect privacy. From another perspective, differential privacy can be considered as a constraint for algorithms that disclose statistical characteristics of statistical databases. The constraint requires that private information in the records of the database not be disclosed.
It should be noted that the federal learning method provided in this embodiment of the present application may be executed by an electronic device, where the electronic device includes but is not limited to: a tagged-party device, an untagged-party device, and/or a central server. The tagged side equipment refers to electronic equipment with a sample tag for training a model, and the tagged equipment may or may not have sample data corresponding to the sample tag. The non-label side device is an electronic device with sample data for training a model, but the non-label side device does not have a sample label corresponding to the sample data, so that the machine learning model needs to be trained by combining the sample data on the non-label side device and the sample label on the label side device. Optionally, the non-labeled device and the labeled device may select a central server, where the central server is used to coordinate a process of training the machine learning model together, and ensure that, in the process of training the machine learning, sample data on the non-labeled device is not leaked to the labeled device and the central server, and a sample label on the labeled device is not leaked to the non-labeled device and the central server, that is, neither the sample data nor the sample label can be leaked in the clear text to other devices except the device itself, and the central server does not know the sample data and the sample label in the clear text state.
The above-mentioned tagged side device, untagged side device and central server all refer to a device terminal or server having a function of executing a computer program, and the device terminal includes: a smart phone, a personal computer, a tablet computer, a personal digital assistant, or a mobile internet device, etc. A server refers to a device that provides computing services over a network, such as: x86 server and non-x 86 server, non-x 86 server includes: mainframe, minicomputer, and UNIX server. It is understood that the user corresponding to the above-mentioned tagged-side device and the user corresponding to the non-tagged-side device are both parties of federal learning, and therefore, the above-mentioned tagged-side device and the non-tagged-side device may also be both referred to as participant devices.
Application scenarios applicable to the federal learning method are described below, where the application scenarios include, but are not limited to: during federal learning, the gradient values obtained by the untagged device can be formulated as
Figure BDA0003701166740000081
Wherein,
Figure BDA0003701166740000082
representing the gradient value, y the sample label, h (θ) (2) ) Denotes a prediction tag, x (2) Representing sample data. Because the label-free equipment stores sample data, and an accurate prediction label, namely h (theta) can be obtained after the model is iterated for multiple times (2) ) And x (2) Are known. With the deepening of the training times, an accurate gradient value is finally obtained, and y is a discrete value of 0 or 1, so that the label-free device can easily use the gradient value, the prediction label and sample data to break the sample label stored on the labeled device, the specific breaking idea can simplify the above formula and consider the formula as a Subset Sum problem, and the Subset Sum problem has a quick solution under the condition of small-scale calculation or sparse search domain. Therefore, the sample label stored on the labeled device can be easily broken according to the gradient value, the prediction label and the sample data, and the federal learning safety is low. In such a scenario, the federate learning method can be used to add a noise value to the obtained gradient random value, so that the non-labeled side device is difficult to use the gradient value, the prediction label and the sample data to break the sample label stored on the labeled side device, thereby effectively improving the security of federate learning.
Please refer to fig. 1, which illustrates a flow chart of a federal learning method provided in an embodiment of the present application; in a federal learning scenario executed by three parties (including a labeled party device, an unlabeled party device, and a central server), the federal learning method can be applied to the central server, i.e., the federal learning method can be executed by the central server, and of course, can also be executed by a labeled party device in a federal learning method executed by two parties, which will be described in detail later. The main idea of the federal learning method is to obtain a noise-added gradient random value by adding a noise value to the obtained gradient random value; and sending the noise-added gradient random value to the equipment on the non-tag side, so that the equipment on the non-tag side is difficult to use the gradient value, the prediction tag and the sample data to break the sample tag stored on the equipment on the tag, and the security of federal learning is effectively improved. The federal learning method may specifically include:
step S110: the central server receives a secret state gradient random value sent by the equipment of the non-tag side, sample data is stored in the equipment of the non-tag side, and the secret state gradient random value is obtained by carrying out homomorphic encryption on a loss value corresponding to the sample data and adding a random number.
