FATE federal privacy set intersection method and system based on vector oblivious valuation
By using the VOLE protocol and OPRF protocol to identify the same elements in the privacy set, the problem of low computational efficiency and insufficient security of the privacy set interception method in FATE federated learning is solved, and the malicious scenarios are realized Efficient privacy requests.
Patent Information
- Application Number
- CN202411031500.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-12
AI Technical Summary
The existing FATE federated learning method of privacy collection interception is inefficient in processing large-scale data and cannot resist malicious scenarios and cannot be applied to malicious scenarios.
Using the FATE federal privacy set interception method based on vector inadvertent valuation, the VOLE protocol is generated by initialization and expansion, combined with the OPRF protocol of pseudo-random functions, the same elements in the set held by the participants are identified to realize the privacy set interception .
It improves computing efficiency, reduces the amount of public key encryption, can be effectively applied in malicious security scenarios, and has higher security.
Smart Images

Figure CN118944872A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a FATE federal privacy set intersection method and system, and in particular to a FATE federal privacy set intersection method and system based on vector oblivious valuation, belonging to the technical field of privacy set intersection. Background Art
[0002] In the prior art, the FATE federated learning framework includes: (1) the data provider Host, which is usually a cooperative organization responsible for providing data to replicate the Guest to complete modeling and help improve the training effect; (2) the data application Guest, which is usually the participant who has business needs to apply data in the actual modeling scenario. In vertical federated learning, the Guest is usually the party with label y; (3) the third-party facilitator Arbiter, which assists multiple parties in completing the joint modeling task, does not provide data, and is mainly responsible for issuing public keys, encryption and decryption, and aggregating models, etc. Figure 2 The FATE framework is divided into four layers from top to bottom, as follows: (1) Application layer: FATEBoard is the FATE visualization interface, which provides users with visualization and measurement of the entire process of model training, allowing users to easily view the key indicators in the model training process; (2) Interconnection layer: FATE Flow is the FATE task scheduler, which connects the database DB, communication module OSX and algorithm container layer container. It is at the core of the framework and is responsible for parsing DSL dependencies and running related job tasks, processing scheduling tasks such as creation and stop, and issuing instructions to various participants; (3) Algorithm container layer: FederatedML is the algorithm local running environment, including the implementation of many common machine learning algorithms and necessary tools to ensure the efficient execution of the algorithm; (4) Platform resource layer: The FATE framework integrates many high-performance computing frameworks and file management components, including Eggroll, a high-performance computing framework in the field of machine learning, Spark, an open source big data processing engine, and HDFS, a distributed file system. The above components together constitute the computing and storage resources of the FATE platform, providing strong support for various machine learning tasks.
[0003] The specific implementation schemes of the privacy intersection technology in the traditional FATE federated learning framework mainly include the following two: (1) Privacy set intersection based on Diffie-Hellman, refer to Figure 3The Diffie-Hellman-based privacy set intersection utilizes the Diffie-Hellman key exchange principle. The first participant P0 and the second participant P1 hold the first privacy intersection set X and the second privacy intersection set Y respectively. Both parties agree on the hash function Hash and select the first random key k0 and the second random key k1 respectively. The specific steps are as follows: In the first step, the first participant P0 applies the hash function Hash to each element of the first privacy intersection set X and calculates the k0 power of the hash result to obtain the first set Send it to the second participant P1; In the second step, the second participant P1 also applies the hash function Hash to each element of the second privacy intersection set Y, and calculates the k1th power of the hash result. At the same time, the second participant P1 calculates the first set sent by the first participant P0 The second set obtained by raising each element to the power of k1 is and the third set Sent to the first participant P0; Step 3, the first participant P0 calculates the second set sent by the second participant P1 The k0 power of each element gives the fourth set And look for the fourth set With the third collection Finally, the first participant P0 informs the first participant P1 of the result of the intersection; (2) Based on the blind signature RSA, the privacy set intersection is obtained, refer to Figure 4 , the first participant P0 and the second participant P1 hold the first privacy intersection set X and the second privacy intersection set Y respectively, and both parties jointly determine the first hash function H1 and the second hash function H2. The second participant P1 selects the public key (n RSA ,e RSA ) and private key (n RSA ,d RSA ), and the public key (n RSA ,e RSA ) is sent to the first participant P0. The specific steps are as follows: Step 1: For each element x in the first private intersection set X, i , the first participant P0 selects a random mask r i Calculate and send the obtained fourth set X' to the second participant P1; Step 2, the second participant P1 uses the private key (n RSA ,d RSA ) Encrypt each element in the fourth set X' to obtain the encrypted set SignedX, because each element in it is randomly masked r iAt the same time, the second participant P1 calculates the fifth set Y", and sends the fifth set Y" together with the encrypted set SignedX to the first participant P0; in the third step, the second participant P1 removes the random mask r in the encrypted set SignedX. i , the sixth set X'' is obtained, and the two parties compare the sixth set X'' and the fifth set Y'' to obtain the common elements of both sets; the main disadvantage of the privacy set intersection based on Diffie-Hellman and the privacy set intersection based on blind signature RSA is that a large amount of public key encryption calculations are used, resulting in a large amount of calculation and a long running time when processing large-scale data sets. In addition, the above two privacy intersection schemes cannot resist malicious adversaries and cannot be applied to malicious scenarios.
