Concealment apparatus, estimation device, display device, estimation system, concealment method, estimation method, and program

Abstract

To enable an input character string to a system for outputting a response in the form of a character string to input of a character string to be the character string in which information subject to concealment is concealed.SOLUTION: The concealment apparatus is provided with concealment means for generating a second character string in which a partial character string detected as a concealment object in a first character string is replaced by another partial character string, and concealment prompt issuing means for inputting the second character string to a device for outputting a response to the second character string.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a concealment device, an estimation device, a display device, an estimation system, a concealment method, an estimation method, and a program. [Background technology]

[0002] There are cases where a system that outputs a string response to a string input, such as a large-scale language model, is used. For example, Patent Document 1 describes a sentence generation device that generates a prompt for inputting a number of characters to a large-scale language model that does not exceed the limit on the number of characters that can be input to the large-scale language model. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 7313757 Summary of the Invention [Problem to be solved by the invention]

[0004] It is preferable that an input string to a system that outputs a response in the form of a string in response to an input of a string can be a string in which the information to be concealed is concealed.

[0005] An example of an object of the present invention is to provide a concealment device, an estimation device, a display device, an estimation system, a concealment method, an estimation method, and a program that can solve the above-mentioned problems. [Means for solving the problem]

[0006] According to a first aspect of the present invention, an anonymization device includes an anonymization means for generating a second string in which a substring detected as a target for anonymization in a first string is replaced with another substring, and an anonymization prompt issuing means for inputting the second string to a device that outputs a response to the second string.

[0007] According to a second aspect of the present invention, an estimation device includes an estimation means for estimating a fourth string that is a response to a first string in which a partial string detected as a target for anonymization has not been replaced, based on a third string that is a response to a second string in which a partial string detected as a target for anonymization of a first string has been replaced with another partial string.

[0008] According to a third aspect of the present invention, a display device includes a display means for displaying a second character string in which a partial character string to be concealed in a first character string has been replaced with another partial character string.

[0009] According to a fourth aspect of the present invention, a display device includes a display means for displaying a fourth string, which is a response to a first string in which a substring of a first string that is a target of anonymity has been replaced with another substring, the fourth string being a response to the first string in which the substring of the first string that is a target of anonymity has not been replaced.

[0010] According to a fifth aspect of the present invention, an estimation system includes an anonymization means for generating a second string in which a partial string detected as a target for anonymization in a first string is replaced with another partial string, and an estimation means for estimating a fourth string that is a response to the first string without replacement of the partial string detected as a target for anonymization, based on a third string that is a response to the second string.

[0011] According to a sixth aspect of the present invention, a method for anonymization includes a computer generating a second string in which a substring of a first string detected as a target for anonymization is replaced with another substring, and inputting the second string to a device that outputs a response to the second string.

[0012] According to a seventh aspect of the present invention, an estimation method includes a computer estimating, based on a third string provided in response to a second string in which a substring of a first string detected as a target for anonymization has been replaced with another substring, a fourth string that is a response to the first string in which the substring detected as a target for anonymization has not been replaced.

[0013] According to an eighth aspect of the present invention, an estimation method includes a computer generating a second string in which a substring of a first string detected as a target for anonymization is replaced with another substring, and estimating, based on a third string responded to the second string, a fourth string that is a response to the first string without replacement of the substring detected as a target for anonymization.

[0014] According to a ninth aspect of the present invention, a program causes a computer to generate a second string in which a substring of a first string detected as a target for anonymization is replaced with another substring, and input the second string to a device that outputs a response to the second string.

[0015] According to a tenth aspect of the present invention, a program causes a computer to execute the following steps: based on a third string returned in response to a second string in which a partial string detected as a target of anonymization in a first string has been replaced with another partial string, a fourth string being a response to the first string in which the partial string detected as a target of anonymization has not been replaced. Effect of the Invention

[0016] According to the present invention, a character string input to a system that outputs a character string response in response to an input of a character string can be a character string in which information that is the subject of anonymity is concealed. [Brief description of the drawings]

[0017] [Figure 1] FIG. 1 illustrates an example of a system configuration according to at least one embodiment. [Diagram 2] FIG. 2 is a diagram illustrating an example of a procedure of a process performed by the estimation system according to at least one embodiment. [Diagram 3] FIG. 1 is a diagram illustrating an example of the configuration of a concealment device according to at least one embodiment. [Figure 4] FIG. 1 is a diagram illustrating an example of a configuration of an estimation device according to at least one embodiment. [Diagram 5] 1A to 1C are diagrams illustrating an example of input and output of data when an estimation system according to at least one embodiment estimates an ideal response by replacing a character string. [Figure 6] FIG. 2 is a diagram illustrating an example of the configuration of a learning system when training a neural network included in an estimation unit according to at least one embodiment. [Figure 7] FIG. 1 is a diagram illustrating an example of a configuration of a learning control device according to at least one embodiment. [Figure 8] FIG. 2 is a diagram illustrating an example of input and output of data in a learning system when training a neural network included in an estimation unit according to at least one embodiment. [Figure 9] 4 is a diagram illustrating an example of input and output of data in an estimation system when an estimator according to at least one embodiment estimates an ideal response. FIG. [Figure 10] FIG. 2 is a diagram illustrating an example of the configuration of a learning system when additional learning is performed on a response device according to at least one embodiment. [Figure 11] FIG. 2 is a diagram showing an example of input and output of data in a learning system when additional learning is performed on a response device according to at least one embodiment. [Figure 12] FIG. 13 is a diagram illustrating another example of the configuration of the concealment device according to at least one embodiment. [Figure 13] FIG. 13 is a diagram illustrating another example of the configuration of an estimation device according to at least one embodiment. [Figure 14] FIG. 1 is a diagram illustrating an example of a configuration of a display device according to at least one embodiment. [Figure 15] FIG. 2 is a diagram showing another example of the configuration of a display device according to at least one embodiment. [Figure 16] FIG. 1 is a diagram illustrating an example of a configuration of an estimation system according to at least one embodiment. [Figure 17] FIG. 2 is a diagram illustrating an example of a processing procedure in a concealment method according to at least one embodiment. [Figure 18] FIG. 4 is a diagram illustrating an example of a processing procedure in an estimation method according to at least one embodiment. [Figure 19] FIG. 11 is a diagram illustrating another example of a processing procedure in the estimation method according to at least one embodiment. [Figure 20] FIG. 1 illustrates an example configuration of a computer in accordance with at least one embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] The following describes embodiments of the present invention, but the following embodiments do not limit the scope of the invention. Furthermore, not all of the combinations of features described in the embodiments are necessarily essential to the solution of the invention. In the following, a letter with a macron may be written with a superscript "-" after the letter. For example, f with a macron is written as f - It can also be written as:

[0019] First Embodiment 1 is a diagram showing an example of a system configuration according to at least one embodiment. In the configuration shown in FIG. 1, an estimation system 1 includes a concealment device 100 and an estimation device 200. Also shown in FIG. 1 is a response device 910 .

[0020] The response device 910 may be configured as a part of the estimation system 1 or may be configured externally to the estimation system 1. In addition, the anonymization device 100, the estimation device 200, and the response device 910, or any one of them may be integrally configured. For example, the anonymization device 100 and the estimation device 200 may be configured as a single device.

[0021] The response device 910 receives an input of a character string and outputs the character string. The response device 910 corresponds to an example of a system that outputs a response by a character string in response to an input of a character string. A character string input to the response device 910, or a character string that is the source of a character string input to the response device 910, is also called a prompt. The character string that is the source of a character string input to the response device 910 here refers to a character string before processing is performed when a certain processing is performed on the character string and the processed character string is input to the response device 910. The character string output by the response device 910 is also called a response. The response device 910 may be configured using a computer.

[0022] The response device 910 may be configured as a Large Language Model (LLM), which is a type of language model. For example, the response device 910 may be configured to output a response in natural language in response to an input of a query in natural language. Furthermore, the input to the response device 910 does not have to be in the form of a query, such as when the response device 910 receives an input of a sentence in natural language and responds with the emotion shown in the sentence. Furthermore, the input to the response device 910, or the output of the response device 910, or both of them, do not have to be limited to a character string in natural language, such as when the response device 910 outputs a program (source code).

[0023] However, the response device 910 is not limited to one configured as a large-scale language model. Various devices that receive an input of a character string and output a character string, and in which the input character string may include a subcharacter string that is the subject of anonymization, can be used as the response device 910.

[0024] The anonymization device 100 performs anonymization on the prompt. Specifically, the anonymization device 100 detects a partial character string to be anonymized in the prompt, and replaces the detected partial character string with another partial character string. The anonymization device 100 then inputs the replaced prompt to the response device 910. The anonymization device 100 may be configured using a computer.

[0025] The prompt after replacement (the prompt in which the substring detected as the target of anonymization is replaced with another substring) is also referred to as an anonymized prompt. The anonymized prompt corresponds to an example of the second string. The anonymized prompt also corresponds to an example of the sixth string. The prompt before replacement (the prompt before the replacement of the substring to be masked) is also referred to as an unmasked prompt. An unmasked prompt is an example of the first string. An unmasked prompt is also an example of the fifth string.

[0026] The estimation device 200 estimates the response of the response device 910 to a non-anonymized prompt based on the response of the response device 910 to the anonymized prompt. The estimation device 200 may be configured using a computer. The response of the response device 910 to the anonymized prompt corresponds to an example of a third string. The response of the response device 910 to the anonymized prompt also corresponds to an example of a seventh string. The response of the response device 910 to the non-anonymized prompt corresponds to an example of a fourth string. The response of the response device 910 to the non-anonymized prompt also corresponds to an example of an eighth string.

[0027] FIG. 2 is a diagram showing an example of a procedure of processing performed by the estimation system 1. 2, the anonymization device 100 acquires a prompt (step S101). The prompt that the anonymization device 100 acquires in step S101 corresponds to a non-anonymized prompt.

[0028] Next, the anonymization device 100 replaces the partial character string that is the target of anonymization in the prompt acquired in step S101 with another partial character string (step S102). Specifically, the anonymization device 100 detects a partial character string that corresponds to the character string designated as the target of anonymization in the prompt acquired in step S101. Then, the anonymization device 100 replaces the detected partial character string with another partial character string. Then, the anonymization device 100 transmits the anonymized prompt to the response device 910 (step S103).

[0029] The response device 910, which has received the anonymized prompt from the anonymization device 100, generates a response to the received anonymized prompt (step S111). Then, the response device 910 transmits the generated response to the estimation device 200 (step S112). The response transmitted by the response device 910 in step S112 corresponds to a response to the anonymized prompt.

[0030] The estimation device 200, which has received a response from the response device 910, estimates a response to the non-anonymized prompt based on the received response (step S121). Then, the estimation device 200 outputs the estimated, non-anonymized response to the prompt (step S122). The estimated response is also referred to as an estimate of the response.

[0031] Fig. 3 is a diagram showing an example of the configuration of the anonymization device 100. In the configuration shown in Fig. 3, the anonymization device 100 includes a first communication unit 110, a first display unit 120, a first operation input unit 130, a first storage unit 180, and a first control unit 190. The first control unit 190 includes a prompt acquisition unit 191, an anonymization unit 192, and an anonymization prompt issuance unit 193.