Please refer to fig. 2, which illustrates an interaction sequence diagram of a three-party execution federal learning method provided in the embodiment of the present application; the embodiment of step S110 described above is, for example: the central server may generate a pair of public key (pk) and private key (secret key) in advance using an asymmetric encryption algorithm, and then send the public key generated on the central server to the untagged device. After receiving a public key pk sent by a central server, a non-tag side device firstly obtains sample data and model parameters which are locally stored, wherein the sample data can be represented as x, and the model parameters can be represented as theta; and then calculating an inner product result between the sample data and the model parameters, and sending the inner product result to the equipment with the label, wherein the inner product result can be represented as u, and then u can be obtained by carrying out inner product calculation on x and theta. After receiving the inner product result sent by the non-tag device, the tagged device calculates a prediction tag corresponding to the inner product result, inputs the inner product result into the machine learning model, and obtains a prediction tag, where the prediction tag may be specifically represented as h (θ), and calculates a loss value between the prediction tag and a sample tag corresponding to the inner product result, where the loss value may be represented as y-h (θ), where y represents the sample tag, and then performs homomorphic encryption on the loss value using a public key of the central server, and obtains a secret state loss value represented as [ L []＝[y-h(θ)]And transmitting the secret state loss value [ L ] to the non-label side equipment]. After the non-tag side equipment receives the secret state loss value sent by the tag side equipment, the secret state gradient is calculated according to the secret state loss value, and the secret state gradient can be expressed as
Figure BDA0003701166740000091
Then, the generated random number is added to the dense gradient value to obtain a dense gradient random value
Figure BDA0003701166740000092
Where R represents the generated random number, and finally, sends the secret gradient random value to the central server
Figure BDA0003701166740000093
Figure BDA0003701166740000094
The central server can receive the secret state gradient random value sent by the equipment without the label
Figure BDA0003701166740000095
After step S110, step S120 is performed: and the central server performs homomorphic decryption on the secret state gradient random value to obtain the gradient random value.
The embodiment of step S120 described above is, for example: central server for secret state gradient random value
Figure BDA0003701166740000096
The homomorphic decryption is carried out, and the obtained gradient random value can be expressed as
Figure BDA0003701166740000097
After step S120, step S130 is performed: and the central server adds a noise value in the gradient random value to obtain a noise-added gradient random value.
The embodiment of step S130 described above is, for example: the central server generates a noise value by using a Local Differential Privacy (LDP) algorithm, and adds the noise value to the gradient random value to obtain a noise-added gradient random value; the local differential privacy algorithm is also referred to as a localized differential privacy algorithm. The local differential privacy algorithm only considers the statistical characteristics of the local data set, so that the risk that local data are leaked to other equipment is better avoided in the multi-party federal learning or multi-party safety calculation process, and the safety of federal learning is effectively improved. The noise value can also be generated by adopting a Global Differential Privacy (GDP) algorithm, and the Global Differential Privacy algorithm considers the statistical characteristics of all data sets, so that the model training and convergence process can be accelerated more quickly in the multi-party federal learning or multi-party safety calculation process. Of course, in a specific practical process, the above-mentioned noise value may also use laplacian (Laplace) noise, or may use CH noise.
After step S130, step S140 is performed: and the central server sends the noise-added gradient random value to the equipment on the non-tag side, so that the equipment on the non-tag side performs federal learning on the local model according to the noise-added gradient random value.
The embodiment of step S140 described above is, for example: the central server sends the noise-added gradient random value to the non-tag side device through a Transmission Control Protocol (TCP) or a User Datagram Protocol (UDP). After receiving the noise-added gradient random value sent by the central server, the non-tag side equipment can subtract the random number from the noise-added gradient random value to obtain a noise-added gradient value, and then the noise-added gradient value is used for updating the weight value of the local model; local models herein include, but are not limited to: a logistic regression algorithm model in a machine learning algorithm. In the above process, only one round of data interaction process is described, and in specific practice, the above steps S110 to S140 need to be repeated until the loss function of the local model converges, for example: and when the loss value of the local model is smaller than a preset proportion or the number of iteration times (epoch) is larger than a preset threshold value, the trained local model can be obtained. The preset proportion can be set according to specific situations, for example, set to be 5% or 10%, etc.; the preset threshold may also be set according to specific situations, for example, set to 100 or 1000, etc.
