Patent Application
20210279581
The present disclosure relates to a prediction model conversion method and a prediction model conversion system for performing prediction processing of a neural network while keeping content secret.
In recent years, various companies have been providing services that use neural networks. For example, a service that identifies types of subjects from uploaded images, or a service that predicts a user's preferences and recommends products based on the user's purchase history, can be given as examples.
Because such services use personal information such as images or purchase histories, it is necessary to protect the user's private information. There is also demand for a system that can provide a service to users without allowing third parties to know information related to the knowledge of the service provider.
For example, NPL 1 (SecureML), NPL 2 (CryptoNets), and NPL 3 (MOBIUS) disclose techniques for prediction processing while keeping data secret.
However, NPL 1 (SecureML) has a problem in that the prediction accuracy is significantly lower than a typical neural network. Additionally, NPL 2 (CryptoNets) and NPL 3 (MOBIUS) have problems in that the amount of computation is extremely high and the prediction accuracy is low.
Accordingly, the present disclosure provides a prediction model conversion method and a prediction model conversion system that improve the efficiency of prediction processing. Furthermore, by employing the above-described configuration, the present disclosure reduces the amount of computation, which improves the processing speed and reduces a drop in the prediction accuracy.
To solve the above-described problems, one aspect of a prediction model conversion method includes: converting a prediction model by converting at least one parameter which is included in the prediction model and is for performing homogenization processing into at least one parameter for performing processing including nonlinear processing, the prediction model being a neural network; and generating an encrypted prediction model that performs prediction processing with input in a secret state remaining secret by encrypting the prediction model that has been converted.
Additionally, to solve the above-described problems, one aspect of a prediction model conversion system includes: a prediction model converter that converts a prediction model by converting at least one parameter which is included in the prediction model and is for performing homogenization processing into at least one parameter for performing processing including nonlinear processing, the prediction model being a neural network; and a prediction model encryptor that generates an encrypted prediction model that performs prediction processing with input in a secret state remaining secret by encrypting the prediction model that has been converted.
According to the prediction model conversion method and prediction model conversion system of the present disclosure, the speed of prediction processing, which can be executed while keeping input secret, can be improved, and a drop in prediction accuracy can be reduced.
These and other advantages and features will become apparent from the following description thereof taken in conjunction with the accompanying Drawings, by way of non-limiting examples of embodiments disclosed herein.
In recent years, various companies have been providing services that use neural networks. Services which identify the type of subjects from uploaded images, services that recommend products that a user may like based on the user's purchase history, or services that predict a user's physical state or mental state based on the user's biometric information (e.g., pulse, blood sugar level, body temperature, or the like) and providing feedback to the user can be given as examples of services that use neural networks.
With such services, the input information from the user, e.g., input information such as images uploaded by the user, the user's purchase history or biometric information, or the like, often contains sensitive information, and it is therefore necessary to protect the user's private information. There is thus a need for technology to perform the training processing and prediction processing of neural networks which enables users to use services without disclosing their private information to service providers.
From the service providers' standpoint as well, there is demand for a system that can provide a service to users without allowing users and third parties aside from users to know information related to the knowledge involved with the service.
As a technique that satisfies these two conditions, for example, a method of performing prediction processing of a neural network using a method called secure computation, which makes it possible to perform computations while keeping data secret, is conceivable. Secure computation is a technique that makes it possible to keep the computation process and results secret from the entity that stores the data. For example, data can be stored on a server managed by a third party, such as in the cloud, and any operation can be executed on the stored data. Because no third party can know the input data, computation process, or computation results, analysis processing of sensitive information, such as personal information, can be outsourced.
For example, NPL 1 (SecureML), NPL 2 (CryptoNets), and NPL 3 (MOBIUS) disclose techniques for prediction processing while keeping data secret.
However, the practicality of the methods described in these documents is questionable, due to the reduced prediction accuracy and the huge amount of computation.
After diligently examining the foregoing issue, the inventors found new parameters for processing including homogenization processing and nonlinear processing, by converting the parameters for homogenization processing in prediction processing. The inventors then found that by using the new parameters, a single layer can execute processing that includes both the homogenization processing and the nonlinear processing in a neural network.
Accordingly, the present disclosure provides a prediction model conversion method and a prediction model conversion system capable of improving the speed of prediction processing, which can be executed while keeping input secret, and reducing a drop in prediction accuracy.
One aspect of the present disclosure is as follows.
A prediction model conversion method according to one aspect of the present disclosure includes: converting a prediction model by converting at least one parameter which is included in the prediction model and is for performing homogenization processing into at least one parameter for performing processing including nonlinear processing, the prediction model being a neural network; and generating an encrypted prediction model that performs prediction processing with input in a secret state remaining secret by encrypting the prediction model that has been converted.
In this manner, by converting the plurality of parameters for performing the homogenization processing into at least one parameter for performing processing including the nonlinear processing, the processing related to the homogenization processing and the nonlinear processing can be performed through simpler processing. As a result, the number of times the processing is performed is reduced, which makes it possible to reduce the amount of computation in the prediction processing. Additionally, reducing the number of times the processing is performed makes it possible to reduce the occurrence of computation error, which in turn makes it possible to reduce a drop in prediction accuracy.
For example, in the prediction model conversion method according to one aspect of the present disclosure, the at least one parameter for performing the homogenization processing may include a plurality of parameters, the at least one parameter for performing the processing including the nonlinear processing may be one parameter, and in the converting, the plurality of parameters for performing the homogenization processing may be converted into one parameter for performing the processing including the nonlinear processing.
Through this, the equations used for the processing including the nonlinear processing can be made into simple equations. As a result, the amount of computation in the prediction processing is reduced, and the speed of the prediction processing is improved.
For example, in the prediction model conversion method according to one aspect of the present disclosure, the homogenization processing may be processing performed by an equation yi=sixi+ti, where xi is an input and yi is an output, si and ti may be the plurality of parameters for performing the homogenization processing, the processing including the nonlinear processing may be processing performed by Equation (1), and ki may be the at least one parameter for performing the processing including the nonlinear processing, and may be determined using si and ti.
Through this, the output after the nonlinear processing can be obtained by inputting the input xi of the homogenization processing into the aforementioned Equation (1). As a result, the amount of computation in the prediction processing is reduced, and the speed of the prediction processing is improved.
For example, in the prediction model conversion method according to one aspect of the present disclosure, ki may be expressed by Equation (2).
Here, u is a theoretical maximum value during computation of the prediction processing, and p is a divisor used in the encrypting.
Through this, an appropriate value can be obtained for the parameter ki, even if the value of si is positive and too large or is negative and too small.
For example, in the prediction model conversion method according to one aspect of the present disclosure, in the generating: the prediction model may be encrypted by distributing, through a secret sharing method, the prediction model that has been converted, and in the distributing of the prediction model, the at least one parameter for performing the processing including the nonlinear processing may be distributed.