[0032] The first communication unit 110 communicates with other devices under the control of the first control unit 190. For example, the first communication unit 110 may transmit a concealed prompt to the response device 910. First display unit 120 includes a display screen such as a liquid crystal panel or an LED (Light Emitting Diode) panel, and displays various images under the control of first control unit 190. For example, first display unit 120 may display a non-encrypted prompt and an encrypted prompt. The first display unit corresponds to an example of a display means. The anonymization device 100 corresponds to an example of a display device in that it includes the first display unit 120.

[0033] The first operation input unit 130 includes input devices such as a keyboard and a mouse, and receives a user operation. For example, the first operation input unit 130 may receive a user operation of inputting a prompt. The prompt here corresponds to a non-anonymized prompt.

[0034] The first storage unit 180 stores various data. The first storage unit 180 is configured using a storage device included in the anonymization device 100. For example, the first storage unit 180 may store data for detecting a partial character string that is to be concealed.

[0035] The first storage unit 180 may store a list of character strings to be detected as substrings to be anonymized as data for detecting substrings to be anonymized. A character string to be detected as a substring to be anonymized is also referred to as a replacement source character string.

[0036] Alternatively, the first storage unit 180 may store, as data for detecting a substring to be concealed, a replacement rule indicating a source string and a string to replace the source string. The string to replace the source string is also referred to as a replacement string.

[0037] The first control unit 190 performs various processes by controlling each unit of the anonymization device 100. The functions of the first control unit 190 may be performed by a CPU (Central Processing Unit) included in the anonymization device 100 reading a program from the first storage unit 180 and executing the program.

[0038] The prompt acquisition unit 191 acquires a prompt. The prompt acquired by the prompt acquisition unit 191 corresponds to a non-anonymized prompt. The method by which the prompt acquisition unit 191 acquires a prompt is not limited to a specific method. For example, the first operation input unit 130 may receive a user operation to input a prompt, and the prompt acquisition unit 191 may generate a character string according to the user operation to acquire the prompt. Alternatively, the first communication unit 110 may receive a prompt from another device, and the prompt acquisition unit 191 may extract the prompt from the data received by the first communication unit 110.

[0039] The anonymization unit 192 anonymizes the prompt. Specifically, the anonymization unit 192 detects a partial character string that is to be anonymized from the prompt acquired by the prompt acquisition unit 191. Then, the anonymization unit 192 replaces the detected partial character string with another partial character string. The anonymization unit 192 corresponds to an example of an anonymization means. The prompt replaced by the anonymization unit 192 corresponds to an anonymized prompt.

[0040] The concealed prompt issuing unit 193 transmits the prompt replaced by the concealment unit 192 to the response device 910 via the first communication unit 110 .

[0041] Fig. 4 is a diagram showing an example of the configuration of the estimation device 200. In the configuration shown in Fig. 4, the estimation device 200 includes a second communication unit 210, a second display unit 220, a second operation input unit 230, a second storage unit 280, and a second control unit 290. The second control unit 290 includes a response acquisition unit 291, an estimation unit 292, and an ideal response output unit 293.

[0042] The second communication unit 210 communicates with other devices under the control of the second control unit 290. For example, the second communication unit 210 may receive, from the response device 910, a response to a concealed prompt. The second display unit 220 has a display screen such as a liquid crystal panel or an LED panel, and displays various images under the control of the second control unit 290. For example, the second display unit 220 may display responses to anonymized prompts and estimates of responses to non-anonymized prompts. The second display unit 220 corresponds to an example of a display means. The estimation device 200 corresponds to an example of a display device in that it includes the second display unit 220.

[0043] The second operation input unit 230 includes input devices such as a keyboard and a mouse, and receives a user operation. For example, the second operation input unit 230 may receive a user operation that specifies a method for outputting an estimated value of a response to a non-anonymized prompt.

[0044] The second storage unit 280 stores various data. The second storage unit 280 is configured using a storage device included in the estimation device 200. For example, when the estimation device 200 replaces a partial string in a response from the response device 910, which corresponds to a string replaced by the anonymization device 100, with a string before the replacement of the replaced string, the second storage unit 280 may store data indicating the replaced string and the string before the replacement of the replaced string. Alternatively, when the estimation device 200 estimates a response to a non-anonymized prompt using a machine learning model, the second storage unit 280 may store the machine learning model.

[0045] The second control unit 290 performs various processes by controlling each unit of the estimation device 200. The functions of the second control unit 290 may be performed by a CPU included in the estimation device 200 reading a program from the second storage unit 280 and executing the program.

[0046] The response acquisition unit 291 acquires a response from the response device 910 to the anonymized prompt. For example, the second communication unit 210 may receive a response from the response device 910, and the response acquisition unit 291 may extract the response from the response device 910 from the received data of the second communication unit 210.

[0047] The estimation unit 292 estimates a response to a non-anonymized prompt based on the response to the anonymized prompt acquired by the response acquisition unit 291. The estimation unit 292 corresponds to an example of an estimation means. A response to a non-anonymized prompt is also referred to as an ideal response. A response that a user of the estimation system 1 desires to give in response to a non-anonymized prompt may be used as the ideal response. Alternatively, if a response from the response device 910 to a non-anonymized prompt can be obtained, an actual response from the response device 910 may be used as the ideal response.

[0048] The ideal response output unit 293 outputs the estimated value of the ideal response by the estimation unit 292. The method by which the ideal response output unit 293 outputs the estimated value of the ideal response is not limited to a specific method. For example, the ideal response output unit 293 may cause the second display unit 220 to display the estimated value of the ideal response. Alternatively, in addition to or instead of displaying the estimated value of the ideal response by the second display unit 220, the ideal response output unit 293 may cause the second communication unit 210 to transmit the estimated value of the ideal response to another device.

[0049] According to the estimation system 1, an input string to a system that outputs a string response in response to an input string can be a string in which the information to be concealed is concealed. Specifically, according to the estimation system 1, the information to be concealed that is included in a prompt can be concealed and transmitted to the response device 910, and an ideal response (a response of the response device 910 to a prompt in which the information is not concealed) can be estimated.

[0050] For example, if the response device 910 is located on the cloud, such as when the response device 910 is a large-scale language model made publicly available on the cloud, there is a possibility that the information may be leaked if a prompt including information that should be concealed (information that is the subject of concealment) is sent directly to the response device 910. In contrast, in the estimation system 1, the anonymization device 100 anonymizes information that should be anonymized and is included in a prompt, thereby reducing the possibility of information leakage. Moreover, in the estimation system 1, the anonymization device 200 is expected to obtain an ideal response by estimating an ideal response (a response of the response device 910 to a prompt in which information is not anonymized).

[0051] Information that is the subject of anonymization is also referred to as anonymized information. The anonymized information is not limited to a specific type of information. For example, information related to privacy, such as personal information, may be anonymized information, or information related to business activities, such as trade secrets, may be anonymized information. Alternatively, both information related to privacy and information related to business activities may be anonymized information.

[0052] The concealment of the concealed information and the estimation of the ideal response will now be described further.

[0053] (Anonymity using hash values) The anonymization unit 192 may replace a substring that is the target of anonymization with a hash value of the substring. For example, the anonymization unit 192 detects a substring in a prompt that corresponds to a string that is specified as the target of anonymization. The anonymization unit 192 then inputs the detected substring to a hash function to calculate a hash value. The anonymization unit 192 replaces the detected substring in the prompt with the hash value of the substring.

[0054] The hash function used by the concealment unit 192 to calculate the hash value is not limited to a specific hash function. In addition, the representation format of the hash value is not limited to a specific format. For example, the anonymization unit 192 may use a hash value represented in hexadecimal numbers, or a hash value represented in decimal numbers.

[0055] In order to make it more difficult for a third party to restore the original substring from the hash value, the anonymization unit 192 may add a string of random numbers to the substring to be anonymized before calculating the hash value. For example, if “apple” is the substring to be anonymized, the anonymization unit 192 may input “apple10”, which is obtained by adding a random number “10” to the end of “apple”, into a hash function to calculate a hash value, and replace “apple” in the prompt with the hash value.

[0056] The position where the anonymization unit 192 adds the character string based on random numbers to the partial character string to be concealed is not limited to a specific position. For example, the anonymization unit 192 may add the character string based on random numbers to the end of the partial character string to be concealed. Alternatively, the anonymization unit 192 may add the character string based on random numbers to the beginning of the partial character string to be concealed. Alternatively, the anonymization unit 192 may add the character string based on random numbers to the partial character string to be concealed, for example, by adding the character string based on random numbers between the first character and the second character of the partial character string to be concealed.

[0057] The concealment unit 192 may determine a character string based on random numbers for each prompt. In this case, a different hash value is assigned to the same substring for each prompt, thereby reducing the possibility of leakage of concealed information. In this case, the same substring is replaced with the same hash value in one prompt. This is expected to allow the response device 910 to understand that the same hash value indicates the same substring, and generate and output a relatively accurate response. In addition, when the response device 910 uses the hash value indicated in the prompt in the response, the estimation device 200 can convert the hash value in the response to the original substring. In this respect, it is expected that the estimation device 200 can estimate the ideal response with a relatively high accuracy.

[0058] Alternatively, the anonymization unit 192 may determine a random string each time it detects a substring to be anonymized. In this case, different hash values ​​are assigned to the same substring even within a single prompt, which further reduces the possibility of leakage of anonymized information.

[0059] In order to assist the response device 910 in recognizing the hash value, the anonymization unit 192 may add a specific character string to the hash value. For example, if "apple" is the substring to be anonymized, the anonymization unit 192 may input "apple" to a hash function to calculate a hash value, prepend the string "Fruits:" to the resulting hash value, and replace "apple" in the prompt with a string in the format "Fruits:(hash value)".

[0060] The training of the answering device 910 may be performed by fine tuning, which updates parameter values ​​of a machine learning model of the answering device 910. For example, the fine tuning function of the answering device 910 may be exposed to a user. Alternatively, the response device 910 may be trained using an In-Context Learning (also called Few-Shot) technique in which sample prompts and sample responses to those prompts are sent to the response device 910.

[0061] By having the response device 910 learn in advance the character string that the concealment unit 192 adds to the hash value, it is expected that the response device 910 can recognize a partial character string that is a combination of the character string and the hash value with higher accuracy. As a result, it is expected that the response device 910 can generate a response with higher accuracy. For example, the response device 910 is expected to learn that a substring in the format "Fruits:(hash value)" indicates a substring that falls into the class "Fruits".

[0062] The character string that the concealment unit 192 adds to the hash value may be a character string obtained using a machine learning model, such as class classification using a machine learning model. For example, the anonymization unit 192 inputs the partial character string "apple" to be anonymized to a machine learning model to obtain the class classification result "Fruits." The anonymization unit 192 also inputs the partial character string "apple" to a hash function to calculate a hash value. The anonymization unit 192 then adds the character string "Fruits" of the class classification result to the beginning of the obtained hash value to obtain a character string in the format of "Fruits: (hash value)," and replaces "apple" in the prompt with the character string in the format of "Fruits: (hash value)."