[0004] In summary, in order to address the computational efficiency and security issues of the private intersection implementation method in the above-mentioned FATE, a private set intersection method with high security and low computational cost is needed to complete the private intersection task. Summary of the invention
[0005] A brief overview of the present invention is provided below in order to provide a basic understanding of certain aspects of the present invention. It should be understood that this overview is not an exhaustive overview of the present invention. It is not intended to identify key or important parts of the present invention, nor is it intended to limit the scope of the present invention. Its purpose is merely to present certain concepts in a simplified form as a prelude to a more detailed description discussed later.
[0006] In view of this, in order to solve the problem that the traditional privacy set intersection method in the prior art cannot be applied to malicious scenarios and is inefficient, the present invention provides a FATE federal privacy set intersection method and system based on vector oblivious valuation.
[0007] Technical solution 1 is as follows: a FATE federated privacy set intersection method based on vector oblivious valuation, including the following steps:
[0008] S1. Based on the noisy parity learning problem, VOLE tuples are generated by initialization and expansion to obtain the VOLE protocol;
[0009] S2. The first participant and the second participant run the VOLE protocol to obtain vectors and scalars, and construct the OPRF protocol in combination with the estimated key of the pseudo-random function;
[0010] S3. Based on the OPRF protocol, identify the same elements in the sets held by the first participant and the second participant, and achieve the intersection of private sets.
[0011] Furthermore, in S1, the initialization Setup stage selects a subset S with a potential of t on the input domain [n] and a random vector y with a length of t, S = {s1, s1, ..., s t}, For the group The first combination vector (a, b) is randomly selected from the set of vectors of length t. Number Field The set of vectors with length k above, a is the first vector, b is the second vector, calculate the third vector c, c = ax + b, x is the data, perform multi-point function secret sharing MPFSS and seed generation Gen operations on the first key K0 and the second key K1, (K0, K1) = MPFSS.Gen (1 λ ,f S,xy ), where (K0, K1) is the second combination vector, 1 λ is the safety parameter, f S,xy is a multi-point function, for each element s in the set S i , the function is in s i The value at x·y i , for any element z not in the set S, the multipoint function f S,xy The value at z is 0, generating the first seed seed0 and the second seed seed1. The number field defined for the above addition and multiplication operations,
[0012] In the Expand phase, when the index σ=0, the first seed seed0 is decompressed. The calculation is done by the spread function spread n Define a sparse vector μ of length n, μ←spread n (S,y), sparse vector μ represents the multi-point function f S,y The range vector spread n (S,y), that is, for each element s in the set S i , the sparse vector μ in s i The value of y i For any element z that is not in the set S, the value of the sparse vector μ at z is 0, and the function evaluation FullEval is performed on the first key K0, v0 = MPFSS.FullEval (0, K0), where v0 is the multi-point function f S,xy The range vector spread n The first secret sharing of (S,xy) outputs the second combination vector in, is the first input vector, is the second input vector, C k,n is a random matrix known to both parties, Number Field The set of all two-dimensional matrices with k rows and n columns above, when σ = 1, decompress the second seed Perform function evaluation FullEval on the first key K0, v1 = MPFSS.FullEval (1, K1), v1 is the multi-point function f S,xy The range vector spread n The second secret sharing of (S,xy) satisfies spread n (S,xy)=v0+v1, get the output vector
[0013] Furthermore, in S2, the first participant P0 and the second participant P1 run the VOLE protocol, the first participant P0 obtains the fourth vector A and the sixth vector C, the second participant P1 obtains the fifth vector B and the scalar Δ, and C=A·Δ+B is satisfied, and the fourth vector A, the fifth vector B and the sixth vector C all belong to Number Field The set of vectors of length m on which the scalar Δ belongs to the number field m is the vector length of the VOLE protocol output determined by the first participant P0 and the second participant P1;
[0014] The first participant P0 and the second participant P1 jointly select a random matrix M * , The first participant P0 defines a matrix M, M∈{0,1} n×m , the i-th row of the matrix M is equal to the random matrix M * The xth i The first participant P0 solves the equation system M·P T =(0,0,…,0) T , T is transposed, and the seventh vector P is obtained, The first participant P0 sends the fourth vector A and the seventh vector P to the second participant P1, and the second participant P1 sets an estimation key K of a pseudo-random function;
[0015] The valuation key K is expressed as:
[0016] K=B+Δ(A+P)
