JP7390715B2 - Information processing device, information processing method and program - Google Patents

Description

The present invention relates to an information processing device, an information processing method, and a program.

Conventionally, techniques for identifying objects have been proposed. For example, Patent Document 1 describes a technology that scans random features of an object, converts the scanned random features according to predetermined rules to generate a label, and uses the label to identify the object. has been done.

US Patent No. 7,577,844

It is thought that the characteristics of an object may not necessarily be measurable. With the technique described in Patent Document 1, if an object has a feature that cannot be scanned, it is considered that the object cannot be identified based on the feature.

Therefore, the present invention provides an information processing device, an information processing method, and a program that make it possible to identify an object based on the characteristic even if the object has a characteristic that is difficult to measure. The purpose is to

An information processing device according to one aspect of the present invention inputs m-dimensional (m: a natural number of 2 or more) state data regarding the state of an object, and generates an n-dimensional (n: a natural number, m>n) first vector. Regarding the first network that outputs and the second network that receives data based on the first vector as input and outputs m-dimensional data, the second network outputs m-dimensional data that restores state data. , a first learning unit that learns the first network and the second network, and receives p-dimensional (p: a natural number of 2 or more, p>n) physical data of the object as input and outputs an n-dimensional second vector. Regarding the third network and the trained second network trained by the first learning unit that receives data based on the second vector as input, the trained second network and a second learning unit that learns the third network so as to output estimation data for estimating the state.

According to this aspect, even if the characteristics included in the state of the object are difficult to measure, the state of the object can be estimated by using the physical data of the object. This estimation allows objects to be identified based on characteristics that are difficult to measure.

In the above aspect, using the trained third network trained by the second learning unit and the trained second network inputting data based on the second vector outputted by the trained third network, The device may further include an estimation unit that generates estimated data using the trained second network by inputting physical data to the trained third network.

According to this aspect, estimation data for estimating the state of the object can be obtained. By using this estimated data, it becomes possible to identify the target object.

The above aspect may further include an identification unit that generates an identifier for the object based on the estimation data generated by the estimation unit.

According to this aspect, by using the generated identifier, it becomes possible to more easily identify the target object.

In the above aspect, the state data may be data in which values corresponding to the state of the object are associated with each of m positions in a two-dimensional space or a three-dimensional space.

According to this aspect, the state data corresponds to the state of the object in two-dimensional space or three-dimensional space. The estimation data is data for estimating the state of the object in the two-dimensional space, similar to the state data. Such estimated data is unlikely to lose its function as an identifier even if the state of the object changes over time, making it possible to identify the object over a longer period of time.

In the above aspect, the physical data may be data based on a physical phenomenon caused by wave interference in the target object.

According to this aspect, the state of the object is reflected in the data based on the physical phenomenon caused by wave interference. Therefore, by using data based on physical phenomena caused by wave interference, the state of the object can be estimated more accurately. As a result, it becomes possible to identify the target object with higher accuracy.

In the above aspect, the physical data may be data expressing the magnetic field dependence of the electrical conduction of the object in a two-dimensional graph of the magnetic field applied to the object and the electrical conduction of the object.

According to this aspect, data based on nanoscale phenomena can be used as physical data. Therefore, even if the features of the object are minute, it is possible to identify the object.

In the above aspect, the physical data may be data that expresses a speckle pattern generated when the object is irradiated with light as a two-dimensional image.

According to this aspect, physical data can be acquired by irradiating the object with light. Therefore, it becomes possible to identify the target object more easily.

An information processing method according to another aspect of the present invention receives m-dimensional (m: a natural number of 2 or more) state data regarding the state of an object as input, and generates an n-dimensional (n: a natural number, m>n) first vector. and a second network that receives data based on the first vector and outputs m-dimensional data, such that the second network outputs m-dimensional data that restores the state data. The first network and the second network are trained, and the third network receives p-dimensional (p: a natural number of 2 or more, p>n) physical data of the object as input and outputs an n-dimensional second vector. network and a trained second network that receives data based on the second vector as input, such that the trained second network outputs estimation data that estimates the state of the object based on the second vector. and learning a third network.

According to this aspect, even if the characteristics included in the state of the object are difficult to measure, the state of the object can be estimated by using the physical data of the object. This estimation allows objects to be identified based on characteristics that are difficult to measure.

A program according to another aspect of the present invention inputs m-dimensional (m: a natural number of 2 or more) state data regarding the state of an object to a computer, and inputs an n-dimensional (n: a natural number, m>n) first Regarding a first network that outputs a vector and a second network that receives data based on the first vector and outputs m-dimensional data, the second network outputs m-dimensional data that restores state data. As shown in FIG. 3 network and a trained second network that receives data based on the second vector as input, the trained second network outputs estimation data that estimates the state of the object based on the second vector. and learning the third network as follows.