In the implementation process, firstly, the secret state gradient random value sent by the non-tag side equipment is received, then homomorphic decryption is carried out on the secret state gradient random value to obtain a gradient random value, then a noise value is added into the gradient random value, and finally the noisy gradient random value is sent to the non-tag side equipment, so that the non-tag side equipment is difficult to crack the sample tag according to the received gradient random value, and the received gradient random value is a numerical value after noise is added. That is, by adding a noise value to the obtained gradient random value, a noise-added gradient random value is obtained; and sending the noise-added gradient random value to the equipment on the non-tag side, so that the equipment on the non-tag side is difficult to use the gradient value, the prediction tag and the sample data to break the sample tag stored on the equipment on the tag, and the security of federal learning is effectively improved.
Please refer to fig. 3, which illustrates an interaction sequence diagram of two parties executing the federal learning method according to an embodiment of the present application; the federal learning method may specifically include:
step S210: and the equipment on the label side receives the secret state gradient random value sent by the equipment on the non-label side, sample data is stored on the equipment on the non-label side, and the secret state gradient random value is obtained by homomorphically encrypting the loss value corresponding to the sample data and adding a random number.
The embodiment of step S210 described above is, for example: and the equipment on the non-tag side acquires the sample data and the model parameters, calculates the inner product result between the sample data and the model parameters, and then sends the inner product result to the equipment on the tag side. After receiving the inner product result sent by the equipment without the label, the equipment with the label first calculates the prediction label corresponding to the inner product result, calculates the loss value between the prediction label and the sample label corresponding to the inner product result, then performs homomorphic encryption on the loss value to obtain a secret state loss value, and finally sends the secret state loss value to the equipment without the label. After receiving the dense-state loss value sent by the equipment with the label, the equipment without the label calculates a dense-state gradient according to the received dense-state loss value, adds a random number generated at random to the dense-state gradient to obtain a dense-state gradient random value, and then sends the dense-state gradient random value to the equipment with the label. The device with the tag can receive the secret state gradient random value sent by the device without the tag.
After step S210, step S220 is performed: and the equipment on the part with the label performs homomorphic decryption on the secret state gradient random value to obtain the gradient random value.
After step S220, step S230 is performed: and the equipment on the label side adds a noise value in the gradient random value to obtain a noise-added gradient random value.
After step S230, step S240 is performed: and the equipment on the label side sends the noise-added gradient random value to the equipment on the non-label side, so that the equipment on the non-label side performs federal learning on the local model according to the noise-added gradient random value.
The implementation principle and implementation manner of the above steps S220 to S240 are similar to those of the steps S120 to S140, and therefore, the implementation principle and implementation manner will not be described herein, and if it is unclear, reference may be made to the description of the steps S120 to S140.
Please refer to fig. 4, which is a schematic flow chart of a federal learning method executed by a non-tag device according to an embodiment of the present application; in a specific practical process, the above non-tag side device may further calculate a dense state loss value according to a dense state prediction tag and a dense state sample tag corresponding to the sample data, instead of calculating the dense state loss value according to an inner product result, and the implementation may specifically include:
step S310: and the non-tag side equipment acquires a dense state loss value, and a dense state gradient value is calculated according to the dense state loss value and sample data stored on the non-tag side equipment, wherein the dense state loss value is calculated according to a dense state prediction tag and a dense state sample tag on the tag side equipment.
The embodiment of the step S310 is, for example: and the equipment of the non-tag party acquires the sample data, homomorphically encrypts the sample data by using the public key of the central server to obtain the sample data of the secret state, and then sends the sample data of the secret state to the equipment of the tag party. The method comprises the steps that the equipment on the label side calculates a dense state prediction label corresponding to dense state sample data, calculates a dense state loss value between the dense state prediction label and the dense state sample label calculated by the equipment on the label side, and then sends the dense state loss value to the equipment on the non-label side. And the non-tag side equipment receives the secret state loss value sent by the tag side equipment.
Step S320: and the non-tag side equipment adds the generated random number to the secret gradient value to obtain a secret gradient random value, and sends the secret gradient random value to the central server.
The embodiment of step S320 is, for example: after the non-tag side equipment receives the secret state loss value sent by the tag side equipment, the secret state gradient is calculated according to the secret state loss value, and the secret state gradient can be expressed as
Figure BDA0003701166740000131
Then, the generated random number is added to the dense gradient value to obtain a dense gradient random value
Figure BDA0003701166740000132