Through this, the prediction model can be kept secret, and the prediction processing can be performed safely. To apply the secret sharing method, integerization processing such as dropping numbers below the decimal point in the prediction model is required, which increases the possibility of computation errors and reduces the prediction accuracy. However, by converting the parameters of the homogenization processing to the parameters of the processing including the nonlinear processing, the stated integerization processing is no longer necessary, and computation errors can be eliminated even when the secret sharing method is used. This reduces the amount of computation and improves the accuracy of the prediction processing, and furthermore reduces a drop in the prediction accuracy.
For example, the prediction model conversion method according to one aspect of the present disclosure may further include determining a divisor used in the secret sharing method in a range greater than a possible value of an element of the prediction model.
In the secret sharing method, using a large numerical value as the divisor (i.e., modulus p) increases the amount of computation, and thus determining the optimal divisor makes it possible to perform the prediction processing with a minimum amount of computation.
For example, in the prediction model conversion method according to one aspect of the present disclosure, the prediction model may be a binarized neural network including a plurality of parameters each having a binary value of −1 or 1.
In this manner, using a binarized neural network as the prediction model makes it possible to shorten the computation time for the matrix product operation. Additionally, because the prediction model is a binarized neural network, the process of converting negative numerical values in the prediction model to positive numerical value is simplified. This makes it possible to reduce a drop in the speed of the prediction processing.
For example, the prediction model conversion method according to one aspect of the present disclosure may further include training the prediction model using training data collected in advance, and a parameter obtained through the prediction processing as the at least one parameter for performing the homogenization processing may be converted in the converting.
Through this, it is easier to create a prediction model suitable for deriving the correct prediction results. The prediction accuracy can therefore be improved.
For example, in the prediction model conversion method according to one aspect of the present disclosure, in the converting, the divisor used in the secret sharing method may be added to a negative numerical value in a plurality of parameters included in the prediction model to convert the negative numerical value to a positive numerical value.
In general, the prediction accuracy increases as the numerical value of a parameter increases, and the speed of the computations increases as the numerical value decreases. Therefore, for example, from the perspective of balancing prediction accuracy with prediction speed, the value of the divisor used in the secret sharing method is determined and added to the negative numerical value. Accordingly, when the converted prediction model is used, a drop in both the prediction accuracy and the prediction speed can be reduced. Furthermore, because all the parameters in the converted prediction model are represented by positive numerical values, the converted prediction model can be distributed through the secret sharing method. The prediction processing can therefore be performed while keeping the input secret.
For example, in the prediction model conversion method according to one aspect of the present disclosure, in the converting, a negative numerical value is converted to a positive numerical value by converting a value in a plurality of parameters included in the prediction model to a set including a sign part indicating a sign of the numerical value as 0 or 1 and a numerical value part indicating an absolute value of the numerical value.
In this conversion processing, when, for example, one of the parameters in the prediction model is −10, −10 is converted into a pair (1,10) with a sign part indicating the sign and a numerical value part indicating the absolute value of the numerical value. As such, because the negative number −10 is converted into a pair of negative numerical values 1 and 10, the parameters in the prediction model are expressed only by positive numerical values. Accordingly, applying the conversion processing to the prediction model makes it possible to distribute the converted prediction model using the secret sharing method.
For example, the prediction model conversion method according to one aspect of the present disclosure may further include: calculating a feature amount from data obtained by sensing; and distributing, through the secret sharing method, the feature amount that has been calculated.
Through this, sensitive information, such as personal information obtained from the user through sensing, can be distributed through the secret sharing method, in the same manner as the prediction model. Thus with the prediction model conversion method according to one aspect of the present disclosure, the input for the prediction processing (here, user information) can be kept secret, i.e., the prediction processing can be performed while protecting the user's private information.
For example, the prediction model conversion method according to one aspect of the present disclosure may further include: executing prediction processing by the prediction model that has been distributed, by inputting, to the prediction model that has been distributed, the feature amount that has been distributed, wherein the executing includes the nonlinear processing, and the nonlinear processing is processing of converting an input to the nonlinear processing into 1 when the input is 0 or a numerical value corresponding to a positive, and into a positive numerical value corresponding to −1 when the input is a numerical value corresponding to a negative.
Through this, the numerical value of the input can be converted so that the converted numerical value falls within a positive numerical value range that guarantees the accuracy of the prediction. This makes it possible to reduce a drop in the speed and the prediction accuracy of the prediction processing.
Note that the following embodiments describe specific examples of the present disclosure. The numerical values, shapes, constituent elements, steps, orders of steps, and the like in the following embodiments are merely examples, and are not intended to limit the present disclosure. Additionally, of the constituent elements in the following embodiments, constituent elements not denoted in the independent claims, which express the broadest interpretation, will be described as optional constituent elements. Additionally, the drawings are not necessarily exact illustrations. Configurations that are substantially the same are given the same reference signs in the drawings, and redundant descriptions may be omitted or simplified.
Additionally, variations on the embodiments conceived by one skilled in the art, other embodiments implemented by combining constituent elements from parts of each embodiment in all of the embodiments, and the like, for as long as they do not depart from the essential spirit thereof, fall within the scope of the present disclosure.
A prediction model conversion method and a prediction model conversion system according to the present embodiment will be described below with reference to the drawings.
First, an overview of the prediction model conversion system will be given.
Prediction model conversion system 400 according to the present embodiment is a prediction model conversion system for performing prediction processing while keeping input secret. More specifically, the prediction model conversion system is a system that uses an encrypted prediction model (“distributed prediction model” hereinafter) and user information encrypted using the same method as the distributed prediction model (“distributed feature amount” hereinafter) to perform prediction processing while maintaining the encryption. In other words, the stated input is encrypted data input to a neural network that executes the prediction processing (here, data processing devices 300, 310, and 320). Note that the prediction model is data necessary for the prediction processing, including parameters, a weighting matrix, and the like used in the prediction processing.
The prediction model conversion system will also be referred to as a “secret prediction system”. Distribution and encryption will also be referred to below as “concealing”.
The configuration of prediction model conversion system 400 according to the present embodiment will be described next with reference to
As illustrated in
Although the present embodiment describes prediction model conversion system 400 illustrated in
For example, prediction model conversion system 400 may include: prediction model converter 104 that converts a prediction model, which is a neural network, by converting at least one parameter included in the prediction model and being for performing homogenization processing into at least one parameter for performing processing including nonlinear processing; and a prediction model encryptor (e.g., prediction model distributor 105) that generates an encrypted prediction model that performs prediction processing with input in a secret state remaining secret by encrypting the prediction model that has been converted.
Note that prediction model distributor 105 is an example of a prediction model encryptor. For example, prediction model distributor 105 encrypts the prediction model by distributing, through a secret sharing method, the prediction model that has been converted.