[0063] (Encryption-based concealment) The anonymization unit 192 may replace the partial character string to be anonymized with an encrypted partial character string. The encryption algorithm used by the anonymization unit 192 to encrypt the partial character string is not limited to a specific type of encryption algorithm. For example, the anonymization unit 192 may use an encryption algorithm included in a common key cryptosystem, or an encryption algorithm included in a public key cryptosystem.

[0064] Moreover, the encryption algorithm used by the concealment unit 192 may be one for which the response device 910 does not have information for decryption. As in the case of anonymization using a hash value, the anonymization unit 192 may add a character string based on random numbers to the partial character string to be anonymized before encrypting it.

[0065] As in the case of anonymization using a hash value, the anonymization unit 192 may add a specific character string to the encrypted partial character string. In this case, the character string that the anonymization unit 192 adds to the encrypted partial character string may be a character string that has been learned in advance by the response device 910. In addition, the character string that the anonymization unit 192 adds to the encrypted partial character string may be a character string obtained by using a machine learning model, such as class classification by a machine learning model.

[0066] (Concealment by replacing with words of the same part of speech) The anonymization unit 192 may replace the word to be anonymized with a word of the same part of speech. For example, if "apple" is a word to be anonymized, the anonymization unit 192 replaces "apple" in the prompt with "car," which is a noun like "apple." The replacement word may be designated in advance by the user. Alternatively, the anonymization unit 192 may have a dictionary and randomly select a replacement word from among words of the same part of speech as the word to be anonymized. By the anonymization unit 192 replacing the word to be anonymized with a word of the same part of speech, it is expected that the response device 910 will be able to grasp the prompt relatively accurately and generate a response with relatively high accuracy.

[0067] (Anonymization by word replacement using feature vectors) The concealment unit 192 may determine the replacement word by using the feature vector. The input and output of a feature extraction model f used by the concealment unit 192 to extract a feature vector can be expressed as shown in equation (1).

[0068]

number

[0069] W represents a set of words. d indicates a d-dimensional real space. The feature extraction model f receives a word w∈W as input and outputs a d-dimensional real vector. The feature extraction model f is trained so that the distance between the feature vectors of words that are semantically similar is small, and words with closer feature vectors are treated as more similar.

[0070] The anonymization unit 192 is configured to anonymize the word w s The distance between the feature vectors is dist(f(ws ),f(w)) is replaced with the top K words w∈W in ascending order. Here, dist is a function indicating the distance between two vectors. K is an integer K≧1. The value of K may be set by the user.

[0071] For example, suppose the word to be concealed is "apple," and the distance between the feature vector of "apple" and the feature vector of "melon" is the smallest, and the distance between the feature vector of "apple" and the feature vector of "orange" is the second smallest. When K=1, the anonymization unit 192 replaces "apple" in the prompt with "melon." On the other hand, when K=2, the anonymization unit 192 replaces "apple" in the prompt with "orange."

[0072] Alternatively, words may be classified into classes in advance, and the anonymization unit 192 may replace words within a class. For example, let us divide a set of words W into C classes W1, W2, . . . , W C Divide it into:

[0073]

number

[0074] The entity that performs this division is not limited to a specific entity. For example, this division may be performed by the anonymization unit 192. Alternatively, this division may be performed by an entity other than the anonymization unit 192, such as a user. Also, based on equation (3), the mean vector of each class f - i is calculated in advance.

[0075]

number

[0076] i indicates an identification number for identifying a class and is an integer of 1≦i≦C. The class identified by the identification number i is also referred to as the i-th class. |W i | is the i-th class W i Indicates the number of words contained in. The entity that calculates the average vector is not limited to a specific entity. For example, the anonymization unit 192 may calculate the average vector. Alternatively, the average vector may be calculated by an entity other than the anonymization unit 192, such as an external device of the anonymization device 100.

[0077] When concealing words, the concealment unit 192 uses the average vector f - 1, f - 2, , f - C Among them, the word w that is the subject of anonymization s The feature vector f(w s ) and the distance dist(f - i ,f(w s )) is the smallest mean vector f - i Detect.

[0078] Then, the anonymization unit 192 calculates the word w that is the target of anonymization in the prompt based on the formula (4). s Let f be the detected mean vector - i Replace it with the top K words w∈W in ascending order of distance l between the feature vectors.

[0079]

number

[0080] The first storage unit 180 may store the feature vector f(w) of each word calculated when calculating formula (3). Then, the concealment unit 192 may store the average vector f(w) of each class stored in the first storage unit 180 when calculating formula (4). - iAnd the feature vector r(w) of each word may be used.

[0081] The anonymization unit 192 calculates a word w that is to be anonymized in the prompt based on the formula (5) instead of the formula (4). s Let f be the detected mean vector - i Alternatively, the word w may be replaced with the top K words w∈W in ascending order of distance l between the feature vectors.

[0082]

number

[0083] r is a vector of random numbers with d dimensions, the same as the dimensions of the feature vector. According to formula (5), using random numbers when selecting words is expected to make it more difficult for a third party to restore the original word from the replaced word, just as in the case of using random numbers when calculating the hash value described above.

[0084] As in the case where random numbers are used in the calculation of the hash value described above, the anonymization unit 192 may determine the value of the random number vector r for each prompt, or may determine the value of the random number vector r each time the anonymization unit 192 detects a substring that is the subject of anonymization.

[0085] The concealment unit 192 may obtain the representative vector for each class by a method other than the method of calculating the average vector, for example, by having the user specify the representative vector for each class. For example, the user may specify representative words for each class in advance, and the anonymization unit 192 or another entity may calculate feature vectors of the specified representative words in advance and prepare them as representative vectors.

[0086] When concealing a word, the concealment unit 192 uses a representative vector for each class instead of the average vector for each class described above. s The feature vector f(w s ) is detected. The anonymization unit 192 then detects the representative vector that is closest to the word w s is replaced with the top K words w∈W in ascending order of distance between the detected representative vector and its feature vector.

[0087] The concealment unit 192 calculates the average vector f - i Alternatively, the concealment unit 192 may use an equation in which the average vector f - i Alternatively, an equation in which the above is replaced with a representative vector for each class may be used.

[0088] (Confidentiality to ensure privacy) The anonymization unit 192 may employ replacement of a partial string of the prompt in such a way as to satisfy a standard defined as a privacy protection standard, such as Local Differential Privacy (LDP). For example, the anonymization unit 192 may replace a substring of the prompt based on a privacy protection standard such as Metric Local Differential Privacy (MLDP) or Utility-optimized MLDP (UMLDP) using an algorithm disclosed as an algorithm for satisfying the privacy protection standard. However, the privacy protection standard used by the concealment unit 192 is not limited to a specific standard.

[0089] (Estimation of ideal responses by string replacement) The estimation unit 292 may convert the words replaced by the anonymization unit 192, among the partial character strings included in the response from the response device 910, to the words before replacement. Here, it is assumed that the partial character string after replacement by the concealment unit 192 is used as is in the response from the response device 910. In the following, a case where the concealment unit 192 and the estimation unit 292 replace words will be described as an example. However, the concealment unit 192 and the estimation unit 292 may replace characters other than words, such as phrases.

[0090] For example, a dictionary d indicating character replacement by the concealment unit 192 as shown in equation (6) is prepared, and the concealment unit 192 and the estimation unit 292 share the dictionary d.

[0091]

number

[0092] w o i indicates the i-th original word. w r i indicates the i-th replacement word. When the estimation unit 292 uses the dictionary d, w o i corresponds to the dictionary value, and w r i corresponds to a dictionary key. The estimation unit 292 estimates the word w in the response from the response device 910. r i If it contains the word w r i The word w o i Replace with.

[0093] The dictionary d may be generated in advance. Alternatively, in a case where the anonymizing unit 192 determines a replacement word according to a predetermined rule, the anonymizing unit 192 may generate the dictionary d and transmit the generated dictionary d to the estimation unit 292 via the first communication unit 110 and the second communication unit 210.

[0094] The same word w exists in dictionary d. r i When there are a plurality of words, the estimation unit 292 estimates the word w included in the response from the response device 910. r i The word w r i Words associated with w o i For example, the estimation unit 292 may replace the word w r i Multiple words associated with w o i Any one of these may be selected at random.

[0095] Alternatively, the estimation unit 292 may estimate the word w included in the response from the response device 910. r i The word w r i Multiple ws associated with o i In this case, the user can select multiple w o i You can choose either one of them. One word w r i Words associated with w o i If there are multiple, then in dictionary d, a word w r i and multiple words w o i The list may be linked to the list of

[0096] When the anonymization unit 192 encrypts a word to be anonymized, w o imay be a secret key. That is, the estimation unit 292 estimates the encrypted word w r i w as the private key o i The estimation unit 292 may input the encrypted word w included in the response from the response device 910 to an encryption algorithm to decrypt the word. r i may be replaced with the decoded word.

[0097] FIG. 5 is a diagram showing an example of input and output of data when the estimation system 1 estimates an ideal response by replacing a character string. 5, the anonymization device 100 transmits the anonymized prompt to the response device 910. In addition, the anonymization device 100 transmits replacement information, which is information related to the replacement of substrings by the anonymization device 100, such as dictionary information, to the estimation device 200. The response device 910 generates a response to the anonymized prompt from the anonymization device 100 and transmits the generated response to the estimation device 200. The estimation device 200 estimates an ideal response based on the replacement information from the concealment device 100 and the response from the response device 910 .

[0098] (Estimation of ideal responses using machine learning models) The estimation unit 292 may be configured to include a machine learning model that outputs a character string in response to an input of a character string. In the following, a case where the estimation unit 292 is configured to include a neural network will be described as an example. However, the machine learning model included in the estimation unit 292 is not limited to a specific type.

[0099] Fig. 6 is a diagram showing an example of the configuration of a learning system when training a neural network included in the estimation unit 292. In the example of Fig. 6, the learning system 2 includes a concealment device 100, an estimation device 200, a learning control device 300, and a response device 910. 6, parts having similar functions to those of the parts in FIG. 1 are given the same reference numerals (100, 200, 910), and detailed description thereof will be omitted here.

[0100] Fig. 7 is a diagram showing an example of the configuration of the learning control device 300. In the configuration shown in Fig. 7, the learning control device 300 includes a third communication unit 310, a third display unit 320, a third operation input unit 330, a third storage unit 380, and a third control unit 390. The third control unit 390 includes a training data acquisition unit 391 and a learning control unit 392.

[0101] The learning control device 300 controls the training of the neural network included in the estimation unit 292 . The learning control device 300 may be configured using a computer. The third communication unit 310 communicates with other devices under the control of the third control unit 390. For example, when the learning control device 300 accumulates training data, the third communication unit 310 may transmit a prompt to the response device 910 and receive a response.

[0102] The third display unit 320 includes a display screen such as a liquid crystal panel or an LED panel, and displays various images under the control of the third control unit 390. For example, the third display unit 320 may display various information related to training, such as the progress of training of the neural network.

[0103] The third operation input unit 330 includes input devices such as a keyboard and a mouse, and receives user operations. For example, the third operation input unit 330 may receive a user operation for setting training-related settings, such as a learning rate.

[0104] The third storage unit 380 stores various data. The third storage unit 380 is configured using a storage device included in the learning control device 300. For example, the third storage unit 380 may store training data for training a neural network.