Where R represents the generated random number, and finally, sends the secret gradient random value to the central server
Figure BDA0003701166740000133
Step S330: the central server receives the secret state gradient random value sent by the non-label side equipment, homomorphic decryption is carried out on the secret state gradient random value to obtain a gradient random value, then, a noise value is added to the gradient random value to obtain a noise-added gradient random value, and the noise-added gradient random value is sent to the non-label side equipment.
Optionally, before receiving the secret gradient random value sent by the non-tag device, the central server may also send a secret loss value to the non-tag device, so that the non-tag device calculates the secret gradient random value according to the secret loss value, where the implementation may include:
step S331: the central server obtains a dense state loss value, wherein the dense state loss value is calculated by the equipment on the part with the label, and the equipment on the part with the label is equipment for storing a sample label corresponding to the sample data.
The embodiment of step S331 described above is, for example: the central server sends the public key of the central server to the equipment with the label through a Hyper Text Transfer Protocol (HTTP) or a Hyper Text Transfer Protocol Secure (HTTPS). After receiving the public key sent by the central server through the HTTP protocol or the HTTPS protocol, the tagged party equipment uses the public key to perform homomorphic encryption on the obtained loss value to obtain a secret state loss value.
Step S332: and the central server sends the secret state loss value to the non-tag side equipment so that the non-tag side equipment calculates a secret state gradient value according to the secret state loss value, and adds the generated random number to the secret state gradient value to obtain and return a secret state gradient random value.
The embodiment of the step S332 is, for example: and the central server sends the secret state loss value to the non-label side equipment through an HTTP (hyper text transport protocol) or an HTTPS (hyper text transport protocol). After receiving the secret state loss value sent by the central server, the non-tag side equipment calculates a secret state gradient value according to the secret state loss value, adds the generated random number to the secret state gradient value to obtain a secret state gradient random value, and finally sends the secret state gradient random value to the central server.
Step S340: and the equipment on the non-tag side receives the noise-added gradient random value sent by the central server, subtracts the random number from the noise-added gradient random value to obtain a noise-added gradient value, and trains a local model by using the noise-added gradient value.
The embodiment of step S340 is, for example: after receiving the noise-added gradient random value sent by the central server, the equipment on the non-tag side can subtract the random number from the noise-added gradient random value to obtain a noise-added gradient value, then train the local model by using the noise-added gradient value, and specifically update the weight value of the local model by using the noise-added gradient value; local models herein include, but are not limited to: a logistic regression algorithm model in a machine learning algorithm. In the above process, only a round of data interaction process is described, and in specific practice, the above steps need to be repeated until the loss function of the local model converges, specifically for example: and when the loss value of the local model is smaller than a preset proportion or the number of iteration times (epoch) is larger than a preset threshold value, the trained local model can be obtained. The preset proportion can be set according to specific situations, for example, set to be 5% or 10%, etc.; the preset threshold may also be set according to specific situations, for example, set to 100 or 1000, etc.
Please refer to fig. 5, which illustrates a schematic structural diagram of a federal learning device provided in an embodiment of the present application; the embodiment of the present application provides a bang learning device 400, is applied to central server, includes:
and a secret state gradient receiving module 410, configured to receive a secret state gradient random value sent by a non-tag side device, where sample data is stored in the non-tag side device, and the secret state gradient random value is obtained by performing homomorphic encryption on a loss value corresponding to the sample data and adding a random number.
And the secret gradient decryption module 420 is configured to perform homomorphic decryption on the secret gradient random value to obtain a gradient random value.
And a gradient random adding module 430, configured to add a noise value to the gradient random value to obtain a noise-added gradient random value.
And a noise gradient sending module 440, configured to send the noise gradient random value to the non-tag side device, so that the non-tag side device performs federal learning on the local model according to the noise gradient random value.
Optionally, in an embodiment of the present application, the gradient random noise adding module includes:
and the differential privacy generation module is used for generating a noise value by using a local differential privacy algorithm and adding the noise value into the gradient random value.
Optionally, in an embodiment of the present application, the federal learning device further includes:
and the dense state loss acquisition module is used for acquiring a dense state loss value, wherein the dense state loss value is calculated by the equipment on the part with the label, and the equipment on the part with the label is equipment for storing a sample label corresponding to the sample data.