According to this aspect, even if the characteristics included in the state of the object are difficult to measure, the state of the object can be estimated by using the physical data of the object. This estimation allows objects to be identified based on characteristics that are difficult to measure.

According to the present invention, it is possible to provide an information processing device, an information processing method, and a program that make it possible to identify an object having characteristics that are difficult to measure based on the characteristics.

FIG. 1 is a functional block diagram showing the configuration of an information processing device according to an embodiment of the present invention. FIG. 2 is a schematic diagram showing an example of a configuration of a network learned by a learning unit according to the present embodiment. It is a diagram showing a network formed by connecting a first network and a second network. FIG. 2 is a diagram illustrating an example of state data representing a two-dimensional electron distribution of an electronic device. It is a figure which shows the network formed by connecting the 3rd network and the learned 2nd network. It is a graph showing magnetic field dependence of fluctuation of electric conductivity in an electronic device concerning this embodiment. 1 is a diagram illustrating an example of a hardware configuration of an information processing device according to an embodiment. FIG. 2 is a flowchart illustrating a process in which the information processing apparatus according to the present embodiment learns a network and generates an identifier for an object based on the learned network. FIG. 2 is a diagram showing the structure of a quantum wire produced in this example. FIG. 2 is a diagram showing the results of measuring the magnetic field dependence of fluctuations in the electrical resistance of a sample. FIG. 3 is a diagram showing estimation data for estimating the fine structure of a quantum wire.

Preferred embodiments of the present invention will be described with reference to the accompanying drawings. In addition, in each figure, those with the same reference numerals have the same or similar configurations.

[First embodiment]
FIG. 1 is a functional block diagram showing the configuration of an information processing device 10 according to an embodiment of the present invention. The information processing apparatus 10 according to the present embodiment receives state data of an object as input and generates a network that generates data for restoring the state data. Furthermore, the information processing device 10 according to the present embodiment has a function of learning a network that receives physical data of a target object and generates estimation data for estimating the state of the target object. Furthermore, the information processing device 10 according to the present embodiment has a function of identifying an inspection target based on the generated estimation data. Note that in this specification, "learning a network" includes "learning weighting parameters of a network."

The target object may be any type of object for which physical data can be obtained, but in this embodiment, the target object is a nanodevice having nano-level defects and impurities. Here, it is assumed that the distribution of defects and impurities in a nanodevice cannot be measured without destroying the nanodevice. The condition data is data regarding the condition of the object, and includes data regarding unmeasurable quantities such as the structure, impurities, and defects of the object. In this embodiment, it is assumed that the state data is data representing the distribution of electrons in the electronic device.

Further, the physical data is data based on measurable physical quantities of the object. More specifically, the physical data may be data based on physical phenomena due to wave interference in the object, for example. In this embodiment, the physical data is data based on a physical phenomenon caused by interference of electron waves in a target object, and specifically, data representing the dependence of electrical conductivity on a magnetic field in a nanodevice.

With reference to FIG. 1, the configuration of the information processing device 10 according to the present embodiment will be described in detail. As shown in FIG. 1, the information processing device 10 includes a storage section 100, a learning section 110, an estimation section 120, and an identification section 130.

The storage unit 100 has a function of storing various types of information. The storage unit 100 stores, for example, physical data of the object and information regarding a learned network (for example, network weighting parameters). The various types of information stored in the storage unit 100 are referred to by the learning unit 110 or the estimation unit 120 as needed.

The learning unit 110 has a function of learning various learning models including deep neural networks. The learning unit 110 may learn various networks such as an autoencoder, a variational autoencoder, a generative adversarial network, a fully connected neural network, a convolutional neural network, a recurrent neural network, or a Boltzmann machine. good.

FIG. 2 is a schematic diagram showing an example of the configuration of the network 20 learned by the learning unit 110 according to the present embodiment. The network 20 according to this embodiment includes a first network 210, a second network 220, and a third network 230. In this embodiment, the first network 210 and the second network 220 are convolutional neural networks, and the third network is a fully connected neural network. Note that the first network 210, second network 220, and third network 230 are not limited to the model and structure described in this embodiment.

The output layer 218 of the first network 210 is connected to the input layer 222 of the second network 220 via a representation space 240. Further, the output layer 238 of the third network 230 is connected to the input layer 222 of the second network 220 via the representation space 240.