Prediction model conversion system 400 may further include, for example, feature amount calculator 202 that calculates a feature amount from data obtained by sensing user information, and feature amount distributor 203 that distributes, through the secret sharing method, the feature amount that has been calculated.
Prediction model conversion system 400 may further include, for example, prediction processor 302 that executes prediction processing by the prediction model that has been distributed by inputting, to the prediction model that has been distributed, the feature amount that has been distributed.
In prediction model conversion system 400, for example, a company or organization secretly sends data required for prediction processing (“prediction model” hereinafter) from data providing device 100 to three cloud servers, i.e., data processing devices 300, 310, and 320. When a user uses a service of the secret prediction system, the user secretly transmits their own information (“feature amount” hereinafter) from user terminal device 200 to the three cloud servers, i.e., data processing devices 300, 310, and 320. By communicating with each other, the three cloud servers compute prediction results while keeping the data secret, with each cloud server using data obtained by the other cloud servers. Each of the three data processing devices 300, 310, and 320 then transmits the obtained prediction results to user terminal device 200. User terminal device 200 decrypts the prediction results received from the three data processing devices 300, 310, and 320.
Note that there may be at least one data providing device 100, and there may be at least one user terminal device 200 as well. Furthermore, although prediction model conversion system 400 includes the three data processing devices 300, 310, and 320 in the example illustrated in
Note that the communication among the devices constituting prediction model conversion system 400 need not be real-time communication. For example, user terminal device 200 may collect a given number of pieces of user information from sensing or request commands for prediction processing performed while remaining in a secret state (also called simply “prediction processing” hereinafter) and then transmit those items to two or more of data processing devices 300, 310, and 320.
Each of the constituent elements of the prediction model conversion system according to the present embodiment will be described in detail hereinafter with reference to the drawings.
Data providing device 100 will be described hereinafter with reference to
As illustrated in
As illustrated in
Data providing device 100 creates a prediction model by training a neural network using the knowledge held by the company, organization, or the like as training data. The knowledge held by the company or organization is, for example, data in which biometric information such as blood pressure, heartbeat, CT scan information, or the like is associated with medical cases corresponding to that biometric information. Data providing device 100 creates the prediction model by training, for example, a binarized neural network (BNN) using that training data. Then, by distributing the created prediction model through a secret sharing method, data providing device 100 transmits the prediction model to the plurality of data processing devices 300, 310, and 320 in a secret state.
The various constituent elements of data providing device 100 will be described next.
Training data storage 101 stores the training data for creating the prediction model required to perform prediction processing while keeping the input secret. The training data is a set including data having the same nature as a feature amount calculated by feature amount calculator 202 of user terminal device 200 (see
Trainer 102 is, for example, a BNN, and creates the prediction model by performing training processing through a predetermined method using the training data stored in training data storage 101. For example, the method described in NPL 4 (Matthieu Courbariaux et al, “Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or −1” (https://arxiv.org/abs/1602.02830)) is used for the training processing.
Prediction model storage 103 stores the prediction model created by trainer 102.
Prediction model converter 104 converts the prediction model obtained through the training processing by trainer 102. Here, prediction model converter 104 performs conversion processing on the prediction model stored in prediction model storage 103. Note that the prediction model includes, for example, parameters, equations, weighting matrices, and the like used in the prediction processing. The prediction processing is executed by prediction processor 302 of each of data processing devices 300, 310, and 320. In the present embodiment, prediction processor 302 is a BNN. The prediction model will be described in detail hereinafter with reference to the drawings.
γ, σ, ε, β, and μ in Equation (B) and Equation (C) are trained parameter vectors in the prediction model of
In the prediction processing, the nonlinear processing is always executed immediately after the homogenization processing. As such, the input data of the nonlinear processing is the output data of the homogenization processing. The positive or negative sign of the input data (i.e., the output data of the homogenization processing) is determined in the nonlinear processing. In other words, the nonlinear processing can be substituted by processing which returns a value that has the same positive and negative signs as the output of the homogenization processing (Equation (D1) and Equation (D2) below). For example, by dividing both sides of Equation (A) in
Here, if the range of values of the input data in the prediction processing is set, the theoretical maximum of the values during the computation in the prediction processing can be computed in advance based on the range of values of the input data and the trained parameters.
Note that ki is a parameter for performing the processing including the nonlinear processing, and is determined using si and ti. Additionally, in Equation (G) in
With the prediction processing according to the present embodiment, the processing associated with the new homogenization processing and the nonlinear processing can be executed using simpler processing. This simple processing is processing including nonlinear processing (called “homogenization+nonlinear processing” hereinafter), and is processing performed using the equation illustrated in
As described above, the at least one parameter for performing the homogenization processing are a plurality of parameters, and the prediction model converter converts the plurality of parameters for performing the homogenization processing into one parameter for performing the homogenization+nonlinear processing.
Additionally, prediction model converter 104 may perform the computations of Equation (A), Equation (B), and Equation (C) illustrated in
By computing the Equations which can be computed in advance before the encryption and taking the result as the new prediction model, the amount of computation and amount of communication of data processing devices 300, 310, and 320 can be reduced, which makes it possible to reduce a drop in the prediction accuracy.
As described above, the secret sharing method cannot handle decimals. As such, when prediction model distributor 105 distributes a prediction model through the secret sharing method, decimals cannot be handled. Accordingly, as illustrated in FIGS. 10A and 10B, prediction model converter 104 multiplies new prediction models s and t, created by performing the computations of Equation (A), Equation (B), and Equation (C) illustrated in
As described above, the secret sharing method cannot handle negative numerical values (i.e., negative integers). As such, when prediction model distributor 105 distributes a prediction model through the secret sharing method, negative numerical values cannot be handled. Accordingly, prediction model converter 104 may add the divisor (i.e., modulus p) used in the secret sharing method to a negative numerical value in a plurality of parameters included in the prediction model in order to convert the negative numerical value to a positive numerical value. For example, as illustrated in
The advance computation of the prediction model will be described in detail hereinafter with reference to
Prediction model distributor 105 is an example of a prediction model encryptor. Prediction model distributor 105 uses a predetermined method to distribute and make secret the converted prediction model created by prediction model converter 104. For example, prediction model distributor 105 encrypts the prediction model which has been converted (the so-called “converted prediction model”) by distributing the prediction model using the secret sharing method, and when distributing the prediction model, distributes the parameters for performing the processing including the nonlinear processing (the so-called “homogenization+nonlinear processing”).