And the secret state loss sending module is used for sending the secret state loss value to the non-tag side equipment so that the non-tag side equipment calculates a secret state gradient value according to the secret state loss value, and adds the generated random number to the secret state gradient value to obtain and return the secret state gradient random value.
Optionally, in an embodiment of the present application, the secret state loss obtaining module includes:
and the public key homomorphic encryption module is used for sending the public key of the central server to the equipment with the label so that the equipment with the label uses the public key to homomorphically encrypt the obtained loss value, and the obtained loss value is returned.
And the secret state loss receiving module is used for receiving the secret state loss value sent by the equipment with the label.
The embodiment of the application provides a bang learning device, is applied to no label side equipment, includes:
and the secret state gradient calculation module is used for acquiring a secret state loss value and calculating a secret state gradient value according to the secret state loss value and sample data stored on the non-tag side equipment, wherein the secret state loss value is obtained by homomorphic encryption.
And the dense gradient random module is used for adding the generated random number to the dense gradient value to obtain a dense gradient random value.
And the secret state gradient sending module is used for sending the secret state gradient random value to the central server so that the central server decrypts the gradient random value from the secret state gradient random value, adds a noise value in the gradient random value, and obtains and returns the noisy gradient random value.
And the noise gradient receiving module is used for receiving the noise gradient random value sent by the central server, subtracting the random number from the noise gradient random value to obtain a noise gradient value, and the noise gradient value is used for carrying out federal learning on the local model.
Optionally, in an embodiment of the present application, the dense gradient calculating module includes:
and the inner product result calculation module is used for acquiring the model parameters and the sample data of the local model and calculating the inner product result between the model parameters and the sample data.
And the inner product result sending module is used for sending the inner product result to the equipment with the label, so that the equipment with the label calculates a prediction label corresponding to the inner product result, calculates a loss value between the prediction label and a sample label corresponding to the inner product result, and then homomorphically encrypts the loss value by using a public key of the central server to obtain and return a secret state loss value.
And the first loss receiving module is used for receiving the secret state loss value sent by the equipment with the label.
It should be understood that the apparatus corresponds to the above-mentioned federal learning method embodiment, and can perform the steps related to the above-mentioned method embodiment, and the specific functions of the apparatus can be referred to the above description, and the detailed description is appropriately omitted here to avoid redundancy. The device includes at least one software function that can be stored in memory in the form of software or firmware (firmware) or solidified in the Operating System (OS) of the device.
Please refer to fig. 6 for a schematic structural diagram of an electronic device according to an embodiment of the present application. An electronic device 500 provided in an embodiment of the present application includes: a processor 510 and a memory 520, the memory 520 storing machine readable instructions executable by the processor 510, the machine readable instructions when executed by the processor 510 performing the method as above.
Embodiments also provide a computer-readable storage medium 530, where the computer-readable storage medium 530 stores thereon a computer program, and when the computer program is executed by the processor 510, the computer program performs the method described above.
The computer-readable storage medium 530 may be implemented by any type of volatile or nonvolatile storage device or combination thereof, such as a Static Random Access Memory (SRAM), an Electrically Erasable Programmable Read-Only Memory (EEPROM), an Erasable Programmable Read-Only Memory (EPROM), a Programmable Read-Only Memory (PROM), a Read-Only Memory (ROM), a magnetic Memory, a flash Memory, a magnetic disk, or an optical disk.
In the embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. The apparatus embodiments described above are merely illustrative, and for example, the flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved.
In addition, functional modules of the embodiments in the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part. Furthermore, in the description of the present specification, reference to the description of "one embodiment," "some embodiments," "an example," "a specific example," or "some examples" or the like means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the embodiments of the present application. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.
In this document, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions.
The above description is only an alternative embodiment of the embodiments of the present application, but the scope of the embodiments of the present application is not limited thereto, and any person skilled in the art can easily conceive of changes or substitutions within the technical scope of the embodiments of the present application, and all the changes or substitutions should be covered by the scope of the embodiments of the present application.