[0017] By evaluating the key K, we get MK T =M·B T +Δ(M·A T +M·PT )=MB T +ΔMA T =MC T , set the pseudo-random function Among them, H is a random oracle, Represents the matrix M * The xth row corresponds to the OPRF protocol, where the first participant P0 obtains the estimate of the data x under the pseudo-random function. The second participant P1 uses the valuation key K to calculate the valuation F of any input random vector y under the random vector function K (y).
[0018] Furthermore, in S3, for each element in the first privacy intersection set X of the first party P0, the first party P0 and the second party P1 execute the OPRF protocol so that the first party P0 obtains the valuation result The second participant P1 obtains the second participant's valuation key k i , using the second party's estimated key k i The elements in the second privacy intersection set Y of the second participant P1 are evaluated, and all evaluation results are sent to the first participant P0. The evaluation results are compared to identify the same elements in the first privacy intersection set X and the second privacy intersection set Y through the first participant P0.
[0019] Technical Solution 2: A FATE federal privacy set intersection system based on vector oblivious valuation, which is used to implement the FATE federal privacy set intersection method based on vector oblivious valuation described in Technical Solution 1, including a framework layer, a component layer, an interface layer and a function layer;
[0020] The framework layer includes the Fate Board front-end interface and the FaderatedML task scheduling component;
[0021] The component layer includes a privacy set intersection host component, a privacy set intersection guest component and a database, and the privacy set intersection host component and the privacy set intersection guest component are connected to the corresponding database;
[0022] The interface layer is used to implement cross-language function calls and data interface conversion between C++ and Python;
[0023] The function layer includes a sender and a receiver of a privacy request protocol VOLE-PSI connected via a transmission control protocol TCP;
[0024] The FaderatedML task scheduling component is connected to the privacy set intersection host component and the privacy set intersection guest component respectively. The privacy set intersection host component and the privacy set intersection guest component are both connected to the interface layer. The interface layer is connected to the sender and the receiver respectively.
[0025] The beneficial effects of the present invention are as follows: the present invention obtains the VOLE protocol through initialization and expansion based on the noisy parity learning problem, runs the VOLE protocol on the first participant and the second participant, sets the valuation key of the pseudo-random function to construct the OPRF protocol, identifies the same elements in the sets held by the first participant and the second participant based on the OPRF protocol, that is, implements the public key encryption operation of privacy set intersection based on vector inadvertent valuation; and compared with the privacy set intersection method based on Diffie-Hellman key exchange and blind signature RSA, which makes extensive use of public key encryption calculations, the present invention has less calculation amount, shorter calculation time and higher calculation efficiency; the present invention can be extended to privacy set intersection under malicious security, while the privacy set intersection method based on Diffie-Hellman key exchange and blind signature RSA can only be used in semi-honest scenarios, and the present invention has stronger security. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:
[0027] Figure 1 This is a flowchart of the FATE federated privacy set intersection method based on vector oblivious valuation;
[0028] Figure 2 This is a schematic diagram of the FATE federated learning framework;
[0029] Figure 3 This is a flowchart of the privacy set intersection method based on Diffie-Hellman;
[0030] Figure 4 This is a flowchart of the privacy set intersection method based on blind signature RSA;
[0031] Figure 5 This is a structural diagram of the FATE federated private set intersection system based on vector oblivious valuation;
[0032] Figure 6 Schematic diagram of the components of the FATE federated private set intersection system based on vector oblivious estimation;
[0033] Figure 7This is a schematic diagram of the principle of function secret sharing;
[0034] Figure 8 This is a flowchart for implementing the intersection of private sets based on OPRF;
[0035] Fig. 9 This is the code diagram of the upload_guest.json file;
[0036] Fig.10 This is the code diagram of the test_intersect_job_dsl.json file;
[0037] Fig.11 This is the first part of the test_intersect_job_vole_conf.json file;
[0038] Fig.12 This is the second part of the test_intersect_job_vole_conf.json file;
[0039] Fig.13 This is a diagram of the third part of the test_intersect_job_vole_conf.json file. DETAILED DESCRIPTION
[0040] In order to make the technical solutions and advantages of the embodiments of the present invention more clearly understood, the exemplary embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than an exhaustive list of all the embodiments. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.