Prediction model distributor 105 creates a prediction model (also called a “distributed prediction model” hereinafter) capable of performing prediction processing in data processing devices 300, 310, and 320 with the feature amount obtained from user terminal device 200 remaining encrypted (i.e., in a secret state). Prediction model distributor 105 creates the distributed prediction model by distributing the prediction model using, for example, the Shamir (2,3) threshold secret sharing method (NPL 5: Adi Shamir, “How to share a secret” (https://cs.jhu.edu/˜sdoshi/crypto/papers/shamirturing.pdf)).
Note that as mentioned above, the secret sharing method is not limited to the method of NPL 5, and the method described in NPL 6 (Ronald Cramer et al., “Share Conversion, Pseudorandom Secret-Sharing and Applications to Secure Computation” (https://rd.springer.com/chapter/10.1007/978-3-540-30576-7_19)) or NPL 7 (Toshinori Araki, and four others, “High-Throughput Semi-Honest Secure Three-Party Computation with an Honest Majority”, (https://eprint.iacr.org/2016/768.pdf)) may be used instead.
The upper bound of the value during the computation in the prediction processing (e.g., the maximum value u) can be computed from the range of values of the input data and the trained parameters. It is necessary that modulus p used in the distribution processing be set so that the value during computation does not exceed p, and thus the upper bound of the value during the computation of the prediction processing (i.e., the maximum value u) is computed in advance, with numbers greater than or equal to that value being determined as modulus p and held in prediction model distributor 105.
Communicator 106 communicates with data processing devices 300, 310, and 320. Communicator 106 transmits the distributed prediction model created by prediction model distributor 105 (the so-called “encrypted prediction model”) to the plurality of data processing devices 300, 310, and 320.
User terminal device 200 senses information pertaining to a user, such as the user's blood pressure, heartbeat, CT scan information, or the like, i.e., private data, calculates a feature amount, and transmits the feature amount to data processing devices 300, 310, and 320. At this time, user terminal device 200 distributes the feature amount through the secret sharing method, for example, to transmit the feature amount to data processing devices 300, 310, and 320 in a secret state. Then, user terminal device 200 requests prediction results corresponding to the calculated feature amount from data processing devices 300, 310, and 320, obtains the prediction results from data processing devices 300, 310, and 320, and uses a service in prediction model conversion system 400. At this time, user terminal device 200 obtains encrypted prediction results from data processing devices 300, 310, and 320, and decrypts and uses the prediction results.
Sensor 201 is configured including one or more measurement devices, which are sensors, for sensing information on the user (user information).
The user information to be sensed may be, for example, the user's vital data such as blood pressure, body temperature, heartbeat, or the like, image information such as a face image, ultrasound information, CT scan information, or the like obtained by capturing an image of, or measuring, the user's body.
Additionally, the user information to be sensed may be, for example, location information obtained by GPS (Global Positioning System), log information indicating a history of the user's operation of an electric device or a moving object such as a vehicle, the user's purchase history information for products and so on, or the like.
The log information may be various types of information obtained or measured in association with, for example, steering operations, acceleration operations, braking operations, operations for shifting gears, or the like in a vehicle, and may be, for example, information that associates a displacement amount, speed, acceleration, or the like to a time of operation.
The user information to be sensed may, for example, be private data, which is personal matters that the user does not want others to know.
Prediction model conversion system 400 is a prediction model conversion system for performing prediction processing of a BNN while keeping the user's private data secret, and is a secret prediction system that calculates a prediction result with the result remaining secret. Here, the descriptions assume that the information about the user sensed by sensor 201 is private data.
Feature amount calculator 202 calculates a feature amount from the user's private data obtained by sensor 201. The feature amount calculated by feature amount calculator 202 can be expressed as a vector containing a plurality of components.
The feature amount includes, for example, a component expressing an indicator related to at least one of a shape, size, weight, condition, and movement of the whole or a part of the user's body.
Note that the part of the user's body that is the subject of the feature amount can be any part of the body, such as the eyes, nose, ears, hands, feet, organs, blood vessels, or the like.
A physical state, and more specifically, the state of the user with respect to various items used in health checkups, the amount of water in the body, blood pressure, oxygen saturation level, or the like can be given as the state of the whole or part of the user's body.
Body motion (i.e., body movement), and more specifically, the number of times the user turns over in bed per unit time, vibrations such as shaking of limbs or facial twitching, microvibrations such as heart rate, breathing rate, or inhalation/exhalation ratio can be given as examples of the movement of the whole or a part of the user's body.
Note that when the private data is a face image of the user, the feature amount is the primary component of the characteristic parameters in the face image, for example. The feature amount may be, for example, information such as the position, area, width, or the like of a given region of the user's face image. Additionally, the feature amount may, for example, be information expressed by vectors that include, as components (e.g., coefficients of each term when expressed as a polynomial expression), a trend in some element measured for the user by sensor 201, corresponding to the time axis, from history information indicating that element.
Note that the feature amount extracted from the user information obtained by sensor 201 can itself be private data.
Feature amount distributor 203 distributes and makes secret the feature amount calculated by feature amount calculator 202, through a predetermined method. Feature amount distributor 203 creates a feature amount that has been distributed (also called a “distributed feature amount” hereinafter) by distributing the feature amount using a method in which data processing devices 300, 310, and 320 can perform prediction processing using the feature amount still in a distributed state (i.e., still in a secret state), e.g., the Shamir (2,3) threshold secret sharing method (NPL 5).
The secret sharing method is a technique for generating a plurality of pieces of distributed information from secret information. The distributed information is created in such a way that the secret information can be recovered from a predetermined combination, but not from other combinations. The predetermined combination can take on a variety of structures, and those structures are called “access structures”. There are many different types of access structures. A threshold-type access structure will be described here as a typical access structure. The threshold-type access structure is expressed by two parameters, namely a number n of the pieces of distributed information to be generated, and a threshold m (m≤n). The secret information can be recovered from at least m pieces of the distributed information, but not from fewer than m pieces of the distributed information. Secret sharing methods with a threshold-type access structure include, for example, the Shamir (2,3) threshold secret sharing method (NPL 5) mentioned above, which includes distribution processing for generating three pieces of distributed information with the secret information as input, and recovery processing for recovering the secret information from two or more pieces of the distributed information.
Note that the secret sharing method is not limited to the method described in NPL 5, and the method described in NPL 6 or NPL 7 may be used. Modulus p used in the distribution processing is determined in advance by the system and held by feature amount distributor 203.
Decryptor 204 receives the prediction results corresponding to the feature amounts transmitted by user terminal device 200 and distributed to data processing devices 300, 310, and 320 from data processing devices 300, 310, and 320 and decrypts the prediction results. These prediction results are results obtained using the feature amount and prediction model distributed using the secret sharing method while remaining in a distributed state, and are so-called “encrypted prediction results”. The method described in any one of NPL 5, NPL 6, and NPL 7, for example, may be used as the method for decrypting the prediction results.