Claims (10)

1.一种联邦学习方法，其特征在于，包括：1. A federated learning method, comprising: 接收无标签方设备发送的密态梯度随机值，所述无标签方设备上存储有样本数据，所述密态梯度随机值是对所述样本数据对应的损失值进行同态加密并加随机数获得的；Receive the random value of the density gradient sent by the device of the unlabeled side, where the sample data is stored on the device of the unlabeled side, and the random value of the density gradient is to perform homomorphic encryption on the loss value corresponding to the sample data and add a random number acquired; 对所述密态梯度随机值进行同态解密，获得梯度随机值；performing homomorphic decryption on the dense gradient random value to obtain a gradient random value; 在所述梯度随机值中添加噪声值，获得加噪梯度随机值；adding a noise value to the gradient random value to obtain a noise-added gradient random value; 向无标签方设备发送所述加噪梯度随机值，以使所述无标签方设备根据所述加噪梯度随机值对本地模型进行联邦学习。The random value of the noised gradient is sent to the device on the unlabeled side, so that the device on the unlabeled side performs federated learning on the local model according to the random value of the added noised gradient. 2.根据权利要求1所述的方法，其特征在于，所述在所述梯度随机值中添加噪声值，包括：2. The method according to claim 1, wherein the adding a noise value to the gradient random value comprises: 使用局部差分隐私算法生成所述噪声值，并在所述梯度随机值中添加所述噪声值。The noise value is generated using a local differential privacy algorithm and added to the gradient random value. 3.根据权利要求1-2任一所述的方法，其特征在于，在所述接收无标签方设备发送的密态梯度随机值之前，还包括：3. The method according to any one of claims 1-2, wherein before the receiving the random value of the density gradient sent by the tagless device, the method further comprises: 获取密态损失值，所述密态损失值是有标签方设备计算出来的，所述有标签方设备是存储有所述样本数据对应的样本标签的设备；Obtaining a dense state loss value, the dense state loss value is calculated by a device with a label, and the device with a label is a device that stores the sample label corresponding to the sample data; 向所述无标签方设备发送所述密态损失值，以使所述无标签方设备根据所述密态损失值计算密态梯度值，并将所述密态梯度值加上生成的随机数，获得并返回所述密态梯度随机值。Send the secret state loss value to the unlabeled device, so that the unlabeled device calculates the secret state gradient value according to the secret state loss value, and adds the secret state gradient value to the generated random number , obtain and return the random value of the dense state gradient. 4.根据权利要求3所述的方法，其特征在于，所述获取密态损失值，包括：4. The method according to claim 3, wherein the obtaining the dense state loss value comprises: 向所述有标签方设备发送公钥，以使所述有标签方设备使用所述公钥对获得的损失值进行同态加密，获得并返回所述密态损失值；Sending a public key to the device with the label, so that the device with the label uses the public key to perform homomorphic encryption on the obtained loss value, and obtain and return the encrypted loss value; 接收所述有标签方设备发送的所述密态损失值。Receive the secrecy loss value sent by the labeled device. 5.一种联邦学习方法，其特征在于，应用于无标签方设备，包括：5. A federated learning method, characterized in that, applied to an unlabeled party device, comprising: 获取密态损失值，根据所述密态损失值和所述无标签方设备上存储的样本数据计算出密态梯度值，所述密态损失值是同态加密获得的；Obtain a dense state loss value, and calculate a dense state gradient value according to the dense state loss value and the sample data stored on the unlabeled device, and the dense state loss value is obtained by homomorphic encryption; 将所述密态梯度值加上生成的随机数，获得密态梯度随机值；Adding the generated random number to the dense state gradient value to obtain the dense state gradient random value; 向中心服务器发送所述密态梯度随机值，以使所述中心服务器从所述密态梯度随机值使用解密出梯度随机值，并在所述梯度随机值中添加噪声值，获得并返回加噪梯度随机值；Send the secret gradient random value to the central server, so that the central server decrypts the gradient random value from the secret gradient random value, adds a noise value to the gradient random value, and obtains and returns the added noise gradient random value; 接收所述中心服务器发送的所述加噪梯度随机值，并将所述加噪梯度随机值减去所述随机数，获得加噪梯度值，所述加噪梯度值用于对本地模型进行联邦学习。Receive the noise-added gradient random value sent by the central server, and subtract the random number from the noise-added gradient random value to obtain a noise-added gradient value, and the noise-added gradient value is used to federate the local model study. 