[0105] The third control unit 390 performs various processes by controlling each unit of the learning control device 300. The functions of the third control unit 390 may be performed by a CPU included in the learning control device 300 reading and executing a program from the third storage unit 380.

[0106] The training data acquisition unit 391 acquires training data. For example, if a user of the estimation system 1 inputs a prompt t c And the prompt c Ideal response to y c Prepare a set of pairs with .

[0107]

number

[0108] Then, the concealment device 100 issues a prompt t c The prompt that has been made anonymous is then sent to the response device 910. Then, the training data acquisition unit 391 generates training data as shown in equation (8).

[0109]

number

[0110] Here, E indicates the conversion (anonymization of a substring) by the anonymization unit 192. E(t c ) indicates the prompt after conversion by the concealment unit 192. f indicates the response by the response device 910. f(E(t c )) indicates the response of the response device 910 to the prompt after conversion by the anonymization unit 192.

[0111] Concat indicates the concatenation of character strings. Concat combines multiple character strings into one character string. When the training data acquisition unit 391 combines character strings using concat, it may insert a specific character indicating a separator between the character strings, such as "[SEP]", between the character strings. y c i is the ideal response y c The data identification number i is indicated in the table.

[0112] The learning control unit 392 trains the neural network using the training data acquired by the training data acquisition unit 391. The learning control unit 392 trains the neural network by inputting the string concat(t c ,E(t c ),f(E(t c ))) i Input the string y c i The neural network is trained to output The neural network included in the estimation unit 292 may be configured as a language model, for example, a language model smaller than a large-scale language model.

[0113] The learning algorithm used by the learning control unit 392 to train the neural network is not limited to a specific one. For example, the learning control unit 392 may train the neural network using backpropagation, but is not limited to this. In addition, the learning control unit 392 may train the neural network using cross-entropy loss as a loss function, but is not limited to this.

[0114] FIG. 8 is a diagram showing an example of input and output of data in the learning system 2 when training the neural network included in the estimation unit 292. In the example of FIG. 8, for example, if the user of the estimation system 1 issues a prompt t c And the prompt cIdeal response to y c The set of pairs {(t c ,y c ) i} i=1 N Then, the encryption device 100 prepares a prompt t c The anonymized prompt E(t c ) to the response device 910. In addition, the encryption device 100 receives a prompt t c And prompt t c The anonymized prompt E(t c ) and prompt t c Ideal response to y c and transmit it to the learning control device 300.

[0115] The response device 910 receives the encrypted prompt E(t c ) to the response f(E(t c )) and transmits it to the learning control device 300. The learning control device 300 prompts c And prompt t c The anonymized prompt E(t c ) and prompt t c Ideal response to y c and prompt E(t c ) to the response f(E(t c )) and the training data {concat(t c ,E(t c ),f(E(t c ))) i ,y c i} i=1 N Generate. The learning control device 300 generates training data {concat(t c ,E(t c ),f(E(t c ))) i ,y c i} i=1 NThe neural network included in the estimation unit 292 is trained using the above.

[0116] FIG. 9 is a diagram showing an example of input and output of data in the estimation system 1 when the estimating unit 292 estimates an ideal response. In the example of FIG. 9, the concealment device 100 c Get the prompt c The anonymized prompt E(t c ) to the response device 910. In addition, the encryption device 100 receives a prompt t c And prompt t c The anonymized prompt E(t c ) to the estimation device 200.

[0117] The response device 910 receives the encrypted prompt E(t c ) to the response f(E(t c )) and transmits it to the estimation device 200. In the estimation device 200, the estimation unit 292 receives a prompt t c The anonymized prompt E(t c ) and the anonymized prompt E(t c ) to the response f(E(t c )) and the concatenated string concat(t c ,E(t c ),f(E(t c The estimation unit 292 inputs the generated character string to the trained neural network to generate an estimate y c ' is calculated.

[0118] (Estimation of ideal responses using trained language models) The estimation unit 292 may include a trained language model as a machine learning model. In this case, the learning parameters of the language model may not be updated, and examples of input and output of the language model may be given to the language model to cause the language model to estimate an ideal response by an in-context learning method.

[0119] In this case, the estimation unit 292 calculates the string concat(t c ,E(t c ),f(E(t c ))) sample and its prompt t c Ideal response to y c We obtain several pairs of samples of and provide them to the language model as input and output samples. 9, the estimation unit 292 estimates the prompt t c Then, the encryption device 100 issues a prompt t c The anonymized prompt E(t c ) and the anonymized prompt E(t c ) of the response device 910 f(E(t c )) and the concatenated string concat(t c ,E(t c ),f(E(t c The estimation unit 292 inputs the generated character string to a language model and calculates an estimate y c ' is calculated.

[0120] The estimation unit 292 directly estimates the prompt t of the target of the ideal response estimation by the Zero-Shot method without providing the input and output examples of the language model to the language model. c Then, the encryption device 100 issues a prompt t c The anonymized prompt E(t c ) and the anonymized prompt E(t c ) of the response device 910 f(E(t c )) and the concatenated string concat(t c ,E(t c ),f(E(t c ))) is input to the language model to obtain the estimate of the ideal response y c ' may be calculated.

[0121] (Additional learning of response device 910) In order to make it easier for a user of the estimation system 1 to obtain a response desired by the estimation device 200 based on an estimated value of an ideal response, additional learning may be performed on the response device 910.

[0122] 10 is a diagram showing an example of the configuration of a learning system when additional learning is performed on a response device 910. In the example of FIG. 10, parts having similar functions to those in FIG. 6 are given the same reference numerals (100, 300, 910), and detailed description thereof will be omitted here.

[0123] FIG. 11 is a diagram showing an example of data input and output in the learning system 3 when additional learning is performed on the response device 910. In FIG. In the example of FIG. 11, for example, if the user of the estimation system 1 issues a prompt t c And the prompt c Ideal response to y c The set of pairs {(t c ,y c ) i} i=1 N Prepare the following. Then, the concealment device 100 issues a prompt t c and the ideal response y c Then, we anonymize each of the prompts E(t c ) and the concealed response E(y c ) and transmits it to the learning control device 300.

[0124] The training data acquisition unit 391 of the learning control device 300 receives the concealed prompt E(t c ) and the concealed response E(y c ), we obtain the training data D shown in equation (9).

[0125]

number

[0126] The learning control unit 392 of the learning control device 300 trains the response device 910 using the training data D. The learning control unit 392 trains the response device 910 using the concealed prompt E(t c ) is input to the response device 910, and the response device 910 outputs the concealed response E(y c The learning parameters of the response device 910 are adjusted so that the response device 910 outputs

[0127] The additional learning of the response device 910 performed by the learning control unit 392 can be considered as fine tuning for adapting the generation of responses by the response device 910 to the replacement of partial character strings by the concealment device 100 and the estimation device 200. The learning algorithm used by the learning control unit 392 to train the response device 910 is not limited to a specific one. For example, the learning control unit 392 may train the response device 910 using LoRA (Low-Rank Adaptation), but is not limited thereto.

[0128] According to the learning system 3, additional learning of the response device 910 can be performed using training data in which the concealed information is replaced with a character string that is not recognizable to a third party. In this way, the learning system 3 can reduce the possibility that the concealed information is leaked to a third party, and can train the response device 910 so that the response device 910 outputs a response according to the concealed information.

[0129] In particular, the encryption device 100 receives a prompt t c and the ideal response y c The anonymization device 100 anonymizes the anonymized prompt E(t c ) and the concealed response E(y c ), there is little possibility that the confidential information will be leaked to a third party. In addition, the anonymization device 100 receives a prompt t c Concealment of and ideal response y c In the anonymization of c) and the concealed response E(y c The learning control device 300 embeds the concealed prompt E(t c ) and the concealed response E(y c ) to train the response device 910, the response device 910 can be trained to output a response according to the concealed information.

[0130] Furthermore, the learning system 3 can increase the likelihood that the response device 910 will be able to grasp the prompt, and in this respect, it is expected that the response device 910 will be able to generate responses with relatively high accuracy. For example, when the anonymization device 100 replaces a partial character string of a prompt that is to be anonymized with a hash value, the presence of the hash value may reduce the accuracy with which the response device 910 analyzes the prompt. The reduced accuracy with which the response device 910 analyzes the prompt may cause the response device 910 to output a prompt that makes it difficult to estimate an ideal response. In response to this, it is expected that by having the learning control unit 392 train the response device 910 using prompts and responses that include hash values, the response device 910 will be able to generate responses to prompts that include hash values ​​with greater accuracy.

[0131] The estimation device 200 may be constructed using the concealment device 100 used for additional learning of the response device 910, and may be used to estimate an ideal response. Here, the process performed by the estimation device 200 depends on the process performed by the anonymization device 100. For example, when the estimation device 200 estimates an ideal response by replacing a partial string, it is considered that the estimation device 200 performs a replacement of the partial string that corresponds to the inverse conversion of the replacement of the partial string performed by the anonymization device 100.

[0132] Therefore, the anonymization device 100 and the estimation device 200 are configured to share information regarding the replacement of substrings, such as by sharing a dictionary. c The anonymized prompt E(tc ) and the ideal response y c The anonymized response E(y c ) may be registered in a dictionary.

[0133] In addition, when the estimation unit 292 includes a machine learning model, the character string concat(t c ,E(t c ),f(E(t c In generating the prompt E(t c ) is used to reflect the processing by the concealment device 100 in the processing by the estimation device 200.

[0134] As described above, the anonymization unit 192 generates an anonymized prompt by replacing a partial character string detected as an object of anonymization in a non-anonymized prompt with another partial character string. The anonymized prompt issuing unit 193 inputs the anonymized prompt to the response device 910.

[0135] According to the anonymization device 100, the prompt for the response device 910 can be a prompt in which the information to be anonymized is anonymized. As a result, according to the anonymization device 100, the possibility that the information to be anonymized will be leaked can be reduced. In addition, according to the anonymization device 100, a partial character string detected as a target for anonymization is replaced with another partial character string, and it is expected that a response according to the information to be anonymized can be obtained.

[0136] Furthermore, the anonymization unit 192 replaces a substring detected as a target for anonymization in a non-anonymized prompt with a hash value of the substring. According to the anonymization device 100, a partial character string to be substituted for a partial character string can be determined to be a random character string through a relatively simple calculation using a hash function. In this respect, according to the anonymization device 100, the anonymization of information can be more reliably performed through a relatively simple calculation.

[0137] Furthermore, the anonymization unit 192 replaces a partial character string detected as a target for anonymization in the first character string with a hash value of a character string in which a random number character string is added to the partial character string. According to the anonymization device 100, it is possible to further reduce the possibility that information to be anonymized will be leaked.

[0138] Furthermore, the anonymization unit 192 replaces a substring detected as a target for anonymization in a non-anonymized prompt with a substring in which a predetermined string has been added to the hash value of the substring. The anonymization device 100 can assist the response device 910 in recognizing the hash value. In particular, the anonymization device 100 can provide the response device 910 with information for recognizing the hash value by adding a predetermined character string to the hash value.

[0139] Furthermore, the response device 910 has previously learned about the character string to be added to the hash value. According to the concealment device 100, it is expected that the response device 910 can appropriately recognize the hash value by using the character string added to the hash value.