Prediction result utilizer 205 uses the prediction results decrypted by decryptor 204. One example of the utilization of the prediction results is presenting the prediction results to the user, i.e., the presentation of the prediction results. The prediction results may be presented as an image or as audio, for example. When the prediction results are presented as an image, the presentation is, for example, in the form of a graph, statistical information, or the like based on the prediction results. When the prediction results are presented as audio, the presentation is done by, for example, outputting audio based on the prediction results. Note that the prediction results may be presented as both an image and as audio. In this case, user terminal device 200 may be provided with, for example, a display that displays images, an audio output device such as a speaker that outputs audio, and other types of user interfaces, and the prediction results may be presented.
Additionally, prediction result utilizer 205 may further perform predetermined operations, information searches, or the like based on the prediction results and present the user with guidance for receiving a medical checkup at a hospital, advice for improving lifestyle habits, a recommended program, or the like.
Communicator 206 communicates with the plurality of data processing devices 300, 310, and 320. Communicator 206 transmits, to data processing devices 300, 310, and 320, the feature amounts created and distributed by feature amount distributor 203. As described in detail below in the section “1.4 Data Processing Devices”, the plurality of data processing devices 300, 310, and 320 execute the prediction processing upon receiving these distributed feature amounts, using the distributed feature amounts which remain in a secret state. Additionally, communicator 106 receives the prediction results computed by data processing devices 300, 310, and 320 and transmits those prediction results to decryptor 204. As described above, these prediction results are encrypted prediction results.
The data processing devices will be described next. As illustrated in
Distributed prediction model storage 301 stores the prediction model which is distributed, received from data providing device 100 (the so-called “distributed prediction model”).
Prediction processor 302 performs the prediction processing using the distributed prediction model stored in distributed prediction model storage 301 and the distributed feature amount received from user terminal device 200. Prediction processor 302 performs the prediction processing using the distributed prediction model and the distributed feature amount still in a distributed state (i.e., still in a secret state), and finds a distributed prediction result. Note that the distributed prediction result is an encrypted prediction result.
The prediction processing will be described in detail next with reference to the drawings.
Prediction processor 302 inputs the distributed feature amount to the distributed prediction model and executes the prediction processing using the distributed prediction model. The prediction processing includes nonlinear processing. The prediction processing is executed through four types of processing: matrix product operation, data distribution homogenization+nonlinear processing, homogenization processing, and maximum value search. In past prediction processing, the data distribution homogenization processing and nonlinear processing are executed separately, but in the prediction processing according to the present embodiment, prediction model converter 104 generates the new parameter k, which enables the homogenization processing and the nonlinear processing to be performed through the simple equation illustrated in
In the prediction processing, the matrix product operation and data distribution homogenization+nonlinear processing are executed for a predetermined number of repetitions, and the maximum value search is then performed to obtain the prediction result (i.e., the distributed prediction result). Note that the flow of the prediction processing illustrated in
Each process in the prediction processing will be described below with reference to the drawings.
The matrix product operation will be described first. The matrix product operation computes a matrix product of the distributed feature amount, which is a distributed input vector, and a distributed weighting matrix included in the distributed prediction model. The distributed weighting matrix and the distributed feature amount will be described below.
The distributed feature amount will be described next. Like the prediction model, the distributed feature amount is an encrypted feature amount obtained by using the secret sharing method to distribute the feature amount calculated from the data obtained from the sensing (also called “sensing data” hereinafter). For example, referring to
The homogenization processing and nonlinear processing of the data distribution obtained from the matrix product operation will be described next. The equation illustrated in
The numerical value corresponding to a positive, expressed using modulus p, for example, may be from 0 to (p−1)/2, and the numerical value corresponding to a negative may be from (p+1)/2 to p−1. Note that which of the values from 0 to p−1 are to be numerical values corresponding to a positive or numerical values corresponding to a negative may be determined as desired.
Additionally, for example, a value having a most significant bit of 0 may be set to a numerical value corresponding to a positive, and a value having a most significant bit of 1 may be set to a value corresponding to negative.
The maximum value search will be described next. In the maximum value search, the element having the maximum value among all the elements in the distributed input vector is searched out. The maximum value search is realized, for example, by comparing magnitude relationships among all elements of the input vector for the maximum value search, and computing a logical product of the comparison results. More specifically, in the maximum value search, the magnitude relationship for each element is compared to all the other elements individually. The comparison result is expressed as a binary value of 0 and 1. For example, if the value of a given element is the same as the value of another element or greater than the value of the other element, the comparison result is expressed as 1, and if the value of a given element is less than the value of another element, the comparison result is expressed as 0. The comparison results of the magnitude relationships between all elements the corresponding other elements are held in a comparison table. In this case, for the element having the maximum value among all the elements, the results of the comparisons of the magnitude relationships with the other elements will be all 1s. Accordingly, when the logical product of the comparison results is computed, the logical product will be 1 only for the element having the maximum value, and will be 0 for all other elements. Using this property makes it possible to extract the element having the maximum value.
As described above, the four types of processing in the present embodiment, namely the matrix product operation, the homogenization+nonlinear processing for the data distribution, the homogenization processing for the data distribution, and the maximum value search, can be constituted by only the sum, product, magnitude relationship comparison, and logical product for the inputs in each type of processing. For example, according to NPL 8 (Takahashi Nishide et al, “Multiparty Computation for Interval, Equality, and Comparison Without Bit-Decomposition Protocol”, “Public Key Cryptography—PKC 2007”, Springer Berlin Heidelberg (https://rd.springer.com/chapter/10.1007/978-3-540-71677-8_23)) the sum, product, magnitude relationship comparison, and logical product of two distributed values can be computed without being decrypted. As such, by using the method of NPL 8 in the prediction processing, the prediction processing can be performed while keeping the input secret, without decrypting the distributed prediction model and the distributed feature amount.
Communicator 303 of data processing device 300 communicates with data providing device 100, user terminal device 200, and the other data processing devices 310 and 320. Communicator 303 receives the distributed prediction model from data providing device 100, and stores the received distributed prediction model in distributed prediction model storage 301. Communicator 303 receives the distributed feature amount from user terminal device 200, and transmits the received distributed feature amount to prediction processor 302. Additionally, communicator 303 transmits, to user terminal device 200, the distributed prediction result calculated by prediction processor 302.
As described above, data processing device 300 can perform the prediction processing without decrypting the distributed prediction model and the distributed feature amount, with those items remaining distributed, i.e., in a secret state.
An example of a prediction model conversion method according to the present embodiment will be described next.
The prediction model conversion method includes: converting a prediction model by converting at least one parameter included in the prediction model and being for performing homogenization processing into at least one parameter for performing processing including nonlinear processing, the prediction model being a neural network (S001); and generating an encrypted prediction model that performs prediction processing with input in a secret state remaining secret by encrypting the prediction model that has been converted (S002).