6.根据权利要求5所述的方法，其特征在于，所述获取密态损失值，包括：6. The method according to claim 5, wherein the obtaining the dense state loss value comprises: 获取所述本地模型的模型参数和所述样本数据，并计算所述模型参数和所述样本数据之间的内积结果；acquiring model parameters of the local model and the sample data, and calculating an inner product result between the model parameters and the sample data; 向有标签方设备发送所述内积结果，以使所述有标签方设备计算所述内积结果对应的预测标签，并计算所述预测标签与所述内积结果对应的样本标签之间的损失值，然后，使用所述中心服务器的公钥对所述损失值进行同态加密，获得并返回密态损失值；Send the inner product result to the labeled side device, so that the labeled side device calculates the predicted label corresponding to the inner product result, and calculates the difference between the predicted label and the sample label corresponding to the inner product result. Loss value, then use the public key of the central server to perform homomorphic encryption on the loss value to obtain and return the encrypted loss value; 接收所述有标签方设备发送的所述密态损失值。Receive the secrecy loss value sent by the labeled device. 7.一种联邦学习装置，其特征在于，应用于中心服务器，包括：7. A federated learning device, characterized in that, applied to a central server, comprising: 密态梯度接收模块，用于接收无标签方设备发送的密态梯度随机值，所述无标签方设备上存储有样本数据，所述密态梯度随机值是对所述样本数据对应的损失值进行同态加密并加随机数获得的；The dense gradient receiving module is used to receive the random value of the gradient of the dense state sent by the device of the unlabeled party. The device of the unlabeled party stores sample data, and the random value of the dense state of the gradient is the loss value corresponding to the sample data. Obtained by homomorphic encryption and adding random numbers; 密态梯度解密模块，用于对所述密态梯度随机值进行同态解密，获得梯度随机值；A dense state gradient decryption module, used for performing homomorphic decryption on the dense state gradient random value to obtain a gradient random value; 梯度随机加噪模块，用于在所述梯度随机值中添加噪声值，获得加噪梯度随机值；Gradient random noise adding module, used for adding noise value to the gradient random value to obtain the noise adding gradient random value; 加噪梯度发送模块，用于向无标签方设备发送所述加噪梯度随机值，以使所述无标签方设备根据所述加噪梯度随机值对本地模型进行联邦学习。The noise-added gradient sending module is configured to send the random value of the noise-added gradient to the device on the unlabeled side, so that the device on the un-labeled side can perform federated learning on the local model according to the random value of the noise-added gradient. 8.根据权利要求7所述的装置，其特征在于，所述梯度随机加噪模块，包括：8. The apparatus according to claim 7, wherein the gradient random noise addition module comprises: 差分隐私生成模块，用于使用局部差分隐私算法生成所述噪声值，并在所述梯度随机值中添加所述噪声值。A differential privacy generating module, configured to use a local differential privacy algorithm to generate the noise value, and add the noise value to the gradient random value. 9.一种电子设备，其特征在于，包括：处理器和存储器，所述存储器存储有所述处理器可执行的机器可读指令，所述机器可读指令被所述处理器执行时执行如权利要求1至6任一项所述的方法。9. An electronic device, comprising: a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and when the machine-readable instructions are executed by the processor, the instructions are executed as follows: The method of any one of claims 1 to 6. 10.一种计算机可读存储介质，其特征在于，该计算机可读存储介质上存储有计算机程序，该计算机程序被处理器运行时执行如权利要求1至6任一项所述的方法。10. A computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the method according to any one of claims 1 to 6 when the computer program is run by a processor.
CN202210693204.2A 2022-06-17 2022-06-17 Method and device for federated learning, electronic equipment and storage medium Pending CN114996772A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210693204.2A CN114996772A (en) 2022-06-17 2022-06-17 Method and device for federated learning, electronic equipment and storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210693204.2A CN114996772A (en) 2022-06-17 2022-06-17 Method and device for federated learning, electronic equipment and storage medium