[0041] Example 1: Reference Figure 1-13 The present embodiment is described in detail, and the FATE federated privacy set intersection method based on vector oblivious valuation specifically includes the following steps:
[0042] S1. Based on the noisy parity learning problem, VOLE tuples are generated by initialization and expansion to obtain the VOLE protocol;
[0043] S2. The first participant and the second participant run the VOLE protocol to obtain vectors and scalars, and construct the OPRF protocol in combination with the estimated key of the pseudo-random function;
[0044] S3. Based on the OPRF protocol, identify the same elements in the sets held by the first participant and the second participant, and achieve the intersection of private sets.
[0045] Furthermore, in S1, the initialization Setup stage selects a subset S with a potential of t on the input domain [n] and a random vector y with a length of t, S = {s1, s1, ..., s t}, For the group The first combination vector (a, b) is randomly selected from the set of vectors of length t. Number Field The set of vectors with length k above, a is the first vector, b is the second vector, calculate the third vector c, c = ax + b, x is the data, perform multi-point function secret sharing MPFSS and seed generation Gen operations on the first key K0 and the second key K1, (K0, K1) = MPFSS.Gen (1 λ ,f S,xy ), where (K0, K1) is the second combination vector, 1 λ is the safety parameter, f S,xy is a multi-point function, for each element s in the set S i , the function is in s i The value at x·y i , for any element z not in the set S, the multipoint function f S,xy The value at z is 0, generating the first seed seed0 and the second seed seed1. The number field defined for the above addition and multiplication operations,
[0046] In the Expand phase, when the index σ=0, the first seed seed0 is decompressed. The calculation is done by the spread function spread n Define a sparse vector μ of length n, μ←spread n (S,y), sparse vector μ represents the multi-point function f S,y The range vector spread n (S,y), that is, for each element s in the set S i , the sparse vector μ in s i The value of y i For any element z that is not in the set S, the value of the sparse vector μ at z is 0, and the function evaluation FullEval is performed on the first key K0, v0 = MPFSS.FullEval (0, K0), where v0 is the multi-point function f S,xy The range vector spread n The first secret sharing of (S,xy) outputs the second combination vector in, is the first input vector, is the second input vector, C k,n is a random matrix known to both parties, Number Field The set of all two-dimensional matrices with k rows and n columns above, when σ = 1, decompress the second seed Perform function evaluation FullEval on the first key K0, v1 = MPFSS.FullEval (1, K1), v1 is the multi-point function f S,xy The range vector spread n The second secret sharing of (S,xy) satisfies spread n (S,xy)=v0+v1, get the output vector
[0047] Furthermore, in S2, the first participant P0 and the second participant P1 run the VOLE protocol, the first participant P0 obtains the fourth vector A and the sixth vector C, the second participant P1 obtains the fifth vector B and the scalar Δ, and C=A·Δ+B is satisfied, and the fourth vector A, the fifth vector B and the sixth vector C all belong to Number Field The set of vectors of length m on which the scalar Δ belongs to the number field m is the vector length of the VOLE protocol output determined by the first participant P0 and the second participant P1;
[0048] The first participant P0 and the second participant P1 jointly select a random matrix M * , The first participant P0 defines a matrix M, M∈{0,1} n×m , the i-th row of the matrix M is equal to the random matrix M * The xth i The first participant P0 solves the equation system M·P T =(0,0,…,0) T , T is transposed, and the seventh vector P is obtained, The first participant P0 sends the fourth vector A and the seventh vector P to the second participant P1, and the second participant P1 sets an estimation key K of a pseudo-random function;
[0049] The valuation key K is expressed as:
[0050] K=B+Δ(A+P)
[0051] By evaluating the key K, we get MK T =M·B T +Δ(M·A T +M·PT )=MB T +ΔMA T =MC T , set the pseudo-random function Among them, H is a random oracle, Represents the matrix M * The xth row corresponds to the OPRF protocol, where the first participant P0 obtains the estimate of the data x under the pseudo-random function. The second participant P1 uses the valuation key K to calculate the valuation F of any input random vector y under the random vector function K Iy).