An example of operations performed by the prediction model conversion system will be described hereinafter.
An example of operations performed by prediction model conversion system 400 will be described next. The operations of prediction model conversion system 400 include (i) a training phase of data providing device 100 training and distributing the prediction model, and (ii) a prediction phase of the plurality of data processing devices 300, 310, and 320 predicting the feature amounts that have been distributed (the so-called “distributed feature amounts”) using the prediction model that has been distributed (the so-called “distributed prediction model”).
Operations of prediction model conversion system 400 in the training phase will be described first.
In training step S101, data providing device 100 (see
The prediction model for performing the prediction processing is created as a result. The created prediction model is stored in prediction model storage 103.
Next, in prediction model conversion step S102, data providing device 100 uses prediction model converter 104 to apply conversion processing to the created prediction model. Specifically, in prediction model conversion step S102, data providing device 100 converts the parameters included in the prediction model of the neural network and used in the homogenization processing (the homogenization parameters si and ti in
Through this, the homogenization processing and the nonlinear processing can be performed using a simple equation, which reduces the amount of computation. Additionally, because the output result of the homogenization+nonlinear processing is the same as the output obtained by executing the nonlinear processing after executing the homogenization processing, a drop in the prediction accuracy can be suppressed.
Next, in prediction model distribution step S103, data providing device 100 distributes the prediction model, converted in prediction mode conversion step S102, using the secret sharing method. A prediction model that has been distributed (the so-called “distributed prediction model”) is obtained as a result.
Next, in step S104, data providing device 100 transmits the distributed prediction model obtained in prediction model distribution step S103 to the plurality of data processing devices 300, 310, and 320.
Next, in step S105, each of data processing devices 300, 310, and 320 stores the distributed prediction model received from data providing device 100 in its own distributed prediction model storage 301.
As described above, in the training phase, data providing device 100 creates the prediction model for performing the prediction processing, and then creates the distributed prediction model by distributing the created prediction model using the secret sharing method. Through this, the prediction model can be transmitted to the plurality of data processing devices 300, 310, and 320, while keeping the prediction model secret.
The prediction phase of prediction model conversion system 400 will be described next.
As illustrated in
Next, in feature amount calculation step S202, user terminal device 200 uses feature amount calculator 202 to calculate a feature amount from the information received from sensor 201. The feature amount is a value indicating a feature of the information received from sensor 201. Referring again to
Next, in feature amount distribution step S203, user terminal device 200 distributes the feature amount calculated in feature amount calculation step S202 using the secret sharing method. A feature amount that has been distributed (the so-called “distributed feature amount”) is obtained as a result. A method for calculating the distributed feature amount will be described here with reference again to
Next, in step S204, user terminal device 200 transmits the distributed feature amounts to the plurality of data processing devices 300, 310, and 320. Specifically, as illustrated in
When each of the plurality of data processing devices 300, 310, and 320 receives the distributed feature amount from user terminal device 200, the data processing device reads out the distributed prediction model stored in its own distributed prediction model storage (distributed prediction model storage 301 illustrated in
In the prediction processing step, the plurality of data processing devices 300, 310, and 320 perform the prediction processing of the binarized neural network (BNN) using the distributed feature amount and the distributed prediction model that remain in the distributed state (remain in a secret state). Note that prediction processing step S205 will be described in detail later.
As a result, the plurality of data processing devices 300, 310, and 320 obtain the distributed prediction results as the results of the respective instances of prediction processing. Note that when the computations of the prediction processing are performed using the method of NPL 8, it is necessary, when performing the prediction processing, for the distributed information held by each of the plurality of data processing devices 300, 310, 320, and the data obtained from prediction processing on the distributed information, to be communicated among the plurality of data processing devices 300, 310, and 320.
Next, as illustrated in
Next, in step S207, user terminal device 200 receives the distributed prediction results transmitted from the plurality of data processing devices 300, 310, and 320, decrypts the received distributed prediction results, and obtains a prediction result.
Finally, in step S208, user terminal device 200 uses the obtained prediction result in prediction result utilizer 205. As described above, user terminal device 200 may, for example, present the prediction result to the user as an image, as audio, or the like, and may also present lifestyle habits improvements, methods for reducing stress, recommended programs, or the like in addition to the prediction result.
As described above, in the prediction phase, data providing device 100 creates the prediction model for performing the prediction processing, and then creates the distributed prediction model by distributing the created prediction model using the secret sharing method. Through this, the prediction model can be transmitted to the plurality of data processing devices 300, 310, and 320, while keeping the prediction model secret. Then, user terminal device 200 decrypts the prediction result, presents the prediction result to the user, and utilizes the prediction result.
Prediction processing step S205 of prediction model conversion system 400 will be described in further detail below.
As illustrated in
In prediction processing step S205, the plurality of data processing devices 300, 310, and 320 perform the prediction processing of the binarized neural network (BNN) using the distributed feature amount and the distributed prediction model which remain in a distributed state (remain in a secret state).
As illustrated in
First, in matrix product operation step S2052, upon receiving a distributed input vector, which is the distributed feature amount (see
Specifically, to describe this using operations performed by data processing device 300 as an example, upon receiving the distributed feature amount from user terminal device 200, data processing device 300 reads out the distributed prediction model stored in distributed prediction model storage 301. Then, data processing device 300 computes the matrix product of the distributed feature amount and the distributed prediction model, and obtains the first distributed vector, which is a first distributed feature amount.
Note that the distributed prediction model (here, the distributed weighting matrix) is obtained by using the secret sharing method to distribute the converted prediction model, which has been converted so that all of the elements are positive numerical values. As described above, the prediction model illustrated in
Next, in homogenization+nonlinear processing step S2053, the plurality of data processing devices 300, 310, and 320 compute a sum for each element included in the first distributed vector (see
More specifically, the second distributed vector yi is obtained by adding the converted homogenization parameter ki to each element xi of the first distributed vector and computing whether or not the result is at least 0, while keeping the values secret, as illustrated in
Next, the plurality of data processing devices 300, 310, and 320 perform matrix product operation step S2052 using the second distributed vector obtained the output in homogenization+nonlinear processing step S2053 and the distributed prediction model. Then, the plurality of data processing devices 300, 310, and 320 execute homogenization+nonlinear processing step S2053 using a third distributed vector obtained from matrix product operation step S2052 as an input. A fourth distributed vector is obtained as a result.
In this manner, the above-described series of steps, namely matrix product operation step S2052 and homogenization+nonlinear processing step S2053, are repeated a predetermined number of times. Referring again
Next, in matrix computation step S2055, the plurality of data processing devices 300, 310, and 320 calculate a matrix product of the fourth distributed vector, which has been obtained as an output by repeating the above-described series of steps S2052 to S2053 a predetermined number of times (here, twice), and a weighting matrix. A fifth distributed vector is obtained as a result.