[0140] In addition, the anonymization unit 192 inputs a partial character string detected as a target for anonymization from among the non-anonymized prompt to a trained machine learning model to obtain a predetermined character string. According to the anonymization device 100, by obtaining a string using a machine learning model, it is possible to obtain a string related to the substring that is the target of replacement, and the obtained string can be added to a hash value.

[0141] Furthermore, the anonymization unit 192 replaces a substring detected as a target for anonymization in a non-anonymized prompt with a substring obtained by encrypting the substring. According to the anonymization device 100, a partial character string to be substituted for a partial character string can be determined to be a random character string through encryption. In this respect, according to the anonymization device 100, the anonymization of information can be performed more reliably.

[0142] In addition, the anonymization unit 192 replaces a substring detected as a target for anonymization from an unanonymized prompt with a substring obtained by encrypting a string obtained by adding a random number string to the substring. According to the anonymization device 100, it is possible to further reduce the possibility that information to be anonymized will be leaked.

[0143] In addition, the anonymization unit 192 replaces a substring detected as a target for anonymization from an unanonymized prompt with a substring obtained by encrypting the substring and adding a predetermined string to the string obtained. The anonymization device 100 can assist the response device 910 in recognizing a character string. In particular, the anonymization device 100 can provide the response device 910 with information for recognizing the encrypted character string by adding a predetermined character string to the encrypted character string.

[0144] Also, the response device 910 has previously learned about the string that is added to the encrypted string. According to the anonymization device 100, it is expected that the response device 910 can appropriately recognize the encrypted character string by using the character string added to the encrypted character string.

[0145] In addition, the anonymization unit 192 inputs a partial character string detected as a target for anonymization from among the non-anonymized prompt to a trained machine learning model to obtain a predetermined character string. According to the anonymization device 100, by obtaining a string using a machine learning model, it is possible to obtain a string related to the substring that is the target of replacement, and to add the obtained string to the encrypted string.

[0146] Furthermore, the anonymization unit 192 replaces a word detected as a target for anonymization among words included in a non-anonymized prompt with a word of the same part of speech as the word detected as the target. According to the anonymization device 100, since the parts of speech before and after the word replacement are the same, it is expected that the response device 910 can grasp the prompt with relatively high accuracy and generate a response with relatively high accuracy.

[0147] The anonymization unit 192 also generates a feature vector of a substring detected as a target for anonymization from among the non-anonymized prompts, and selects one of the candidates for the replacement substring based on the generated feature vector. The anonymization device 100 can replace a substring with a substring that is relatively similar based on some criteria. This is expected to enable the response device 910 in the anonymization device 100 to grasp the prompt with relatively high accuracy and generate a response with relatively high accuracy.

[0148] Furthermore, the anonymization unit 192 employs a replacement for a substring of a non-anonymized prompt that satisfies standards defined as privacy protection standards. The anonymization device 100 can provide a prompt that can protect privacy.

[0149] In addition, the response device 910 is trained using anonymized prompts and strings in which the anonymizer 192 has performed substring replacements on the responses desired as output by the response device 910 to unanonymized prompts.

[0150] According to the anonymization device 100, additional learning of the response device 910 can be performed using training data in which the anonymized information is replaced with a character string that is not recognizable to a third party. In this way, in the learning system 3, it is possible to reduce the possibility that the anonymized information is leaked to a third party, and to train the response device 910 so that the response device 910 outputs a response according to the anonymized information. Furthermore, the anonymization device 100 can increase the likelihood that the response device 910 will be able to understand the prompt, and in this respect, it is expected that the response device 910 will be able to generate a response with relatively high accuracy.

[0151] Moreover, the first display unit 120 displays a concealed prompt in which a partial character string that is a target of concealment in the concealed prompt has been replaced with another partial character string. According to the anonymization device 100, the user can refer to the anonymized prompt and check whether the information has been appropriately anonymized.

[0152] In addition, the estimation unit 292 estimates a response to a non-anonymized prompt in which a substring detected as a target for anonymization has not been replaced, based on a response to a non-anonymized prompt in which a substring detected as a target for anonymization has been replaced with another substring. According to the estimation device 200, it is possible to obtain a response from the response device 910 using a concealed prompt, and it is possible to reduce the possibility that concealed information will be leaked to a third party. Moreover, according to the estimation device 200, it is possible to estimate a response according to the concealed information.

[0153] In addition, the estimation unit 292 replaces a substring specified as a replacement string in the substring replacement rules, among the substrings of a response to a concealed prompt, with the substring specified as a string before replacement. The estimation device 200 can perform replacement that is equivalent to the inverse conversion of the partial character string replacement performed by the concealment device 100, and in this respect, it is expected that the ideal response can be estimated with high accuracy.

[0154] Furthermore, when a partial string replacement rule specifies that there are multiple pre-replacement strings for one replacement string, the estimation unit 292 replaces the partial string specified as a replacement string in the replacement rule among the partial strings in the response to the concealed prompt with any one of the multiple pre-replacement strings. According to the estimation device 200, even if the replacement of partial character strings by the anonymization device 100 does not have a one-to-one correspondence relationship, it is possible to estimate an ideal response.

[0155] In addition, the estimation unit 292 inputs data including a prompt before replacement of a substring detected as a target for anonymization, the prompt after replacement, and a response to the input of the prompt as input data, and inputs data including the prompt before replacement of a substring detected as a target for anonymization, the prompt after replacement, and the response to the input of the prompt into a machine learning model trained using training data in which the response to the prompt before replacement of a substring detected as a target for anonymization is used as correct answer data, and obtains an estimate of the response to the prompt before replacement of a substring detected as a target for anonymization. According to the estimation device 200, it is possible to estimate an ideal response by using a machine learning model. As a result, the estimation device 200 can estimate an ideal response even in a case where, for example, a string replacement rule by the anonymization device 100 cannot be obtained.

[0156] In addition, the second display unit 220 displays responses to non-anonymized prompts in which a substring that is the subject of anonymization has been replaced with another substring, which responses are estimated based on responses to anonymized prompts in which a substring that is the subject of anonymization has been replaced with another substring. The estimation device 200 allows the user to refer to an estimated value of the ideal response.

[0157] Second Embodiment 12 is a diagram showing another example of the configuration of the anonymization device according to at least one embodiment. In the configuration shown in FIG. 12, the anonymization device 610 includes an anonymization unit 611 and an anonymization prompt issuing unit 612.

[0158] In this configuration, the anonymization unit 611 generates a second string in which a substring detected as a target for anonymization in the first string is replaced with another substring. The anonymization prompt issuing unit 612 inputs the second string to a device that outputs a response to the second string. The anonymization unit 611 corresponds to an example of an anonymization means, and the anonymization prompt issuing unit 612 corresponds to an example of an anonymization prompt issuing means.

[0159] According to the anonymization device 610, an input character string to a device that outputs a response to a second character string can be a character string in which the information to be anonymized is anonymized. As a result, according to the anonymization device 610, it is possible to reduce the possibility that the information to be anonymized will be leaked. In addition, according to the anonymization device 610, a partial character string detected as a target for anonymization is replaced with another partial character string, and thus it is expected that a response according to the information to be anonymized can be obtained.

[0160] The anonymization unit 611 can be realized, for example, by using the functions of the anonymization unit 192 in Fig. 3. The anonymization prompt issuing unit 612 can be realized, for example, by using the functions of the anonymization prompt issuing unit 193 in Fig. 3.

[0161] <Third embodiment> 13 is a diagram showing another example of the configuration of an estimation device according to at least one embodiment. In the configuration shown in FIG.

[0162] With this configuration, the estimation unit 621 estimates a fourth string, which is a response to a first string in which the partial string detected as a target for anonymization has not been replaced, based on a third string that is a response to a second string in which a partial string detected as a target for anonymization in the first string has been replaced with another partial string. The estimation unit 621 corresponds to an example of an estimation means.

[0163] According to the estimation device 620, it is possible to obtain a response using the anonymized second character string, and it is possible to reduce the possibility that the anonymized information is leaked to a third party. Moreover, according to the estimation device 620, it is possible to estimate a response according to the information to be anonymized. The estimation unit 621 can be realized, for example, by using the functions of the estimation unit 292 in FIG.

[0164] <Fourth embodiment> 14 is a diagram showing an example of the configuration of a display device according to at least one embodiment. In the configuration shown in FIG. With this configuration, the display unit 631 displays a second character string in which a partial character string that is to be concealed in the first character string has been replaced with another partial character string.

[0165] The display unit 631 corresponds to an example of a display means. Display device 630 allows the user to refer to the second character string and check whether information has been appropriately concealed. Display unit 631 can be realized, for example, by using the function of first display unit 120 in FIG.

[0166] <Fifth embodiment> 15 is a diagram showing another example of the configuration of a display device according to at least one embodiment. In the configuration shown in FIG. With this configuration, the display unit 641 displays a fourth string, which is a response to the first string in which the substring that is the target of anonymization has not been replaced, estimated based on a third string, which is a response to a second string in which a substring that is the target of anonymization of the first string has been replaced with another substring.

[0167] The display unit 641 corresponds to an example of a display means. The display device 640 allows the user to see the estimated fourth character string. The display unit 641 can be realized, for example, by using the function of the second display unit 220 in FIG.

[0168] Sixth Embodiment 16 is a diagram illustrating an example of the configuration of an estimation system according to at least one embodiment. In the configuration illustrated in FIG. 16, an estimation system 6590 includes a concealment unit 651 and an estimation unit 652.

[0169] In this configuration, the anonymization unit 651 generates a second string in which a partial string detected as a target for anonymization in the first string is replaced with another partial string, and the estimation unit 652 estimates a fourth string that is a response to the first string in which the partial string detected as a target for anonymization is not replaced, based on a third string that is a response to the second string.

[0170] According to the estimation system 650, it is possible to obtain a response using the anonymized second character string, and it is possible to reduce the possibility that the anonymized information is leaked to a third party. Moreover, according to the estimation system 650, it is possible to estimate a response according to the information to be anonymized. The concealment unit 651 can be realized, for example, by using the functions of the concealment unit 192 in Fig. 3. The estimation unit 652 can be realized, for example, by using the functions of the estimation unit 292 in Fig. 4.

[0171] Seventh Embodiment FIG. 17 is a diagram illustrating an example of a processing procedure in a concealment method according to at least one embodiment. The anonymization method shown in FIG. 17 includes performing anonymization (step S611) and issuing an anonymized prompt (step S612).

[0172] In performing the anonymization (step S611), the computer generates a second string in which a partial character string detected as a target for anonymization in the first string is replaced with another partial character string. In issuing the anonymous prompt (step S612), the computer inputs a second string into a device that outputs a response to the second string.

[0173] According to the anonymization method shown in Fig. 17, an input character string to a device that outputs a response to a second character string can be a character string in which the information to be anonymized is anonymized. As a result, according to the anonymization method shown in Fig. 17, it is possible to reduce the possibility that the information to be anonymized will be leaked. Moreover, according to the anonymization method shown in Fig. 17, a partial character string detected as an object to be anonymized is replaced with another partial character string, and it is expected that a response according to the information to be anonymized can be obtained.