The method by which the learning unit 110 learns the network 20 will be described in more detail below with reference to FIGS. 3 to 6. FIG. 3 is a diagram showing a network 22 formed by connecting a first network 210 and a second network 220. The network 22 shown in FIG. 3 is a variational autoencoder in which the first network 210 is an encoder and the second network 220 is a decoder. In this embodiment, the first network 210 and the second network 220 are convolutional neural networks (CNN). For example, batch normalization may be applied to the CNN, and a Leaky ReLU function may be used as an activation function. Note that the neural networks forming the first network 210 and the second network 220 are not limited to CNN.

As shown in FIG. 3, the first network 210 includes an input layer 212, a first set of layers 214, a second set of layers 216, and an output layer 218. The input layer 212 is connected to the layer closest to the input of the first set 214, the layer closest to the output of the first set 214 is connected to the layer closest to the input of the second set 216, and the layer closest to the input of the first set 214 is connected to the layer closest to the input of the second set 216. The layer closest to the output of 216 is connected to output layer 218 .

The input layer 212 receives m-dimensional (m: a natural number of 2 or more) state data. The number of dimensions of the input layer 212 may be the same as the number of dimensions of the state data. The number of dimensions of the input layer 212 is not particularly limited, but may be, for example, 3600 dimensions. The input layer 212 receives as input, for example, state data of a 3600-dimensional image having 60 pixels in the vertical direction and 60 pixels in the horizontal direction. Further, in this embodiment, the state data is data in which values corresponding to the state of the object are associated with each of m positions in the two-dimensional space. Specifically, the state data is data representing the distribution of electrons, which associates the electron density of the object with each of m positions in the two-dimensional space.

The first set 214 compresses the state data input to the input layer 212. Although five layers are shown in FIG. 3 as the first set 214, the number of layers in the first set 214 is not particularly limited, and may be, for example, 256 layers. Further, the number of dimensions of each layer of the first set 214 is not particularly limited, but may be, for example, 900 dimensions (30 vertically x 30 horizontally).

The second set 216 further compresses the data of the first set 214. Although five layers are shown in FIG. 3 as the second set 216, the number of layers in the second set 216 is not particularly limited and may be, for example, 512 layers. The number of dimensions of each layer of the second set 216 is not particularly limited, but may be, for example, 225 dimensions (15 vertically x 15 horizontally).

The output layer 218 compresses the data compressed by the second set 216 into an n-dimensional (n: natural number, m>n) first vector and outputs it to the representation space. The number of dimensions of the first vector output by the output layer 218 is not particularly limited, but may be seven dimensions, for example.

The second network 220 includes an input layer 222, a third set 224 of layers, a fourth set 226 of layers, and an output layer 228. The input layer 222 receives data based on a seven-dimensional first vector. Furthermore, data based on reparameterization tricks may be input to the input layer 222.

The third set 224 has a function of generating data by expanding the dimensions of the data input to the input layer 222. Furthermore, the fourth set 226 has a function of expanding the dimensions of the data generated in the third set 224 to generate data. Third set 224 may have a symmetrical configuration with second set 216, and fourth set 226 may have a symmetrical configuration with first set 214, for example.

The output layer 228 outputs m-dimensional data. That is, in this embodiment, the number of dimensions of the output layer 228 of the second network 220 is the same as the number of dimensions of the input layer 212 of the first network 210.

Next, a process in which the learning unit 110 according to this embodiment learns the first network 210 and the second network 220 will be described. The learning unit 110 inputs state data to the input layer 212 of the first network 210. In this embodiment, the state data is data in which values corresponding to the state of the object are associated with each of a plurality of positions in a two-dimensional space. More specifically, in this embodiment, the state data is data representing the two-dimensional electron distribution of the nanodevice.

Note that the state data may be actually measured data or may be data generated by numerical calculation using a known method. For example, when the nanodevice is a nano-sized metal wire having a plurality of circular defects, the electron distribution may be calculated using a known method to generate state data of the electron distribution in the metal wire. Alternatively, a database may be used to generate state data. For example, state data may be generated using a database that associates metal wire information (for example, the shape of the metal wire, the number of defects, the shape of the defects, or the position of the defects) with the electron distribution of the metal wire. good.

FIG. 4 shows an example of state data 300 representing a two-dimensional electron distribution of a nanodevice. In FIG. 4, the whiter the color, the higher the electron density, and the darker the color, the lower the electron density. Two circular defects (holes) are formed in the nanodevice according to this embodiment. Therefore, in the state data 300, circular first defects 302 and second defects 304 in which no electrons are present are formed. In this embodiment, state data representing the electron distribution of the nanodevice in which circular defects are formed is input to the input layer 212 of the first network 210.

The learning unit 110 inputs state data to the input layer 212 of the first network 210 and controls the first network 210 and the second network 220 so that the output layer 228 of the second network 220 outputs data for restoring the state data. Learn. This makes it possible, for example, to compress 3600-dimensional state data into a 7-dimensional first vector and generate data for restoring the state data from the first vector. In other words, 3600-dimensional state data can be expressed with 7-dimensional data. After learning the first network 210 and the second network 220, the learning unit 110 causes the storage unit 100 to store the weighting parameters of these networks.