Publications (1)

Publication Number Publication Date
CN114996772A true CN114996772A (en) 2022-09-02

Family

ID=83034935

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210693204.2A Pending CN114996772A (en) 2022-06-17 2022-06-17 Method and device for federated learning, electronic equipment and storage medium

Country Status (1)

Country Link
CN (1) CN114996772A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115809473A (en) * 2023-02-02 2023-03-17 富算科技(上海)有限公司 Method and device for acquiring information value of longitudinal federal learning
CN115829051A (en) * 2022-11-17 2023-03-21 北京八分量信息科技有限公司 Longitudinal federal logistic regression algorithm

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130170640A1 (en) * 2011-04-29 2013-07-04 International Business Machines Corporation Fully Homomorphic Encryption
CN111507481A (en) * 2020-04-17 2020-08-07 腾讯科技（深圳）有限公司 A federated learning system
WO2021073234A1 (en) * 2019-10-16 2021-04-22 支付宝(杭州)信息技术有限公司 Method and device for jointly training logistic regression model by multiple computing units
CN113065046A (en) * 2021-04-28 2021-07-02 北京理工大学 Product defect detection equipment and method
CN114282692A (en) * 2022-03-08 2022-04-05 富算科技(上海)有限公司 Model training method and system for longitudinal federal learning

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130170640A1 (en) * 2011-04-29 2013-07-04 International Business Machines Corporation Fully Homomorphic Encryption
WO2021073234A1 (en) * 2019-10-16 2021-04-22 支付宝(杭州)信息技术有限公司 Method and device for jointly training logistic regression model by multiple computing units
CN111507481A (en) * 2020-04-17 2020-08-07 腾讯科技（深圳）有限公司 A federated learning system
CN113065046A (en) * 2021-04-28 2021-07-02 北京理工大学 Product defect detection equipment and method
CN114282692A (en) * 2022-03-08 2022-04-05 富算科技(上海)有限公司 Model training method and system for longitudinal federal learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
RAFIK H ET AL: "《Towards Secure Big Data Analysis via Fully Homomorphic Encryption Algorithms》", 《MDPI》, 6 April 2022 (2022-04-06) *
黄立棋: "《基于宽度学习的联邦算法模型的设计与实现》", 《中国优秀硕士学位论文全文数据库信息科技辑》, 15 March 2022 (2022-03-15) *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115829051A (en) * 2022-11-17 2023-03-21 北京八分量信息科技有限公司 Longitudinal federal logistic regression algorithm
CN115809473A (en) * 2023-02-02 2023-03-17 富算科技(上海)有限公司 Method and device for acquiring information value of longitudinal federal learning

Similar Documents

Publication Publication Date Title
TWI733106B (en) Model-based prediction method and device
CN111368319B (en) Block chain-based data security access method in federated learning environment
US11449753B2 (en) Method for collaborative learning of an artificial neural network without disclosing training data
US11176469B2 (en) Model training methods, apparatuses, and systems
CN114595835B (en) Model training method and device based on federal learning, equipment and storage medium
CN113542228B (en) Federated learning-based data transmission method, device and readable storage medium
CN110138802A (en) User's characteristic information acquisition methods, device, block chain node and storage medium
CN113221105A (en) Robustness federated learning algorithm based on partial parameter aggregation
CN114386069B (en) A federated learning model training method based on conditional privacy set intersection
CN112818369B (en) Combined modeling method and device
US11410081B2 (en) Machine learning with differently masked data in secure multi-party computing
CN111210003A (en) Vertical federated learning system optimization method, device, device and readable storage medium
CN115841133A (en) Method, device and equipment for federated learning and storage medium
CN115409198B (en) Distributed prediction method and system
US12348611B2 (en) Method for processing multi-source data
JP6915375B2 (en) Polynomial-based homomorphic encryption
CN112989399A (en) Data processing system and method
Das et al. A secure softwarized blockchain-based federated health alliance for next generation IoT networks
CN114462626A (en) Federal model training method and device, terminal equipment and storage medium
CN107196918B (en) Data matching method and device
WO2022141014A1 (en) Security averaging method based on multi-user data
CN115865323A (en) Pearson correlation coefficient calculation method based on secret sharing and OT protocol
CN114996772A (en) Method and device for federated learning, electronic equipment and storage medium
CN117521102A (en) Model training method and device based on federal learning
CN115130121A (en) Method for training longitudinal logistic regression model under privacy calculation of third party

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
TA01 Transfer of patent application right

Effective date of registration: 20230814

Address after: Room 416, 4th Floor, Building 1, No. 655 Yinxiang Road, Jiading District, Shanghai, 200000

Applicant after: SHANGHAI FUDATA TECHNOLOGY Co.,Ltd.

Address before: 510640 room 741, Zone C, 2nd floor, No. 38, Shengtang street, Cencun, Tianhe District, Guangzhou, Guangdong Province

Applicant before: Shanghai Fushu Technology Co.,Ltd. Guangzhou Branch

TA01 Transfer of patent application right