[0174] Eighth Embodiment FIG. 18 is a diagram illustrating an example of a processing procedure in an estimation method according to at least one embodiment. The estimation method shown in FIG. 18 includes performing estimation (step S621).

[0175] In making the estimation (step S621), the computer estimates a fourth string that is a response to the first string in which the substring detected as a target for anonymization has not been replaced, based on a third string that is a response to a second string in which a substring of the first string that has been detected as a target for anonymization has been replaced with another substring.

[0176] According to the estimation method shown in Fig. 18, a response can be obtained using the anonymized second character string, and the possibility that the anonymized information will be leaked to a third party can be reduced. Moreover, according to the estimation method shown in Fig. 18, a response according to the anonymized information can be estimated.

[0177] <Ninth embodiment> FIG. 19 is a diagram showing another example of the process steps in the estimation method according to at least one embodiment. The estimation method shown in FIG. 19 includes performing concealment (step S631) and performing estimation (step S632).

[0178] In performing the anonymization (step S631), the computer generates a second string in which a partial character string detected as a target for anonymization in the first string is replaced with another partial character string. In making an estimation (step S632), the computer estimates a fourth string, which is a response to the first string without replacing the substring detected as the target of anonymization, based on a third string that is a response to the second string.

[0179] According to the estimation method shown in Fig. 19, a response can be obtained using the anonymized second character string, and the possibility that the anonymized information is leaked to a third party can be reduced. Moreover, according to the estimation method shown in Fig. 19, a response according to the anonymized information can be estimated.

[0180] The above-mentioned concealment device can also be used in a system that searches for a method of treatment to be administered to a patient or a medicine to be administered to a patient from information in a medical record that indicates the patient's condition. For example, an electronic medical record system that electronically manages medical records transmits the information in the electronic medical record to a search system that searches for a method of treatment or a medicine to be administered in response to a doctor's instruction. The search system creates a first character string for searching for a treatment method or a medicine using a sentence in the information in the electronic medical record, and transmits the created character string to an estimation system such as that described above. Alternatively, the search system may include the estimation system such as that described above. The search system transmits a fourth character string acquired from the estimation system to the electronic medical record system. The electronic medical record system receives the fourth character string and displays the received fourth character string.

[0181] The above-mentioned concealment device can also be used in a system for searching for the sentence for a criminal entity from documents showing the progress of a case committed by the criminal entity. For example, a management system that electronically manages documents showing the progress of a case transmits information about the documents showing the progress of a case to a search system that searches for sentences, in response to instructions from a prosecutor, a judge, or a lawyer. The search system creates a first character string for searching for the sentence using a sentence in the document showing the progress of the case, and transmits the created character string to an inference system such as that described above. Alternatively, the search system may include an inference system such as that described above. The search system transmits a fourth character string obtained from the inference system to the management system. The management system receives the fourth character string and displays the received fourth character string.

[0182] The above-mentioned anonymization device can also be used in a system that searches for products to be recommended to a consumer from the consumer's purchase history. For example, a management system that electronically manages information representing a consumer's purchase history transmits the information of the consumer's purchase history to a search system that searches for products to be recommended to the consumer. The search system creates a first character string for searching for products to be recommended to the consumer by using a sentence in the information of the consumer's purchase history, and transmits the created character string to an estimation system such as that described above. Alternatively, the search system may include the estimation system such as that described above. The search system transmits a fourth character string acquired from the estimation system to the management system. The management system receives the fourth character string and displays the received fourth character string.

[0183] FIG. 20 illustrates an example configuration of a computer according to at least one embodiment. In the configuration shown in FIG. 20, a computer 700 includes a CPU 710, a main memory device 720, an auxiliary memory device 730, an interface 740, and a non-volatile recording medium 750.

[0184] Any one or more of the above-mentioned anonymization device 100, estimation device 200, learning control device 300, anonymization device 610, estimation device 620, display device 630, display device 640, and estimation system 650, or a part thereof, may be implemented in the computer 700. In this case, the operation of each of the above-mentioned processing units is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, develops it in the main storage device 720, and executes the above-mentioned processing according to the program. In addition, the CPU 710 secures a storage area corresponding to each of the above-mentioned storage units in the main storage device 720 according to the program. Communication between each device and other devices is executed by the interface 740 having a communication function and performing communication according to the control of the CPU 710. In addition, the interface 740 has a port for the non-volatile recording medium 750, and reads information from the non-volatile recording medium 750 and writes information to the non-volatile recording medium 750.

[0185] When the anonymization device 100 is implemented in a computer 700, the operations of the first control unit 190 and each of its units are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing in accordance with the program.

[0186] Furthermore, the CPU 710 allocates a storage area for the first storage unit 180 in the main storage device 720 in accordance with the program. Communication with other devices by the first communication unit 110 is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Display of images by the first display unit 120 is performed by the interface 740 having a display device and displaying various images under the control of the CPU 710. Acceptance of a user operation by the first operation input unit 130 is performed by the interface 740 having an input device and accepting the user operation under the control of the CPU 710.

[0187] When the estimation device 200 is implemented in a computer 700, the operations of the second control unit 290 and each of its units are stored in the form of a program in an auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing in accordance with the program.

[0188] Furthermore, the CPU 710 allocates a storage area for the second storage unit 280 in the main storage device 720 in accordance with the program. Communication with other devices by the second communication unit 210 is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Display of images by the second display unit 220 is performed by the interface 740 having a display device and displaying various images under the control of the CPU 710. Reception of a user operation by the second operation input unit 230 is performed by the interface 740 having an input device and accepting the user operation under the control of the CPU 710.

[0189] When the learning control device 300 is implemented in a computer 700, the operations of the third control unit 390 and each of its units are stored in the form of a program in an auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing according to the program.

[0190] Furthermore, CPU 710 secures a storage area for third storage unit 380 in main storage device 720 in accordance with the program. Communication with other devices by third communication unit 310 is performed by interface 740 having a communication function and operating under control of CPU 710. Display of images by third display unit 320 is performed by interface 740 having a display device and displaying various images under control of CPU 710. Acceptance of user operations by third operation input unit 330 is performed by interface 740 having an input device and accepting user operations under control of CPU 710.

[0191] When the anonymization device 610 is implemented in the computer 700, the operations of the anonymization unit 611 and the anonymization prompt issuing unit 612 are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing in accordance with the program.

[0192] Furthermore, the CPU 710 secures a storage area in the main storage device 720 for the anonymization device 610 to perform processing in accordance with the program. Communication between the anonymization device 610 and other devices is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Interaction between the anonymization device 610 and a user is performed by the interface 740 having an input device and an output device, presenting information to the user via the output device under the control of the CPU 710, and accepting user operations via the input device.

[0193] When the estimation device 620 is implemented in the computer 700, the operation of the estimation unit 621 is stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing in accordance with the program.

[0194] Furthermore, the CPU 710, in accordance with the program, secures a storage area in the main storage device 720 for the estimating device 620 to perform processing. Communication between the estimating device 620 and other devices is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Interaction between the estimating device 620 and a user is performed by the interface 740 having an input device and an output device, presenting information to the user via the output device under the control of the CPU 710, and accepting user operations via the input device.

[0195] When the display device 630 is implemented in the computer 700, the operation of the display device 630 is stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing in accordance with the program.

[0196] Furthermore, the CPU 710, in accordance with the program, secures a storage area in the main storage device 720 for the display device 630 to perform processing. Communication between the display device 630 and other devices is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Interaction between the display device 630 and a user is performed by the interface 740 having an input device and an output device, presenting information to the user via the output device under the control of the CPU 710, and accepting user operations via the input device.

[0197] When the display device 640 is implemented in the computer 700, the operation of the display device 640 is stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing in accordance with the program.

[0198] Furthermore, the CPU 710 allocates a storage area in the main storage device 720 for the display device 640 to perform processing in accordance with the program. Communication between the display device 640 and other devices is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Interaction between the display device 640 and a user is performed by the interface 740 having an input device and an output device, presenting information to the user via the output device under the control of the CPU 710, and accepting user operations via the input device.

[0199] When the estimation system 650 is implemented in the computer 700, the operations of the concealment unit 651 and the estimation unit 652 are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing according to the program.

[0200] Furthermore, the CPU 710, in accordance with the program, secures a storage area in the main storage device 720 for the estimation system 650 to perform processing. Communication between the estimation system 650 and other devices is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Interaction between the estimation system 650 and a user is performed by the interface 740 having an input device and an output device, presenting information to the user via the output device under the control of the CPU 710, and accepting user operations via the input device.

[0201] Any one or more of the above-mentioned programs may be recorded in non-volatile recording medium 750. In this case, interface 740 may read the program from non-volatile recording medium 750. Then, CPU 710 may directly execute the program read by interface 740, or may temporarily store the program in main storage device 720 or auxiliary storage device 730 and execute it.

[0202] Note that the processes of each part may be performed by recording a program for executing all or part of the processes performed by the anonymization device 100, the estimation device 200, the learning control device 300, the anonymization device 610, the estimation device 620, the display device 630, the display device 640, and the estimation system 650 in a computer-readable recording medium, and reading and executing the program recorded in the recording medium into a computer system. Note that the term "computer system" here includes hardware such as an OS (Operating System) and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs (Read Only Memory), and CD-ROMs (Compact Disc Read Only Memory), as well as storage devices such as hard disks built into computer systems. The above-mentioned program may be for realizing part of the above-mentioned functions, or may be capable of realizing the above-mentioned functions in combination with a program already recorded in the computer system.

[0203] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and designs that do not deviate from the gist of the present invention are also included.

[0204] A part or all of the above-described embodiments can be described as, but is not limited to, the following supplementary notes.

[0205] (Appendix 1) an anonymization means for generating a second string by replacing a substring detected as an anonymization target in the first string with another substring; a prompt generating means for inputting the second character string to a device that outputs a response to the second character string; A concealment device comprising:

[0206] (Appendix 2) The anonymization means replaces a substring detected as a target for anonymization in the first string with a hash value of the substring. 2. The encryption device according to claim 1.

[0207] (Appendix 3) the anonymization means replaces a substring detected as a target for anonymization in the first string with a hash value of a string obtained by adding a random number to the substring, 2. The encryption device according to claim 1.

[0208] (Appendix 4) the anonymization means replaces a substring detected as a target for anonymization in the first string with a substring obtained by adding a predetermined character string to a hash value of the substring, 2. The encryption device according to claim 1.

[0209] (Appendix 5) a device that receives the second character string and outputs the response is trained on the predetermined character string to be added to the hash value; 5. The encryption device according to claim 4.

[0210] (Appendix 6) The anonymization means inputs a substring detected as a target for anonymization in the first string into a trained machine learning model to obtain the predetermined string. 6. The concealment device according to claim 4 or 5.

[0211] (Appendix 7) The anonymization means replaces a substring detected as a target for anonymization in the first string with a substring obtained by encrypting the substring. 2. The encryption device according to claim 1.

[0212] (Appendix 8) The anonymization means replaces a substring detected as a target for anonymization in the first string with a substring obtained by encrypting a string obtained by adding a character string based on random numbers to the substring, 2. The encryption device according to claim 1.

[0213] (Appendix 9) the anonymization means replaces a substring detected as a target for anonymization in the first string with a substring obtained by encrypting the substring and adding a predetermined character string to the resulting string; 2. The encryption device according to claim 1.