[0052] Furthermore, in S3, for each element in the first privacy intersection set X of the first party P0, the first party P0 and the second party P1 execute the OPRF protocol so that the first party P0 obtains the valuation result The second participant P1 obtains the second participant's valuation key k i , using the second party's estimated key k i The elements in the second privacy intersection set Y of the second participant P1 are evaluated, and all evaluation results are sent to the first participant P0. The evaluation results are compared to identify the same elements in the first privacy intersection set X and the second privacy intersection set Y through the first participant P0;
[0053] Specifically, the privacy set intersection algorithm based on Vector-Oblivious Linear Evaluation (VOLE) consists of three parts, namely, the generation of VOLE tuples, the construction of OPRF from VOLE tuples, and the construction of PSI from OPRF. VOLE tuple is a secure two-party computation primitive, which refers to the first input vector of equal length input by the first participant P0. and the second input vector The second participant P1 inputs data x. After both parties execute the VOLE protocol, the second participant P1 obtains the output vector The VOLE protocol is usually constructed based on the problem of learning with parity noise (LPN) and multi-point function secret sharing (MPFSS).
[0054] The noisy parity learning problem is expressed as: randomly generate the first matrix A, Number Field A set of two-dimensional matrices with k rows and l columns, randomly generating a secret row vector s, Number Field A vector set of length k is used to randomly generate an error vector e. Number Field The set of vectors of length l above satisfies HW(e)=r·l, where HW is the Hamming weight, r is the noise ratio, and l is the length of vector e. The vector b is calculated, b=s·A+e, then the matrix-vector pair (A,b) is computationally indistinguishable from the randomly generated matrix-vector pair (A',b').
[0055] refer to Figure 7 , the truth function f is split into the first calculation key key0 and the second calculation key key1, which are held by the first calculation party f0 and the first calculation party f1 respectively, so that each calculation party can calculate a secret share of the function result using the input data x and the calculation key. Its advantage is that the function calculation process can be performed locally by each participant. A (nt) multi-point function secret sharing means that the truth function f is a discrete function with a domain width of n, and there are t points in the domain with non-zero values. Multi-point function secret sharing mainly consists of three parts: seed generation Gen, function valuation Eval and result reconstruction Rec. The result reconstruction Rec generally has addition on the domain.