Next, in homogenization processing step S2056, homogenization processing is executed on the fifth distributed vector obtained from matrix computation step S2055. A sixth distributed vector is obtained as a result.
Finally, in maximum value search step S2057, the element having the maximum value among the sixth distributed vectors obtained from homogenization processing step S2056 is searched out. The distributed prediction result is obtained as a result.
As described thus far, with the prediction model conversion method according to the present embodiment, the homogenization parameter included in the prediction model of the neural network is converted, and the new parameter ki is generated, so that the homogenization processing and the nonlinear processing can be executed through simple processing, which makes it possible to execute a plurality of instances of processing using, for example, the simple equation illustrated in
A prediction model conversion method and a prediction model conversion system according to the present disclosure have been described based on embodiments. However, the present disclosure is not limited to the foregoing embodiments. Variations on the embodiments conceived by one skilled in the art, other embodiments implemented by combining constituent elements from the embodiments, and the like, for as long as they do not depart from the essential spirit thereof, fall within the scope of the present disclosure. The present disclosure is also inclusive of the following cases.
(1) Although the foregoing embodiment describes an example in which data providing device 100 uses prediction model converter 104 to convert negative numerical values, among a plurality of parameters (also called “elements” hereinafter”) included in the prediction model, to positive numerical values. Prediction model converter 104 may perform the following conversion processing on the prediction model. Prediction model converter 104 may convert a negative numerical value to a positive numerical value by converting a value in a plurality of parameters included in the prediction model to a set including a sign part indicating a sign of the numerical value as a 0 or a 1 and a numerical value part indicating an absolute value of the numerical value. For example, for a given element x (where x is an integer), assume that x=ab (where a is the sign part of x, and b is the numerical value part indicating the absolute value of x). If the given element x is 0 or a positive numerical value, 0 is substituted for the sign part a, whereas if the given element x is a negative numerical value, 1 is substituted for the sign part a. The absolute value of x is substituted for the numerical value part b. The given element x is converted into a set (a,b) of a and b. By applying the above conversion processing to all the elements included in the prediction model, negative numerical values in the prediction model can be converted to positive numerical values. Therefore, all the elements included in the prediction model after the conversion processing are expressed only as positive numerical values (here, positive integers). Through this, prediction model distributor 105 can distribute the prediction model following the conversion processing using the secret sharing method.
(2) Although the foregoing embodiment does not specify the specific method for determining modulus p of the secret sharing method, the optimal modulus p may be determined through the following operations. If the ranges of values of the prediction model and input data are known, the upper bound of the values during the computation in the prediction processing can be found. For example, focusing on the first matrix product operation in the prediction processing, if the range of input vector values is from 0 to 255, the number of input vector dimensions is a, and the number of output vector dimensions is b, then the range of output vector values is from −255a to 255a, and the upper bound of the values during the computation in this matrix product operation is 255a. Furthermore, in the next homogenization+nonlinear processing, if the range of input vector values is from −255a to 255a, the number of input vector dimensions is b, the number of output vector dimensions is b, and the maximum value of the parameter k in the homogenization+nonlinear processing is c, the value of the output vector is −1 or 1, and the upper bound of the values during the computation is 255a+c. Thus by determining the upper bound of the values during each instance of processing in the prediction processing, the upper bound of the values during the computation in the overall prediction processing can be computed. For example, if the upper bound that has been found is u, the optimal modulus p that minimizes the amount of computation can be selected by selecting the smallest prime number greater than or equal to 2u+1 as modulus p of the secret sharing method.
(3) Additionally, because the amount of computation depends on the bit length of modulus p, if the bit length is the same, the amount of computation will be the same regardless of whether a small prime number is selected as modulus p or a large prime number is selected as modulus p. However, there are some algorithms in the secret sharing method that can be processed more efficiently using a larger prime number even when the bit length is the same, and thus the smallest prime number greater than or equal to 2u+1 may be selected as described above, or the largest prime number with the same bit length as that prime number may be selected as modulus p. This may further improve the efficiency.
(4) Although the foregoing embodiment describes processing using matrix product operation, homogenization processing, homogenization+nonlinear processing, and maximum value search processing as an example of the prediction processing, processing such as convolution and pooling may be used. An example of prediction processing using such processing is illustrated in
Upon obtaining the distributed feature amount from user terminal device 200, the plurality of data processing devices 300, 310, and 320 starts processing which repeats a predetermined number of times (step S301).
First, in step S302, the plurality of data processing devices 300, 310, and 320 perform the processing of convolution using distributed input vectors and distributed convolution parameters, and obtain the first distributed vector. The convolution processing can be performed by a combination of matrix product and addition.
Next, in step S303, the plurality of data processing devices 300, 310, and 320 perform the homogenization+nonlinear processing using the first distributed vector obtained as an output in step S302 and the converted homogenization parameters obtained by converting the homogenization parameters, and obtain the second distributed vector.
Next, in step S304, the plurality of data processing devices 300, 310, and 320 perform the processing of convolution using the second distributed vector obtained as an output in step S303 and distributed convolution parameters, and obtain the third distributed vector.
Next, in step S305, the plurality of data processing devices 300, 310, and 320 perform the pooling processing on the third distributed vector obtained as output in step S304, and obtain the fourth distributed vector. Pooling includes the processing for calculating the maximum (Max Pooling), average (Average Pooling), or sum (Sum Pooling) of a defined area, as illustrated in
Next, in step S306, the plurality of data processing devices 300, 310, and 320 perform the homogenization+nonlinear processing using the fourth distributed vector obtained as an output in step S305 and the converted homogenization parameters obtained by converting the homogenization parameters, and obtain the fifth distributed vector.
In the present embodiment, after repeating step S302 to step S306 a predetermined number of times (step S307), the matrix product operation is computed using an nth distributed vector, which is the output of the last instance of homogenization+nonlinear processing, and the distributed prediction model (step S308), the homogenization processing is performed using that output and the homogenization parameters (step S309), and finally, the maximum value search processing is performed (step S310). The distributed prediction result is obtained as a result.
(5) As an example of the maximum value search processing by prediction processor 302, the foregoing embodiment describes a processing method in which, for each element, the magnitude relationship is compared with all other elements, and the element for which the theoretical product of the comparison results is 1 is determined to be the element having the maximum value. However, the method is not limited thereto. For example, in the maximum value search processing, the element having the maximum value (called the “maximum value element” hereinafter) may be obtained through the following processing. The first element (“element A”) of the plurality of elements of the input vector for the maximum value search processing is set as a tentative maximum value element, and the magnitude relationships between element A and the remaining elements are compared in sequence. If an element (“element B”) is found that is larger than element A, which is the tentative maximum value element, element B is taken as the new tentative maximum value element, and the magnitude relationships between element B and the remaining elements are compared in sequence. Assuming that element B is the tentative maximum value element at the point where all the elements have been compared, the numerical value and number of element B is the output of the maximum value search processing.
