CN-121997370-A - Privacy protection function calculation method based on tropical algebra neural network fitting

CN121997370ACN 121997370 ACN121997370 ACN 121997370ACN-121997370-A

Abstract

The invention discloses a privacy protection function calculation method based on tropical algebra neural network fitting. The method comprises the steps of performing piecewise linear fitting on continuous functions requiring secret calculation by using a tropical algebraic neural network, wherein the tropical algebraic neural network comprises a linear layer, a minimum additive layer and a maximum additive layer which are sequentially formed, taking a fitted tropical algebraic neural network model as a function requiring protection, enabling a model party to hold fitting parameters of the model, compiling a calculation process of the fitted tropical algebraic neural network model into a Boolean circuit formed by logic gates by the model party, performing encryption confusion on the circuit, generating a confusion circuit, sending the confusion circuit to a data party and the like. The method and the device have the advantages that obvious advantages are shown in privacy protection function calculation by introducing tropical algebra neural network fitting and garbled circuit technology, and the core challenges of the current safe multiparty calculation in nonlinear function processing are effectively solved.

Inventors

LUO YE
OUYANG JINZHI
ZHOU LI

Assignees

厦门大学
阿米智能(厦门)科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260108

Claims (10)

1. A privacy protection function calculation method based on tropical algebra neural network fitting, the method comprising: Performing piecewise linear fitting on a continuous function requiring secret calculation by using a tropical algebraic neural network, wherein the tropical algebraic neural network comprises a linear layer, a minimum additive layer and a maximum additive layer which are sequentially formed; Compiling the calculation process of the fitted tropical algebraic neural network model into a Boolean circuit formed by logic gates by a model party, encrypting and confusing the circuit, generating a confusing circuit and transmitting the confusing circuit to a data party; The data party obtains own private input data from the model party The corresponding encrypted tag; the data party locally uses the obtained encryption tag to evaluate the received confusion circuit layer by layer to obtain the encryption tag output by the fitted tropical algebraic neural network model; the data side decrypts locally based on the output encrypted label to obtain a final function calculation result y (x).
2. A privacy preserving function calculating method based on tropical algebraic neural network fitting according to claim 1, The linear layer is used for fitting and outputting a plurality of linear basis functions aiming at given input data; The minimum adding layer is used for carrying out min-plus combination with bias on all linear basis functions; The maximum adding layer is used for carrying out max-plus combination with bias on the linear base function after all min-plus combinations with bias.
3. A privacy preserving function calculating method based on tropical algebraic neural network fitting according to claim 2, The linear layer, given input data First layer output The linear basis functions: in the formula, Is a first fitting parameter; the minimum additive layer comprises Cores, each core performs min-plus combination with bias for all linear basis functions: in the formula, Is a second fitting parameter for controlling the bias of different kernels; The maximum addition layer comprises Each core, each core is all Max-plus combination with bias is performed: in the formula, And a third fitting parameter for controlling the bias of the different kernels.
4. A privacy preserving function calculating method based on tropical algebraic neural network fitting according to claim 1, The data party obtains own private input data from the model party The corresponding encrypted tag, specifically: the data party obtains own private input data from the model party through an careless transmission protocol The corresponding encrypted tag.
5. A privacy preserving function computing device based on tropical algebraic neural network fitting, comprising: The fitting module is used for carrying out piecewise linear fitting on the continuous function requiring secret calculation by using a tropical algebraic neural network, wherein the tropical algebraic neural network comprises a linear layer, a minimum additive layer and a maximum additive layer which are sequentially formed; the compiling and transmitting module is used for compiling the calculation process of the fitted tropical algebraic neural network model into a Boolean circuit formed by logic gates by a model party, encrypting and confusing the circuit, generating a confusing circuit and transmitting the confusing circuit to a data party; a receiving module for the data party to obtain its own private input data from the model party The corresponding encrypted tag; the solving module is used for locally using the obtained encryption tag by the data party, evaluating the received confusion circuit layer by layer to obtain the encryption tag output by the fitted tropical algebraic neural network model; And the decryption module is used for locally decrypting the data party based on the output encryption tag to obtain a final function calculation result y (x).
6. A privacy preserving function calculating apparatus based on tropical algebraic neural network fitting of claim 5, The tropical algebraic neural network comprises a linear layer, a minimum addition layer and a maximum addition layer, wherein, The linear layer is used for fitting and outputting a plurality of linear basis functions aiming at given input data; The minimum adding layer is used for carrying out min-plus combination with bias on all linear basis functions; The maximum adding layer is used for carrying out max-plus combination with bias on the linear base function after all min-plus combinations with bias.
7. A privacy preserving function calculating apparatus based on tropical algebraic neural network fitting of claim 6, The linear layer, given input data First layer output The linear basis functions: in the formula, Is a first fitting parameter; the minimum additive layer comprises Cores, each core performs min-plus combination with bias for all linear basis functions: in the formula, Is a second fitting parameter for controlling the bias of different kernels; The maximum addition layer comprises Each core, each core is all Max-plus combination with bias is performed: in the formula, And a third fitting parameter for controlling the bias of the different kernels.
8. A privacy preserving function calculating apparatus based on tropical algebraic neural network fitting of claim 5, The receiving module is specifically configured to: the data party obtains own private input data from the model party through an careless transmission protocol The corresponding encrypted tag.
9. A privacy preserving function computing device based on tropical algebraic neural network fitting, comprising: At least one processor, and A memory communicatively coupled to the at least one processor, wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 4.
10. A computer readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the method of any one of claims 1 to 4.