FIG. 5 is a diagram showing a network 24 formed by connecting the third network 230 and the learned second network 220 learned by the method described with reference to FIG. 3. The learning unit 110 uses the network 24 shown in FIG. 5 to learn the third network 230. Here, the configuration of the learned second network 220 is substantially the same as the configuration of the second network 220 described with reference to FIG. 3, so the description thereof will be omitted here.

The third network 230 includes an input layer 232, a first intermediate layer 234, a second intermediate layer 236, and an output layer 238. In this embodiment, the third network 230 is a fully connected network. For example, dropout may be applied to the fully connected network. Further, the activation function applied to the fully connected network is not particularly limited, but for example, a ReLU function may be applied.

P-dimensional (p: a natural number of 2 or more, p>n) physical data is input to the input layer of the third network 230. The number of dimensions of the physical data is the same as the number of dimensions of the input layer 232. The number of dimensions of the input layer 232 of the third network 230 is not particularly limited, but may be, for example, 100 dimensions. Furthermore, the number of layers of the input layer 232 is not limited to one layer, and may be composed of a plurality of layers, for example, may be two layers. At this time, the number of dimensions of the second layer of the input layer 232 may be, for example, 8192 dimensions. Further, the number of dimensions of the first intermediate layer 234, the second intermediate layer 236, and the output layer 238 is not particularly limited, but may be, for example, 8192 dimensions.

Next, a method for learning the third network 230 by the learning unit 110 using the network 24 shown in FIG. 5 will be described. The learning unit 110 inputs physical data to the input layer 232 of the third network 230. In this embodiment, the learning unit 110 inputs physical data representing the magnetic field dependence of the electrical conductivity of the nanodevice to the input layer 232.
The electronic device according to this embodiment is a quantum wire having circular holes (hereinafter also referred to as "antidots"), and for each of a plurality of quantum wire models in which the antidots are arranged differently, The magnetic field dependence of electrical conductivity (physical data) is prepared. In the present embodiment, the learning unit 110 subtracts the average of all prepared physical data from the magnetic field dependence of electrical conductivity, rather than the magnetic field dependence of electrical conductivity itself, and uses the result obtained by subtracting the result from the third network 230. and train the network. This improves the learning efficiency compared to the case where the network is trained using the magnetic field dependence of electrical conductivity itself.

FIG. 6 is a graph showing the magnetic field dependence of fluctuations in electrical conductivity in the electronic device according to the present embodiment. In FIG. 6, the horizontal axis represents the magnetic field, and the vertical axis represents fluctuations in electrical conductivity. The electrical conductivity fluctuation shown in FIG. 6 is a graph obtained by subtracting the average value of all physical data from the calculated magnetic field dependence of electrical conductivity. The number of points on the plotted graph may be, for example, about 100 points.

In this embodiment, it is assumed that the magnetic field dependence of electrical conductivity is measured at an extremely low temperature of about several tens of mK. This makes it possible to measure universal conductance fluctuations (UCF) that appear as magnetoresistive effects in mesoscopic systems.

The UCF is determined by the shape of the sample containing impurities. Therefore, the UCF can be calculated numerically based on the shape of the sample. However, the shape of the sample cannot be calculated from the magnetic field dependence of the measured electrical resistance. In this embodiment, by machine learning the relationship between the UCF pattern (magnetic field dependence of electrical resistance) and state data regarding the shape of the sample, we generate estimation data for estimating the shape of the sample from the measured values of the UCF. It is possible to generate. Note that the physical data regarding the magnetic field dependence of electrical conductivity or state data regarding the shape of the sample used for learning may be data obtained through actual measurement, or may be calculated using known techniques. It may also be data obtained by

When the physical data of the graph shown in FIG. 6 is input to the input layer 232 of the third network 230, an n-dimensional second vector is output from the output layer 238. The output layer 238 may output, for example, a seven-dimensional second vector to the representation space 240. When the output second vector is input to the input layer 222 of the trained second network 220, m-dimensional data is output from the output layer 228.

The learning unit 110 trains the third network 230 so that the trained second network 220 generates estimation data for estimating the state of the object based on the second vector output by the third network 230. Specifically, the third network is trained using physical data and state data corresponding to the physical data as teacher data. The physical data may be data obtained by measuring physical quantities of the object, and may be, for example, data on magnetic field dependence of electrical conductivity. Further, the state data may be the state of the object (for example, data representing the fine structure of a nanodevice) for which the physical data was measured.