[0056] Example 2: Reference Figure 1-Figure 13 This embodiment is described in detail, the FATE federated privacy set intersection system based on vector oblivious valuation is used to implement the FATE federated privacy set intersection method based on vector oblivious valuation described in Example 1, including a framework layer, a component layer, an interface layer and a function layer;
[0057] The framework layer includes the Fate Board front-end interface and the FaderatedML task scheduling component;
[0058] The component layer includes a privacy set intersection host component, a privacy set intersection guest component and a database, and the privacy set intersection host component and the privacy set intersection guest component are connected to the corresponding database;
[0059] The interface layer is used to implement cross-language function calls and data interface conversion between C++ and Python;
[0060] The function layer includes a sender and a receiver of a privacy request protocol VOLE-PSI connected via a transmission control protocol TCP;
[0061] The FaderatedML task scheduling component is connected to the privacy set intersection host component and the privacy set intersection guest component respectively, the privacy set intersection host component and the privacy set intersection guest component are both connected to the interface layer, and the interface layer is connected to the sender and the receiver respectively;
[0062] Specifically, in this embodiment, the host and the guest in the privacy set intersection application scenario are two data providers with equal relationship, and jointly calculate the privacy set intersection. Figure 5 The system operation workflow is represented as follows: the user publishes a privacy set intersection task, which is passed to the privacy set intersection host component and the privacy set intersection guest component in the component layer through the component coordination module FaderatedML in FATE. The two parties interact and initialize the initial parameters of the privacy set intersection. The data structure of the privacy set is converted into a form that can be recognized by the interface layer function. Then, the C++ function interface defined in the interface layer is called, and the privacy set intersection is completed using the privacy intersection algorithm efficiently implemented by the C++ function interface. The privacy intersection result is returned layer by layer, and in the process of returning, it is gradually reconstructed into Table type data that FATE can recognize through the interface layer and the component layer, and finally it is visualized on the FATE Board front-end interface;
[0063] refer to Figure 6 , the privacy intersection component class diagram corresponds to Figure 5The component layer part describes the integration and calling relationship between the privacy set intersection components in FATE. The white box diagram is the existing Python class for coordinating privacy set intersection in FATE, and the gray box diagram is the privacy set intersection Host class and privacy set intersection Guest class implemented by the present invention based on the inheritance relationship of the privacy set intersection components in FATE. The upper-level task scheduling module interacts with IntersectModelBase, the Python class for coordinating privacy set intersection in FATE, to select a specific privacy set intersection implementation. The privacy set intersection components in FATE all implement the Intersect interface, which is the privacy set intersection algorithm interface in FATE. Therefore, there can be an IntersectModelBase. Called in the same way, VoleIntersectHost is the private set intersection Host-side Python class implemented by the present invention, and VoleIntersectGuest is the private set intersection Guest-side Python class implemented by the present invention. Both implement the run_intersect function in the Intersect interface, and thus can be called by IntersectModelBase. ModelBase is the base class of the component class in the FATE framework, VoleIntersect is the VOLE-based private set intersection algorithm implemented by the present invention in FATE, and Interface is the private set intersection algorithm interface class specified by the FATE framework;
[0064] C++-Python function interface implementation correspondence Figure 5 The interface layer of VOLE-PSI is mainly used to encapsulate the efficient privacy set intersection privacy protocol VOLE-PSI implemented in C++ into a library function that can be called in Python. The above process uses the cross-language binding library pybind11 between C++ and Python. The cross-language binding library pybind11 can automatically connect most APIs such as C++ and Python in a safe and concise form, and will only cause less performance loss. The specific implementation steps are as follows: First, the privacy set intersection initialization function init_psi and the privacy set intersection execution function run_psi in the privacy intersection protocol VOLE-PSI are integrated into one function. Only one call is required to complete the privacy set intersection calculation. Secondly, the PYBIND11_MODULE macro provided by the cross-language binding library pybind11 is used to bind the integrated function. The PYBIND11_MODULE macro can automatically implement the vector data type vector in C++ language. <u64>The vector type numpy provided by the Python language scientific computing library numpy <u64>Finally, use the C++ language cross-platform compilation tool CMake to compile the encapsulated functions into dynamic library files, and copy the dynamic library files to the Python language environment path in FATE, so that FATE components can call the dynamic library;
[0065] The specific steps for the Host to perform the private set intersection operation are as follows: First, after receiving the private set, send the size of its own private set to the Guest to perform the private set intersection operation; second, accept the size of the Guest's private set to perform the private set intersection operation, and convert the private set in the Table format of the FATE data structure into numpy <u64>format, calling the dynamic library to perform the intersection of private sets, and converting numpy <64> The intersection result of the format is converted into the Table format in FATE. Finally, the proportion of the intersection elements in the original set is counted, and the intersection of the private sets is returned.