[0214] (Appendix 10) a device that receives the second character string and outputs the response is trained on the predetermined character string that is added to the encrypted character string; 10. The anonymization device according to claim 9.

[0215] (Appendix 11) The anonymization means inputs a substring detected as a target for anonymization in the first string into a trained machine learning model to obtain the predetermined string. 11. The concealment device according to claim 9 or 10.

[0216] (Appendix 12) the anonymization means replaces a word detected as a target for anonymization among words included in the first character string with a word of the same part of speech as the word detected as a target for anonymization, 2. The encryption device according to claim 1.

[0217] (Appendix 13) the anonymization means generates a feature vector of a substring detected as a target for anonymization in the first string, and selects one of candidates for a replacement substring based on the generated feature vector. 2. The encryption device according to claim 1.

[0218] (Appendix 14) The anonymizing means adopts a replacement for the substring of the first string that satisfies a standard defined as a privacy protection standard. 14. The anonymization device according to any one of claims 1 to 13.

[0219] (Appendix 15) a device for outputting a response to the second string is trained using the second string and a string obtained by performing substring replacement on a string desired as an output of the device for the first string by the concealment means; 15. The encryption device according to claim 1,

[0220] (Appendix 16) an estimation means for estimating a fourth character string, which is a response to a first character string in which a partial character string detected as a target of anonymization has not been replaced, based on a third character string returned in response to a second character string in which a partial character string detected as a target of anonymization has been replaced with another partial character string; An estimation device comprising:

[0221] (Appendix 17) the estimation means replaces a substring specified as a replacement target string in a substring replacement rule, among the substrings of the third string, with a substring specified as a string before replacement; 17. The estimation apparatus of claim 16.

[0222] (Appendix 18) When a substring replacement rule specifies that there are a plurality of pre-replacement strings for one replacement string, the estimation means replaces, among the substrings of the third string, a substring specified as a replacement string in the replacement rule with any one of the plurality of pre-replacement strings. 18. The estimation apparatus of claim 17.

[0223] (Appendix 19) the estimation means receives as input data data including a fifth string which is a string before replacement of a partial string detected as a target of concealment, a sixth string which is a string after the replacement, and a seventh string which is a response to an input of the sixth string, and inputs data including the first string, the second string, and the third string into a machine learning model trained using training data in which an eighth string which is a response to the fifth string is received as correct answer data, and acquires an estimate of the fourth string. 17. The estimation apparatus of claim 16.

[0224] (Appendix 20) A display means for displaying a second character string in which a subcharacter string corresponding to the masking target in the first character string has been replaced with another subcharacter string. A display device comprising:

[0225] (Appendix 21) a display means for displaying a fourth character string which is a response to the first character string in which a partial character string corresponding to the anonymization target in the first character string has not been replaced, the fourth character string being estimated based on a third character string which is a response to a second character string in which a partial character string corresponding to the anonymization target in the first character string has been replaced with another partial character string; A display device comprising:

[0226] (Appendix 22) an anonymization means for generating a second string by replacing a substring detected as an anonymization target in the first string with another substring; an estimation means for estimating a fourth character string that is a response to the first character string without replacement of the partial character string detected as a target of anonymization, based on a third character string that is a response to the second character string; An estimation system comprising:

[0227] (Appendix 23) The anonymization means replaces a substring detected as a target for anonymization in the first string with a hash value of the substring. 23. The estimation system of claim 22.

[0228] (Appendix 24) the anonymization means replaces a substring detected as a target for anonymization in the first string with a hash value of a string obtained by adding a random number to the substring, 23. The estimation system of claim 22.

[0229] (Appendix 25) the anonymization means replaces a substring detected as a target for anonymization in the first string with a substring obtained by adding a predetermined character string to a hash value of the substring, 23. The estimation system of claim 22.

[0230] (Appendix 26) a device that receives the second character string and outputs the response is trained on the predetermined character string to be added to the hash value; 26. The estimation system of claim 25.

[0231] (Appendix 27) The anonymization means inputs a substring detected as a target for anonymization in the first string into a trained machine learning model to obtain the predetermined string. 27. The estimation system of claim 25 or 26.

[0232] (Appendix 28) The anonymization means replaces a substring detected as a target for anonymization in the first string with a substring obtained by encrypting the substring. 23. The estimation system of claim 22.

[0233] (Appendix 29) The anonymization means replaces a substring detected as a target for anonymization in the first string with a substring obtained by encrypting a string obtained by adding a character string based on random numbers to the substring, 23. The estimation system of claim 22.

[0234] (Appendix 30) the anonymization means replaces a substring detected as a target for anonymization in the first string with a substring obtained by encrypting the substring and adding a predetermined character string to the resulting string; 23. The estimation system of claim 22.

[0235] (Appendix 31) a device that receives the second character string and outputs the response is trained on the predetermined character string that is added to the encrypted character string; 31. The estimation system of claim 30.

[0236] (Appendix 32) The anonymization means inputs a substring detected as a target for anonymization in the first string into a trained machine learning model to obtain the predetermined string. 32. The estimation system of claim 30 or 31.

[0237] (Appendix 33) the anonymization means replaces a word detected as a target for anonymization among words included in the first character string with a word of the same part of speech as the word detected as a target for anonymization, 23. The estimation system of claim 22.

[0238] (Appendix 34) the anonymization means generates a feature vector of a substring detected as a target for anonymization in the first string, and selects one of candidates for a replacement substring based on the generated feature vector. 23. The estimation system of claim 22.

[0239] (Appendix 35) The anonymizing means adopts a replacement for the substring of the first string that satisfies a standard defined as a privacy protection standard. 35. The estimation system of any one of claims 22 to 34.

[0240] (Appendix 36) a device for outputting a response to the second string is trained using the second string and a string obtained by performing substring replacement on a string desired as an output of the device for the first string by the concealment means; 36. The estimation system of any one of claims 22 to 35.

[0241] (Appendix 37) the estimation means replaces a substring specified as a replacement target string in a substring replacement rule, among the substrings of the third string, with a substring specified as a string before replacement; 37. The estimation system of any one of claims 22 to 36.

[0242] (Appendix 38) When a substring replacement rule specifies that there are a plurality of pre-replacement strings for one replacement string, the estimation means replaces, among the substrings of the third string, a substring specified as a replacement string in the replacement rule with any one of the plurality of pre-replacement strings. 38. The estimation system of any one of claims 22 to 37.

[0243] (Appendix 39) the estimation means receives as input data data including a fifth string which is a string before replacement of a partial string detected as a target of concealment, a sixth string which is a string after the replacement, and a seventh string which is a response to an input of the sixth string, and inputs data including the first string, the second string, and the third string into a machine learning model trained using training data in which an eighth string which is a response to the fifth string is received as correct answer data, and acquires an estimate of the fourth string. 37. The estimation system of any one of claims 22 to 36.

[0244] (Appendix 40) The computer generating a second string in which a substring of the first string detected as a target for anonymization is replaced with another substring; inputting the second string into a device that outputs a response to the second string; A method for concealing information, comprising:

[0245] (Appendix 41) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a hash value of the substring. 41. The method of concealment described in Supplementary Note 40.

[0246] (Appendix 42) In generating the second character string, a substring of the first character string detected as a target for anonymization is replaced with a hash value of a character string obtained by adding a character string generated by random numbers to the substring. 41. The method of concealment described in Supplementary Note 40.

[0247] (Appendix 43) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by adding a predetermined character string to a hash value of the substring. 41. The method of concealment described in Supplementary Note 40.

[0248] (Appendix 44) a device that receives the second character string and outputs the response is trained on the predetermined character string to be added to the hash value; 44. The method of claim 43.

[0249] (Appendix 45) In generating the second character string, a substring of the first character string detected as a target for anonymization is input to a trained machine learning model to obtain the predetermined character string. 45. The method of concealment according to claim 43 or 44.

[0250] (Appendix 46) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting the substring. 41. The method of concealment described in Supplementary Note 40.

[0251] (Appendix 47) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting a character string obtained by adding a character string based on random numbers to the substring. 41. The method of concealment described in Supplementary Note 40.

[0252] (Appendix 48) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting the substring and adding a predetermined character string to the resulting string. 41. The method of concealment described in Supplementary Note 40.

[0253] (Appendix 49) a device that receives the second character string and outputs the response is trained on the predetermined character string that is added to the encrypted character string; 49. The method of concealment described in Supplementary Note 48.

[0254] (Appendix 50) In generating the second character string, a substring of the first character string detected as a target for anonymization is input to a trained machine learning model to obtain the predetermined character string. 49. The method of claim 48 or 49.

[0255] (Appendix 51) In generating the second character string, a word detected as a target for anonymization among words included in the first character string is replaced with a word having the same part of speech as the word. 41. The method of concealment described in Supplementary Note 40.

[0256] (Appendix 52) In generating the second character string, a feature vector of a substring detected as a target for anonymization in the first character string is generated, and one of candidates for a replacement substring is selected based on the generated feature vector. 41. The method of concealment described in Supplementary Note 40.

[0257] (Appendix 53) In generating the second character string, a replacement that satisfies a standard defined as a privacy protection standard is adopted as a replacement for a substring of the first character string. 53. The method of claim 40, wherein the first and second sub-items are identical to those of the first and second sub-items.

[0258] (Appendix 54) a device for outputting a response to the second string is trained using the second string and a string obtained by performing substring replacement on a string desired as an output of the device for the first string by the concealment means; 54. The method of claim 40, wherein the first and second sub-items are identical to those of the first and second sub-items.

[0259] (Appendix 55) The computer Based on a third character string that is a response to a second character string in which a partial character string detected as a target of anonymization in a first character string is replaced with another partial character string, a fourth character string that is a response to the first character string in which the partial character string detected as a target of anonymization is not replaced is estimated. The estimation method includes:

[0260] (Appendix 56) In the estimation, a substring of the third string that is specified as a replacement target string in a substring replacement rule is replaced with a substring that is specified as a pre-replacement string. Estimation method described in Appendix 55.

[0261] (Appendix 57) In the estimation, when a substring replacement rule specifies a plurality of pre-replacement strings for one replacement string, the substring of the third string specified as a replacement string in the replacement rule is replaced with any one of the plurality of pre-replacement strings. Estimation method described in Appendix 56.

[0262] (Appendix 58) In the estimation, data including a fifth string which is a string before replacement of a partial string detected as a target of concealment, a sixth string which is a string after the replacement, and a seventh string which is a response to the input of the sixth string is used as input data, and an eighth string which is a response to the fifth string is used as correct answer data. Data including the first string, the second string, and the third string is input to a machine learning model trained using training data to obtain an estimate of the fourth string. Estimation method described in Appendix 57.

[0263] (Appendix 59) The computer Displaying a second string in which a substring of the first string that is the subject of anonymization has been replaced with another substring Display methods including.

[0264] (Appendix 60) The computer Displaying a fourth character string, which is a response to the first character string in which a subcharacter string corresponding to the anonymization target in the first character string has been replaced with another subcharacter string, estimated based on a third character string, which is a response to the second character string in which the subcharacter string corresponding to the anonymization target in the first character string has not been replaced. Display methods including.