In the present embodiment, the learning unit 110 trains the third network 230 so that the trained second network 220 generates estimation data for estimating the state of the object. More specifically, the learning unit 110 trains the third network 230 so that the estimated data generated by the trained second network 220 approaches the state data corresponding to the physical data input to the third network 230. learn. After completing the learning, the learning unit 110 causes the storage unit 100 to store information regarding the learned weighting parameters of the third network 230 and the like.

In this embodiment, the physical data input to the third network 230 is compressed into seven-dimensional data, and estimated data is generated based on the compressed data. In this way, by once compressing the dimensions of the physical data, it is possible to reduce the number of parameters required for learning, and the third network 230 can be trained at a higher speed.

Returning to FIG. 1, the functions of the information processing device 10 will be explained. The estimation unit 120 has a function of generating estimated data using the trained third network and the trained second network 220. More specifically, the estimation unit 120 inputs the physical data of the object to the trained third network 230 and causes the trained second network 220 to output estimated data. The estimation unit 120 transmits the generated estimation data to the identification unit 130.

The identification unit 130 has a function of identifying a target object based on the estimation data generated by the estimation unit 120. For example, the identification unit 130 may identify the object by comparing the state data of the object to be identified with the generated estimation data.

The identification unit 130 can also generate an object identifier based on the estimation data. The identification unit 130 may generate the estimated data itself as an identifier, or may generate an identifier by converting the estimated data using a known hash function or the like.

FIG. 7 is a diagram showing an example of the hardware configuration of the information processing device 10 according to the present embodiment. The information processing device 10 includes a processor 10a such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit) that corresponds to an arithmetic unit, a RAM (Random Access Memory) 10b that corresponds to a storage unit 100, and a RAM (Random Access Memory) 10b that corresponds to a storage unit 100. It has a ROM (Read Only Memory) 10c, a communication section 10d, an input section 10e, and a display section 10f. These components are connected to each other via a bus so that they can transmit and receive data. Note that in this example, a case will be described in which the information processing device 10 is composed of one computer, but the information processing device 10 may be realized by combining a plurality of computers. Further, the configuration shown in FIG. 7 is an example, and the information processing device 10 may have configurations other than these, or may not have some of these configurations. Here, the calculation section includes a learning section 110, an estimation section 120, and an identification section 130.

The processor 10a is a control unit that controls the execution of programs stored in the RAM 10b or ROM 10c, and performs calculations and processing of data. The processor 10a is a calculation unit that executes a program for learning a network and performing processing using the learned network. The processor 10a receives various data from the input section 10e and the communication section 10d, and displays the calculation results of the data on the display section 10f or stores them in the RAM 10b.

The RAM 10b is a storage unit in which data can be rewritten, and may be formed of, for example, a semiconductor storage element. The RAM 10b may store data such as programs executed by the processor 10a and training data for learning. Note that these are just examples, and the RAM 10b may store data other than these, or some of them may not be stored.

The ROM 10c is a storage section from which data can be read, and may be composed of, for example, a semiconductor storage element. The ROM 10c may store, for example, programs and data that cannot be rewritten.

The communication unit 10d is an interface that connects the information processing device 10 to other devices. The communication unit 10d may be connected to a communication network such as the Internet.

The input unit 10e receives data input from the user, and may include, for example, a keyboard and a touch panel.

The display unit 10f visually displays the calculation results by the processor 10a, and may be configured by, for example, an LCD (Liquid Crystal Display). The display unit 10f may display the generated estimated data and the like.

The program may be provided by being stored in a computer-readable storage medium such as the RAM 10b or ROM 10c, or may be provided via a communication network connected by the communication unit 10d. In the information processing device 10, the various operations described using FIG. 1 are realized by the processor 10a executing programs. Note that these physical configurations are merely examples, and do not necessarily have to be independent configurations. For example, the information processing device 10 may include an LSI (Large-Scale Integration) in which a processor 10a, a RAM 10b, and a ROM 10c are integrated.

FIG. 8 is a flowchart showing a process in which the information processing apparatus 10 according to the present embodiment learns a network and generates an object identifier using the learned network. Processing by the information processing apparatus 10 according to this embodiment will be described below with reference to FIG. 8.

First, the learning unit 110 of the information processing device 10 learns the first network 210 and the second network 220 (step S101). Specifically, as described with reference to FIG. 3, the learning unit 110 operates so that the second network 220 generates m-dimensional data that restores the state data input to the first network 210. The first network 210 and the second network 220 are learned. The learned information regarding the first network 210 and the second network 220 is stored in the storage unit 100.