[0066] The specific steps for the guest to perform the private set intersection operation are as follows: First, receive the private set and accept the size of the private set of the host party who is intersecting the private set. Second, send the size of the private set to the host party who is intersecting the private set, and convert the private set in the table format of the FATE data structure into numpy. <u64>format, calling the dynamic library to perform the intersection of private sets, and converting numpy <64> The intersection result of the format is converted into the Table format in FATE. Finally, the proportion of the intersection elements in the original set is counted; the intersection of the private sets is returned;
[0067] Test the FATE federated privacy set intersection system based on vector oblivious valuation, refer to Fig. 9 , describe the upload_guest.json file in detail, prepare the data set required for the test, and recommend the .csv format. The privacy collection content is written in the id column. The present invention uses Python language to write functions for Host and Guest, and randomly generates a test data set of size 100,000, which are saved in intersect_guest_100000_1.csv and intersect_host_100000_1.csv respectively. Then, write the upload data set configuration file, and use flow to upload the data set using intersect_guest_100000_1.csv as an example;
[0068] refer to Fig.10 , describe the test_intersect_job_dsl.json file in detail, which specifies the components that need to be called by the task, the order of component combination and the data transmission relationship between components. The task will call three modules, namely the file reading Reader module, the data conversion DataTransform module and the privacy intersection Intersect module, namely reader_0, data_transform_0 and intersect_0. The input instruction of each module specifies the input of the module, and the output instruction specifies the output of the module. The file also specifies that reader_0 outputs data "data" as the data input of data_transform_0, data_transform_0 outputs data "data" and model "model", among which "data" is the input of intersect_0 and intersect_0 outputs data "data";
[0069] refer to Figure 11-13 , the test_intersect_job_vole_conf.json file specifies the running parameters for each component during the test. Fig.11 The first line indicates that the DSL version used is version 2, and then specifies that there is only one guest party, whose party_id is 9999, and only one host party, whose party_id is 10000, and the task is submitted by the host party; Fig.12 The "component_parameters" section defines the runtime parameters of three components. The host-side participant No. 0, whose reader_0 component reads the metadata intersect_host_100000_1 from the FATE namespace experiment and passes it to data_transform_0. Then data_transform_0 constructs the metadata into a dense table dense and passes it to intersect_0. Intersect_0 uses the vole privacy intersection scheme, which is the scheme implemented by the present invention. Then "vole_params" defines the runtime parameters, including the random seed salt is "12345", the number of threads threat_num is 1, whether it is a network connection service party is_server is yes (true), whether it satisfies malicious security is_malicious intersection is no (false), the statistical security parameter stat_sec_param is 40, and whether the privacy intersection only outputs the intersection key only_output_key is no;
[0070] Fig.13 Participant No. 0 on the guest side reads the metadata intersect_guest_100000_1 from the FATE namespace experiment and passes it to data_transform_0. Then data_transform_0 constructs the metadata into a dense table dense and passes it to intersect_0. Intersect_0 uses the vole privacy intersection scheme, which is the scheme implemented by the present invention. Then "vole_params" defines the runtime parameters, including the random seed salt of "12345", the number of threads threat_num of 1, whether it is a network connection service party is_server is no, whether it satisfies malicious security is_malicious intersection is no, the statistical security parameter stat_sec_param is 40, and whether the privacy intersection only outputs the intersection key only_output_key is no.
[0071] Although the present invention has been described according to a limited number of embodiments, it will be apparent to those skilled in the art, with the benefit of the above description, that other embodiments may be envisioned within the scope of the invention thus described. In addition, it should be noted that the language used in this specification is selected primarily for readability and teaching purposes, rather than for explaining or defining the subject matter of the present invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the present invention is illustrative, not restrictive, with respect to the scope of the present invention, which is defined by the appended claims.
Claims
1. The FATE federated privacy set intersection method based on vector oblivious valuation is characterized by: The following steps are involved: S1. Based on the noisy parity learning problem, VOLE tuples are generated by initialization and expansion to obtain the VOLE protocol; S2. The first participant and the second participant run the VOLE protocol to obtain vectors and scalars, and construct the OPRF protocol in combination with the estimated key of the pseudo-random function; S3. Based on the OPRF protocol, identify the same elements in the sets held by the first participant and the second participant, and achieve the intersection of private sets.