(6) The maximum value search processing by prediction processor 302 may also be used to find the maximum value element through the following processing. For example, for all the elements of the input vector for the maximum value search processing, the magnitude relationship between neighboring elements is compared, and the smaller elements are excluded. The maximum value element can be obtained by repeating this processing and determining that the last remaining element is the maximum value element.
(7) Although the foregoing embodiment describes an example of processing in which user terminal device 200 utilizes the prediction result, the processing may be that described below. After receiving and decrypting the prediction result, the user may send information pertaining to the correctness and usefulness of the prediction result to data providing device 100.
(8) Although the foregoing embodiment describes an example of processing in which user terminal device 200 utilizes the prediction result, the processing may be that described below. After receiving and decrypting the prediction results from the plurality of data processing devices 300, 310, and 320, user terminal device 200 may send information pertaining to the correctness and usefulness of the prediction results to data providing device 100, along with information pertaining to the user who input information to the prediction model conversion system (called “user information” hereinafter).
(9) Although the foregoing embodiment describes an example of the processing performed by data providing device 100, the following processing may be performed. Data providing device 100 may re-train the prediction model based on a set of user information received from user terminal device 200 and information pertaining to the prediction results, or based on the information pertaining to the prediction results only. Data providing device 100 then distributes the prediction model, newly created through the re-training, using the secret sharing method, and transmits the model to the plurality of data processing devices 300, 310, and 320 as a new distributed prediction model. The plurality of data processing devices 300, 310, and 320 store the received new prediction model in the prediction model storage and update the prediction model.
(10) Although in the foregoing embodiment, the homogenization processing and the nonlinear processing can be performed using a simple equation, and the amount of computation can be reduced, by converting the homogenization parameters, the prediction processing may be converted in the following manner. Because the matrix product operation, the convolution operation, and so on, as well as the homogenization processing, are linear computations, the matrix product operation and the homogenization processing, or the convolution operation and the homogenization processing, can be performed simultaneously. Rather than combining the homogenization processing and the nonlinear processing, data providing device 100 may generate new prediction model by combining the matrix product operation and the homogenization processing or the convolution operation and the homogenization processing, and distribute that processing to data processing devices 300, 310, and 320. In this case, the prediction model is converted by generating new parameters using the weighting matrix, convolution parameters, and homogenization parameters, which are the parameters of the matrix product operation.
(11) Additionally, after converting the matrix product operation and the homogenization processing, and the convolution operation and the homogenization processing, into processing which can be performed simultaneously, the nonlinear processing may also be combined to convert the processing into prediction processing in which the matrix product operation, the homogenization processing, and the nonlinear processing, as well as the convolution operation, the homogenization processing, and the nonlinear processing, can be performed using a simple equation.
(12) Each device in the foregoing embodiments is specifically a computer system constituted by a microprocessor, ROM (Read Only Memory), RAM (Random Access Memory), a hard disk unit, a display unit, a keyboard, a mouse, and the like. A computer program is recorded in the RAM or hard disk unit. Each device realizes the functions thereof by the microprocessor operating in accordance with the computer program. Here, the computer program is constituted by a combination of a plurality of command codes that indicate commands made to a computer to achieve a predetermined function.
(13) Some or all of the constituent elements constituting the devices in the foregoing embodiments may be implemented by a single integrated circuit through system LSI (Large-Scale Integration). “System LSI” refers to very-large-scale integration in which multiple constituent elements are integrated on a single chip, and specifically, refers to a computer system configured including a microprocessor, ROM, RAM, and the like. A computer program is recorded in the RAM. The system LSI circuit realizes the functions of the devices by the microprocessor operating in accordance with the computer program.
The parts of the constituent elements constituting the foregoing devices may be implemented individually as single chips, or may be implemented with a single chip including some or all of the devices.
Although the term “system LSI” is used here, other names, such as IC (Integrated Circuit), LSI, super LSI, ultra LSI, and so on may be used, depending on the level of integration. Further, the manner in which the circuit integration is achieved is not limited to LSIs, and it is also possible to use a dedicated circuit or a general purpose processor. It is also possible to employ a FPGA (Field Programmable Gate Array) which is programmable after the LSI circuit has been manufactured, or a reconfigurable processor in which the connections or settings of the circuit cells within the LSI circuit can be reconfigured.
Further, if other technologies that improve upon or are derived from semiconductor technology enable integration technology to replace LSI circuits, then naturally it is also possible to integrate the function blocks using that technology. Biotechnology applications are one such foreseeable example.
(14) Some or all of the constituent elements constituting the foregoing devices may be constituted by IC cards or stand-alone modules that can be removed from and mounted in the apparatus. The IC card or module is a computer system constituted by a microprocessor, ROM, RAM, and the like. The IC card or module may include the above very-large-scale integration LSI circuit. The IC card or module realizes the functions thereof by the microprocessor operating in accordance with the computer program. The IC card or module may be tamper-resistant.
(15) The present disclosure may be realized by the methods described above. This may be a computer program that implements these methods on a computer, or a digital signal constituting the computer program.
Additionally, the present disclosure may also be computer programs or digital signals recorded in a computer-readable recording medium such as a flexible disk, a hard disk, a CD-ROM, an MO (Magneto-Optical Disc), a DVD, a DVD-ROM, a DVD-RAM, a BD (Blu-ray (registered trademark) Disc), semiconductor memory, or the like. The constituent elements may also be the digital signals recorded in such a recording medium.
Additionally, the present disclosure may be realized by transmitting the computer program or digital signal via a telecommunication line, a wireless or wired communication line, a network such as the Internet, a data broadcast, or the like.
Additionally, the present disclosure may be a computer system including a microprocessor and memory, where the memory records the above-described computer program and the microprocessor operates in accordance with the computer program.
Additionally, the present disclosure may be implemented by another independent computer system, by recording the program or the digital signal in the recording medium and transferring the recording medium, or by transferring the program or the digital signal over the network or the like.
(16) The above-described embodiments and variations may be combined as well.
The present disclosure can be applied in systems that protect privacy by ensuring data processing devices do not handle a user's sensitive information in plain text.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2019-003238 | Jan 2019 | JP | national |
This is a continuation application of PCT International Application No. PCT/JP2019/050376 filed on Dec. 23, 2019, designating the United States of America, which is based on and claims priority of Japanese Patent Application No. 2019-003238 filed on Jan. 11, 2019. The entire disclosures of the above-identified applications, including the specifications, drawings and claims are incorporated herein by reference in their entirety.
| Number | Date | Country | |
|---|---|---|---|
| Parent | PCT/JP2019/050376 | Dec 2019 | US |
| Child | 17318513 | US |