Next, the learning unit 110 learns the third network 230 (step S103). More specifically, as described with reference to FIG. 5, the learning unit 110 trains the third network 230 so that the trained second network 220 generates estimation data for estimating the state of the object. learn. Information regarding the learned third network is stored in the storage unit 100.

Next, the estimation unit 120 generates estimation data (step S105). More specifically, in a state where the learned third network 230 learned in step S103 and the learned second network 220 learned in step S101 are connected, the estimation unit 120 Physical data is input to the input layer of the 3rd network 230, and the learned second network 220 is made to output estimated data.

Next, the identification unit 130 generates an identifier for the object (step S107). More specifically, the identification unit 130 generates an identifier for the object based on the estimated data generated in step S105.

In this manner, the information processing apparatus 10 according to the present embodiment can generate an identifier that can be considered a fingerprint unique to an object (for example, an electronic device). The generated identifier can also be used for data encryption and user authentication. Furthermore, the generated identifier can also be applied to various security-related techniques.

Furthermore, in this embodiment, even if the state of the object (electron distribution, etc.) cannot be calculated from physical data, an identifier for the object is generated based on estimation data that estimates the state of the object from physical data. be able to.

BACKGROUND ART Conventionally, PUF (Physically Unclonable Function) has been proposed as a method of generating a device identifier using a physical phenomenon unique to an object. The PUF includes a method that uses initialization information of SRAM (Static Random Access Memory) or a delay time specific to the circuit, or a method that uses random fluctuations in the operating current of a transistor such as random telegraph noise. All of these methods utilize physical quantities that change over time due to environmental changes or equipment deterioration, making it difficult to maintain a unique identifier over a long period of time.

The information processing device 10 according to the present embodiment can generate estimated data for identifying a target object using physical data based on physical characteristics that change over time, and can generate an identifier unique to the target object. In this embodiment, even if the physical property of magnetic field dependence of electrical conductivity changes over time, the structure of the quantum wire is maintained to the extent that the object can be identified. Therefore, a robust identifier can be created based on physical data that changes over time. It is possible to generate Therefore, it is possible to generate an identifier that allows the object to be identified over a long period of time.

With the recent evolution of IoT (Internet of Things), the need for security technology to identify devices such as PUF is increasing. In IoT, devices can be used in a near-disposable manner, so it is important that devices be provided at low prices. On the other hand, the present invention is preferably applied to highly rare devices for which authenticity is essentially important. The present invention can be used, for example, to identify the quantum computer that performed the quantum calculation in quantum calculations where a single processing cost is high, and to ensure the reliability of the distributor when performing quantum key distribution. It can be used for.

The technique of the present invention can also be applied when an object has characteristics that are difficult to measure. For example, if the internal structure of an object is a characteristic of the object, the object may have to be destroyed in order to measure the characteristic. In order to detect impurities or defects in a target element, for example, it is necessary to slice the element and observe the cross-sectional structure. As described above, if impurities or defects inside the element cannot be detected without destroying the object, the object cannot be an individually identifiable device. For this reason, the present invention, which is capable of estimating the internal state of an object, is important.
According to the information processing device 10 according to the present embodiment, by measuring physical data such as the electrical conductivity of the object, it is possible to estimate estimation data including the characteristics (for example, internal structure) of the object. By using this estimated data, it becomes possible to generate an object identifier based on features that are difficult to measure.

Furthermore, according to the present embodiment, by measuring the physical data of the object using a non-destructive method, it is possible to generate estimated data including the characteristics of the object, and to generate an identifier for the object. Therefore, even if an object has characteristics that cannot be measured without destroying it, an identifier can be generated or the object can be identified without destroying the object.

(Example)
An example of identifying a nanodevice sample using the above-described method will be described. In this example, a quantum wire sample was produced by the lift-off method, and estimated data was generated for the produced sample by the method described above to identify the sample.

First, a trained first network and a trained second network were obtained using the method described with reference to FIG. Specifically, the first network and the second network are connected so that image data of a quantum wire in which two holes are formed is input to the first network, and data for restoring the image data is output from the second network. I learned that. The dimensions of the image data used in this example are 3600 dimensions, 60 pixels vertically and 60 pixels horizontally. By learning the first network and the second network, 3600-dimensional image data can be expressed as a 7-dimensional vector. Note that, since the feature of the quantum wire of this example is the position of the two holes, it is sufficient that the dimension of the vector expressing the feature is seven dimensions. Note that it is also possible to change the dimension of the vector depending on conditions such as the number of holes.