2. The FATE federal privacy set intersection method based on vector oblivious valuation according to claim 1 is characterized in that: In S1, the initialization Setup stage selects a subset S with a potential of t on the input domain [n] and a random vector y with a length of t, S = {s1, s1, ..., s t }, For the group The first combination vector (a, b) is randomly selected from the set of vectors of length t. Number Field The set of vectors with length k above, a is the first vector, calculate the third vector c, c = ax + b, a is the first vector, b is the second vector, x is the data, perform multi-point function secret sharing MPFSS and seed generation Gen operations on the first key K0 and the second key K1, (K0, K1) = MPFSS.Gen (1 λ ,f S,xy ), where (K0, K1) is the second combination vector, 1 λ is the safety parameter, f S,xy is a multi-point function, for each element s in the set S i , the function is in s i The value at x·y i , for any element z not in the set S, the multipoint function f S,xy The value at z is 0, generating the first seed seed0 and the second seed seed1. The number field defined for the above addition and multiplication operations, In the Expand phase, when the index σ=0, the first seed seed0 is decompressed. The calculation is done by the spread function spread n Define a sparse vector μ of length n, μ←spread n (S,y), sparse vector μ represents the multi-point function f S,y The range vector spread n (S,y), that is, for each element s in the set S i , the sparse vector μ in s i The value of y i For any element z that is not in the set S, the value of the sparse vector μ at z is 0, and the function evaluation FullEval is performed on the first key K0, v0 = MPFSS.FullEval (0, K0), where v0 is the multi-point function f S,xy The range vector spread n The first secret sharing of (S,xy) outputs the second combination vector in, is the first input vector, is the second input vector, C k,n is a random matrix known to both parties, Number Field The set of all two-dimensional matrices with k rows and n columns above, when σ=1, decompress the second seed seed1→ Perform function evaluation FullEval on the first key K0, v1 = MPFSS.FullEval (1, K1), v1 is the multi-point function f S,xy The range vector spread n The second secret sharing of (S,xy) satisfies spread n (S,xy)=v0+v1, get the output vector 3. The FATE federal privacy set intersection method based on vector oblivious valuation according to claim 2 is characterized in that: In S2, the first participant P0 and the second participant P1 run the VOLE protocol, the first participant P0 obtains the fourth vector A and the sixth vector C, the second participant P1 obtains the fifth vector B and the scalar Δ, and C=A·Δ+B is satisfied, and the fourth vector A, the fifth vector B and the sixth vector C all belong to Number Field The set of vectors of length m on which the scalar Δ belongs to the number field m is the vector length of the VOLE protocol output determined by the first participant P0 and the second participant P1; The first participant P0 and the second participant P1 jointly select a random matrix M * , The first participant P0 defines a matrix M, M∈{0,1} n×m , the i-th row of the matrix M is equal to the random matrix M * The xth i The first participant P0 solves the equation system M·P T =(0,0,…,0) T , T is transposed, and the seventh vector P is obtained, The first participant P0 sends the fourth vector A and the seventh vector P to the second participant P1, and the second participant P1 sets an estimation key K of a pseudo-random function; The valuation key K is expressed as: K=B+Δ(A+P) By evaluating the key K, we get MK T =M·B T +Δ(M·A T +M·P T )=MB T +ΔMA T =MC T , set the pseudo-random function Among them, H is a random oracle, Represents the matrix M * The xth row corresponds to the OPRF protocol, where the first participant P0 obtains the estimate of the data x under the pseudo-random function. The second participant P1 uses the valuation key K to calculate the valuation F of any input random vector y under the random vector function K (y).
4. The FATE federal privacy set intersection method based on vector oblivious valuation according to claim 3 is characterized in that: In S3, for each element x in the first privacy intersection set X of the first participant P0, i , the first participant P0 and the second participant P1 execute the OPRF protocol so that the first participant P0 obtains the valuation result The second participant P1 obtains the second participant's valuation key k i , using the second party's estimated key k i The elements in the second privacy intersection set Y of the second participant P1 are evaluated, and all evaluation results are sent to the first participant P0. The evaluation results are compared to identify the same elements in the first privacy intersection set X and the second privacy intersection set Y through the first participant P0.
5. The FATE federated privacy set intersection system based on vector oblivious valuation is characterized by: A method for implementing the FATE federated privacy set intersection method based on vector oblivious valuation as described in any one of claims 1 to 4, comprising a framework layer, a component layer, an interface layer and a function layer; The framework layer includes the Fate Board front-end interface and the FaderatedML task scheduling component; The component layer includes a privacy set intersection host component, a privacy set intersection guest component and a database, and the privacy set intersection host component and the privacy set intersection guest component are connected to the corresponding database; The interface layer is used to implement cross-language function calls and data interface conversion between C++ and Python; The function layer includes a sender and a receiver of a privacy request protocol VOLE-PSI connected via a transmission control protocol TCP; The FaderatedML task scheduling component is connected to the privacy set intersection host component and the privacy set intersection guest component respectively. The privacy set intersection host component and the privacy set intersection guest component are both connected to the interface layer. The interface layer is connected to the sender and the receiver respectively.