Description

Privacy protection function calculation method based on tropical algebra neural network fitting Technical Field The invention relates to the technical field of privacy protection, in particular to a privacy protection function calculation method based on tropical algebra neural network fitting. Background In the current data-driven era, privacy protection has become a non-negligible core problem in various computing tasks. With the rapid development of technologies such as artificial intelligence, big data and cloud computing, when various organizations and organizations cooperatively process sensitive data (such as user behavior information, medical records and financial data), necessary computing tasks are needed to be completed on the premise of not revealing original data content and original computing functions. This demand has motivated extensive attention and intensive research into secure multiparty computing technology. Secure multi-party computing is a type of cryptographic method that allows multiple parties to jointly complete a function computation with their respective inputs kept private, and ensures that the final computation results are correct and secret to the parties. Wherein, the two-party security calculation is a basic form of multiparty security calculation and is applicable to the situation of only two participants. It shows wide prospects in applications such as privacy searching, joint machine learning, and cross-organizational data mining. Multiparty secure computing has evolved into a variety of implementations including secret sharing, homomorphic encryption, zero knowledge proof, and garbled circuits. In practical applications, the infrastructure used to construct secure multiparty computing protocols has a relatively sophisticated and efficient scheme in handling linear or near linear functions (e.g., addition, scalar multiplication, matrix multiplication, etc.). For example, by combining secret sharing with Beaver triplet (Beaver triplet) preprocessing, each party can complete multiplication operation on the premise of only exchanging a small amount of mask difference, and communication and calculation overhead can be effectively controlled. It is with this idea that many multi-party computing frameworks break down linear computations into additions and multiplications, thus achieving efficient linear computations in practice. However, when it is required to support nonlinear functions (e.g., activation functions, piecewise functions, log/exponential/regularization terms, etc.) under privacy protection, the efficiency overhead of existing schemes is greatly increased, which is one of the main bottlenecks in the application of current multiparty security computing in the actual function computing direction. The following problems are particularly acute: 1. High communication and computational complexity Nonlinear functions typically require the use of comparisons, conditional branches, approximate interpolation or higher order polynomial expansion, look-up tables, or garbled circuits. For example, implementing non-linear gates (e.g., reLU, sigmoid) in Garbled Circuit (gar) methods requires more encryption gates, data conversion and transmission, and in a secret sharing framework, additional protocols (e.g., secure comparison protocols) may be required to support. Because of the many interaction steps of these protocols, the high encryption overhead, the need for additional pre-processing or auxiliary servers, the overall efficiency tends to be far lower than that of a linear gate. 2. Poor expansibility and severe customization of scheme For different nonlinear functions, different precision/approximation, different input fields/value ranges, it is almost always necessary to individually design specific protocols or optimization strategies. In other words, there is no general, black box privacy preserving computing framework that can efficiently and flexibly cover private computing of any function or any network structure. Therefore, a great deal of customization work is needed for each function fitting scene, and popularization or engineering realization is difficult. 3. Function privacy is not fully protected The existing mainstream multiparty computing schemes almost use "protection of input of each party" as a core, and less importance is placed on confidentiality of a computing function itself (i.e., a function model, parameters, structures, coefficients, etc.). In many real-world scenarios, the function model itself may contain business secrets, algorithmic policies, trained parameters (e.g., weights in AI model, cost function forms, etc.), potentially exposing core assets if only data privacy is preserved while functional details are revealed. Disclosure of Invention Therefore, the present invention is directed to a method for calculating a privacy protection function based on a tropical algebraic neural network fitting, which can solve at least one technical problem me