In this example, the dependence of electrical conductivity on the magnetic field (physical data) was calculated and prepared for each of a plurality of quantum wire models having different antidot arrangements. Preferably, normalized data is input to the network in order to improve learning efficiency. In this example, physical data was obtained by calculating the magnetic field dependence of electrical resistance for each of a plurality of models (quantum wire models) with different anti-dot arrangements. Next, for each of the calculated physical data, the normalized magnetic field dependence of the electrical resistance was obtained by subtracting the average of all physical data. The third network was trained using physical data representing the magnetic field dependence of the normalized electrical resistance and image data representing the fine structure of the quantum wire as training data. As a result, by inputting physical data into the trained third network, a network was obtained that can output estimation data for estimating the fine structure of the quantum wire from the trained second network.

The lift-off method used to prepare the sample will be explained. An organic substance (resist) was coated on a silicon substrate, and bonds between polymers of the organic substance were partially decomposed by an electron beam using an electron beam lithography device, and the organic substance in a desired pattern portion was peeled off. A metal containing titanium and gold was deposited on a silicon substrate on which a desired pattern was formed, and the resist was removed to obtain a structure in which metal was deposited in a desired pattern. In this way, a quantum wire with a width of about 280 nm was produced.

Thereafter, spot firing was performed at two points on the produced quantum wire using an electron beam drawing device to form circular holes (hereinafter also referred to as "anti-dots"). Here, spot firing is a method of irradiating an electron beam in a concentrated manner at one point. In this example, antidots with a diameter of about 50 nm were formed.

FIG. 9 is a diagram showing the structure of the quantum wire 310 produced in this example. A first antidot 312 and a second antidot 314 are formed in the quantum wire 310 . Regarding these two antidots, the distance in the length direction of the quantum wires 310 was set to about 180 nm.

In this example, in order to observe fluctuations in universal conductance based on quantum interference effects in quantum wires, a dilution refrigerator was used to cool the sample to a temperature around 25 mK. Around this temperature, the electrical resistance R of the sample was measured while changing the magnetic field B between 0 and 4 (T).

FIG. 10 is a diagram showing the results of measuring the magnetic field dependence of the fluctuation δR of the electrical resistance of the sample. The fluctuation δR of the electrical resistance is a value obtained by linearly fitting the magnetic field dependence of the measured electrical resistance to obtain a linear function, and subtracting the linear function from the magnetic field dependence of the electrical resistance. FIG. 10 shows the results of measurements while increasing the magnetic field. Even when estimating the state of an object using actual measurement results, it is preferable that the data input to the network be normalized. When estimating the state, in this example, normalized data was obtained by subtracting a linear function from the measured data. Using this normalized data as input, estimated data was obtained using a trained second network and a trained third network.

FIG. 11 is a diagram showing estimation data 320 for estimating the fine structure of a quantum wire. As shown in FIG. 11, a first estimated dot 322 and a second estimated dot 324 are shown. The first estimated dot 322 and the second estimated dot 324 correspond to the first anti-dot 312 or the second anti-dot 314 shown in FIG. 9, respectively. As shown in FIG. 11, in particular, the first estimated dot 322 is clearly shown. In this way, in this example, the structure of the quantum wire 310, specifically, the position of the antidot formed on the quantum wire 310, could be estimated based on the magnetic field dependence of the electrical resistance of the quantum wire. By using this estimation result, quantum wires can be identified.

[Second embodiment]
In the second embodiment, descriptions of points that are substantially the same as those of the first embodiment will be omitted, and points that are different from the first embodiment will be mainly described.

In the second embodiment, the physical data is data that expresses a speckle pattern generated when a target object is irradiated with light as a two-dimensional image. A method for acquiring this physical data will be briefly explained. First, a laser beam is used to irradiate a target object, for example, a flat material. Physical data is image data of a transmitted light speckle pattern generated when the irradiated light passes through the object or a reflected light speckle pattern generated when the irradiated light is reflected by the object.

Also in the second embodiment, as described with reference to FIG. 3, the first network 210 and the second network 220 are trained by inputting physical data to the first network 210. Next, as described with reference to FIG. 5, the third network 230 is trained so that estimation data for estimating the state of the object is generated from the trained second network 220. The estimated data may be data regarding the structure of the surface of the object (for example, distribution of unevenness, etc.), data regarding the distribution of impurities inside the object, or data regarding the elements constituting the object. It may also be data regarding the distribution of .

In this way, in the second embodiment, the identifier of the object can be generated based on the physical data of the optical speckle pattern obtained by irradiating the object with light. Since easily obtained physical data can be used, it becomes possible to generate an object identifier more easily.

An application example of the method according to this embodiment will be described. For example, a tag for identification may be attached to an article. The tag can be irradiated with a laser, estimated data can be estimated using the physical data of the obtained speckle pattern, and the tag can be identified using the estimated data. Based on the identification result, the tagged article can be identified.

The embodiments described above are intended to facilitate understanding of the present invention, and are not intended to be interpreted as limiting the present invention. Each element included in the embodiment, its arrangement, material, conditions, shape, size, etc. are not limited to those illustrated, and can be changed as appropriate. Further, it is possible to partially replace or combine the structures shown in different embodiments.

In the first embodiment described above, the physical data is mainly data regarding the magnetic field dependence of the electrical conductivity of the object, but the physical data is not limited to this. The physical data may be, for example, data regarding measurable electrical, thermal, or magnetic properties of the object.

In the first embodiment described above, the state data is mainly data regarding the electron distribution of the electronic device, but the state data is not limited to this. The condition data may be, for example, data regarding the distribution of nanoscale impurities in the material. In this case, a Variational AutoEncoder or a Generative Adversarial Network may be used in the process corresponding to step S101.

Further, in the above embodiment, the description has been made mainly on the assumption that the state data represents the object in a two-dimensional space. The state data is not limited to this, and may represent a high-dimensional space of three or more dimensions. Even when the state data represents a three-dimensional space, the state data may be data in which values corresponding to the state of the object are associated with each of a plurality of positions. State data representing a three-dimensional space is used, for example, when estimating a three-dimensional state of a substance such as a three-dimensional structure.

10... Information processing device, 10d... Communication unit, 10e... Input unit, 10f... Display unit, 100... Storage unit, 110... Learning unit, 120... Estimating unit, 130... Identification unit, 210... First network, 212... Input Layer, 214...first set, 216...second set, 218...output layer, 220...second network, 222...input layer, 224...third set, 226...fourth set, 228...output layer, 230...th 3 network, 232... input layer, 234... first intermediate layer, 236... second intermediate layer, 238... output layer, 240... representation space, 300... state data, 302... first defect, 304... second defect, 310 ...Quantum wire, 312...First antidot, 314...Second antidot, 320...Estimation data, 322...First estimation dot, 324...Second estimation dot

Claims (9)

a first network that receives m-dimensional (m: a natural number of 2 or more) state data regarding the state of an object and outputs an n-dimensional (n: a natural number, m>n) first vector; and a second network that inputs data based on the data and outputs m-dimensional data,
a first learning unit that learns the first network and the second network so that the second network outputs m-dimensional data that restores the state data;
A third network that receives p-dimensional (p: a natural number of 2 or more, p>n) physical data of the object and outputs an n-dimensional second vector, and receives data based on the second vector as input. , and the learned second network learned by the first learning unit,
a second learning unit that learns the third network so that the trained second network outputs estimation data for estimating the state of the object based on the second vector;
An information processing device comprising: Using a learned third network learned by the second learning unit and the learned second network inputting data based on the second vector outputted by the learned third network, an estimating unit that generates the estimated data by the trained second network by inputting the physical data to the trained third network;
The information processing device according to claim 1, further comprising: an identification unit that generates an identifier for the object based on the estimation data generated by the estimation unit;
The information processing device according to claim 2, further comprising: The state data is data in which values corresponding to the state of the object are associated with each of m positions in a two-dimensional space or a three-dimensional space.
The information processing device according to any one of claims 1 to 3. The physical data is data based on a physical phenomenon caused by wave interference in the target object,
The information processing device according to any one of claims 1 to 4. The physical data is data that expresses the magnetic field dependence of the electrical conduction of the object as a two-dimensional graph of the magnetic field applied to the object and the electrical conduction of the object.
The information processing device according to claim 5. The physical data is data that expresses a speckle pattern generated when the object is irradiated with light as a two-dimensional image.
The information processing device according to claim 5. a first network that receives m-dimensional (m: a natural number of 2 or more) state data regarding the state of an object and outputs an n-dimensional (n: a natural number, m>n) first vector; and a second network that inputs data based on the data and outputs m-dimensional data,
Learning the first network and the second network so that the second network outputs m-dimensional data that restores the state data;
A third network that receives p-dimensional (p: a natural number of 2 or more, p>n) physical data of the object and outputs an n-dimensional second vector, and receives data based on the second vector as input. , the trained second network, and
Learning the third network so that the trained second network outputs estimation data for estimating the state of the object based on the second vector;
information processing methods, including to the computer,
a first network that receives m-dimensional (m: a natural number of 2 or more) state data regarding the state of an object and outputs an n-dimensional (n: a natural number, m>n) first vector; and a second network that inputs data based on the data and outputs m-dimensional data,
Learning the first network and the second network so that the second network outputs m-dimensional data that restores the state data;
A third network that receives p-dimensional (p: a natural number of 2 or more, p>n) physical data of the object and outputs an n-dimensional second vector, and receives data based on the second vector as input. , the trained second network, and
Learning the third network so that the trained second network outputs estimation data for estimating the state of the object based on the second vector;
A program to run.