[0265] (Appendix 61) The computer generating a second string in which a substring of the first string detected as a target for anonymization is replaced with another substring; estimating a fourth character string that is a response to the first character string without replacing the subcharacter string detected as the target of anonymization, based on a third character string that is a response to the second character string; The estimation method includes:

[0266] (Appendix 62) In generating the second character string, a substring of the first character string detected as a target for anonymization is replaced with a hash value of the substring. Estimation method described in Appendix 61.

[0267] (Appendix 63) In generating the second character string, a substring of the first character string detected as a target for anonymization is replaced with a hash value of a character string obtained by adding a character string generated by random numbers to the substring. Estimation method described in Appendix 61.

[0268] (Appendix 64) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by adding a predetermined character string to a hash value of the substring. Estimation method described in Appendix 61.

[0269] (Appendix 65) a device that receives the second character string and outputs the response is trained on the predetermined character string to be added to the hash value; Estimation method described in Appendix 64.

[0270] (Appendix 66) In generating the second character string, a substring of the first character string detected as a target for anonymization is input to a trained machine learning model to obtain the predetermined character string. The estimation method described in Appendix 64 or Appendix 65.

[0271] (Appendix 67) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting the substring. Estimation method described in Appendix 61.

[0272] (Appendix 68) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting a character string obtained by adding a character string based on random numbers to the substring. Estimation method described in Appendix 61.

[0273] (Appendix 69) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting the substring and adding a predetermined character string to the resulting string. Estimation method described in Appendix 61.

[0274] (Appendix 70) a device that receives the second character string and outputs the response is trained on the predetermined character string that is added to the encrypted character string; Estimation method described in Appendix 69.

[0275] (Appendix 71) In generating the second character string, a substring of the first character string detected as a target for anonymization is input to a trained machine learning model to obtain the predetermined character string. The estimation method described in Appendix 69 or Appendix 70.

[0276] (Appendix 72) In generating the second character string, a word detected as a target for anonymization among words included in the first character string is replaced with a word having the same part of speech as the word. Estimation method described in Appendix 61.

[0277] (Appendix 73) In generating the second character string, a feature vector of a substring detected as a target for anonymization in the first character string is generated, and one of candidates for a replacement substring is selected based on the generated feature vector. Estimation method described in Appendix 61.

[0278] (Appendix 74) In generating the second character string, a replacement that satisfies a standard defined as a privacy protection standard is adopted as a replacement for a substring of the first character string. 74. The estimation method according to any one of claims 61 to 73.

[0279] (Appendix 75) a device for outputting a response to the second string is trained using the second string and a string obtained by performing substring replacement on a string desired as an output of the device for the first string by the concealment means; 75. The estimation method according to any one of claims 61 to 74.

[0280] (Appendix 76) In the estimation, a substring of the third string that is specified as a replacement target string in a substring replacement rule is replaced with a substring that is specified as a pre-replacement string. 66. The estimation method of any one of appendixes 61 to 75.

[0281] (Appendix 77) In the estimation, when a substring replacement rule specifies a plurality of pre-replacement strings for one replacement string, the substring of the third string specified as a replacement string in the replacement rule is replaced with any one of the plurality of pre-replacement strings. 77. The estimation method of any one of appendix 61 to 76.

[0282] (Appendix 78) In the estimation, data including a fifth string which is a string before replacement of a partial string detected as a target of concealment, a sixth string which is a string after the replacement, and a seventh string which is a response to the input of the sixth string is used as input data, and an eighth string which is a response to the fifth string is used as correct answer data. Data including the first string, the second string, and the third string is input to a machine learning model trained using training data to obtain an estimate of the fourth string. 8. The estimation method of any one of appendix 61 to 77.

[0283] (Appendix 79) On the computer, generating a second string in which a substring detected as a target for anonymization in the first string is replaced with another substring; inputting the second string into a device that outputs a response to the second string; A program that executes the following.

[0284] (Appendix 80) In generating the second character string, a substring of the first character string detected as a target for anonymization is replaced with a hash value of the substring; 79. The program of claim 79,

[0285] (Appendix 81) In generating the second character string, a substring of the first character string detected as a target for anonymization is replaced with a hash value of a character string obtained by adding a character string generated by random numbers to the substring; 79. The program of claim 79,

[0286] (Appendix 82) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by adding a predetermined character string to a hash value of the substring; 79. The program of claim 79,

[0287] (Appendix 83) a device that receives the second character string and outputs the response is trained on the predetermined character string to be added to the hash value; 83. The program according to claim 82.

[0288] (Appendix 84) In generating the second character string, a substring of the first character string detected as a target for anonymization is input to a trained machine learning model to obtain the predetermined character string; The program according to claim 82 or 83,

[0289] (Appendix 85) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting the substring; 79. The program of claim 79,

[0290] (Appendix 86) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting a character string obtained by adding a character string based on random numbers to the substring; 79. The program of claim 79,

[0291] (Appendix 87) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting the substring and adding a predetermined character string to the substring; 79. The program of claim 79,

[0292] (Appendix 88) A device that receives the second character string and outputs the response is previously trained on the predetermined character string that is added to the encrypted character string. 88. The program according to claim 87.

[0293] (Appendix 89) In generating the second character string, a substring of the first character string detected as a target for anonymization is input to a trained machine learning model to obtain the predetermined character string; The program according to claim 87 or 88,

[0294] (Appendix 90) generating the second character string includes replacing a word detected as a target for anonymization among words included in the first character string with a word of the same part of speech as the word; 79. The program of claim 79,

[0295] (Appendix 91) In generating the second character string, a feature vector of a substring detected as a target for anonymization in the first character string is generated, and one of candidates for a replacement substring is selected based on the generated feature vector; 79. The program of claim 79,

[0296] (Appendix 92) In generating the second character string, a replacement that satisfies a standard defined as a privacy protection standard is adopted as a replacement for a substring of the first character string; 92. The program of claim 79, further comprising:

[0297] (Appendix 93) a device for outputting a response to the second string is trained using the second string and a string obtained by performing substring replacement on a string desired as an output of the device for the first string by the concealment means; 93. A program according to any one of claims 79 to 92.

[0298] (Appendix 94) On the computer, Based on a third character string that is a response to a second character string in which a partial character string detected as a target of anonymization in a first character string is replaced with another partial character string, a fourth character string that is a response to the first character string in which the partial character string detected as a target of anonymization is not replaced is estimated. A program that executes the following.

[0299] (Appendix 95) The estimation includes replacing a substring of the third string, the substring being specified as a replacement target string in a substring replacement rule, with a substring being specified as a pre-replacement string; 95. The program of claim 94,

[0300] (Appendix 96) In the estimation, when a substring replacement rule has a plurality of pre-replacement strings for one replacement string, the substring of the third string that is specified as a replacement string in the replacement rule is replaced with any one of the plurality of pre-replacement strings; The program of claim 95,

[0301] (Appendix 97) The estimation includes inputting data including a fifth string which is a string before replacement of a partial string detected as a target for concealment, a sixth string which is a string after the replacement, and a seventh string which is a response to an input of the sixth string as input data, and inputting data including the first string, the second string, and the third string into a machine learning model trained using training data in which an eighth string which is a response to the fifth string is used as correct answer data, to obtain an estimate of the fourth string; The program of claim 96,

[0302] (Appendix 98) On the computer, Displaying a second string in which a substring of the first string that is the subject of anonymization has been replaced with another substring A program that executes the following.

[0303] (Appendix 99) On the computer, Displaying a fourth character string, which is a response to the first character string in which a subcharacter string corresponding to the anonymization target in the first character string has been replaced with another subcharacter string, estimated based on a third character string, which is a response to the second character string in which the subcharacter string corresponding to the anonymization target in the first character string has not been replaced. A program that executes the following.

[0304] (Appendix 100) On the computer, generating a second string in which a substring of the first string detected as a target for anonymization is replaced with another substring; estimating a fourth character string that is a response to the first character string without replacing the subcharacter string detected as the target of anonymization, based on a third character string that is a response to the second character string; A program that executes the following.

[0305] (Appendix 101) In generating the second character string, a substring of the first character string detected as a target for anonymization is replaced with a hash value of the substring; The program of claim 100,

[0306] (Appendix 102) In generating the second character string, a substring of the first character string detected as a target for anonymization is replaced with a hash value of a character string obtained by adding a character string generated by random numbers to the substring; The program of claim 100,

[0307] (Appendix 103) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by adding a predetermined character string to a hash value of the substring; The program of claim 100,

[0308] (Appendix 104) a device that receives the second character string and outputs the response has previously learned about the predetermined character string to be added to the hash value; 104. The program of claim 103.

[0309] (Appendix 105) In generating the second character string, a substring of the first character string detected as a target for anonymization is input to a trained machine learning model to obtain the predetermined character string; The program according to claim 103 or 104,

[0310] (Appendix 106) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting the substring; The program of claim 100,

[0311] (Appendix 107) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting a character string obtained by adding a character string based on random numbers to the substring; The program of claim 100,

[0312] (Appendix 108) In generating the second character string, a substring detected as a target for anonymization in the first character string is replaced with a substring obtained by encrypting the substring and adding a predetermined character string to the substring; The program of claim 100,

[0313] (Appendix 109) A device that receives the second character string and outputs the response is previously trained on the predetermined character string that is added to the encrypted character string. 101. The program of claim 100.

[0314] (Appendix 110) In generating the second character string, a substring of the first character string detected as a target for anonymization is input to a trained machine learning model to obtain the predetermined character string; The program according to claim 108 or 109,

[0315] (Appendix 111) generating the second character string includes replacing a word detected as a target for anonymization among words included in the first character string with a word of the same part of speech as the word; The program of claim 100,

[0316] (Appendix 112) In generating the second character string, a feature vector of a substring detected as a target for anonymization in the first character string is generated, and one of candidates for a replacement substring is selected based on the generated feature vector; The program of claim 100,

[0317] (Appendix 113) In generating the second character string, a replacement that satisfies a standard defined as a privacy protection standard is adopted as a replacement for a substring of the first character string; 13. The program of claim 100, further comprising:

[0318] (Appendix 114) a device for outputting a response to the second string is trained using the second string and a string obtained by performing substring replacement on a string desired as an output of the device for the first string by the concealment means; 14. A program according to any one of claims 100 to 113.

[0319] (Appendix 115) The estimation includes replacing a substring of the third character string, the substring being specified as a replacement target string in a substring replacement rule, with a substring being specified as a pre-replacement string; 15. The program of claim 100, further comprising:

[0320] (Appendix 116) In the estimation, when a substring replacement rule has a plurality of pre-replacement strings for one replacement string, the substring of the third string that is specified as a replacement string in the replacement rule is replaced with any one of the plurality of pre-replacement strings; 116. The program of claim 100, further comprising:

[0321] (Appendix 117) The estimation includes inputting data including a fifth string which is a string before replacement of a partial string detected as a target for concealment, a sixth string which is a string after the replacement, and a seventh string which is a response to an input of the sixth string as input data, and inputting data including the first string, the second string, and the third string into a machine learning model trained using training data in which an eighth string which is a response to the fifth string is used as correct answer data, to obtain an estimate of the fourth string; 117. The program of any one of appendices 100 to 116, [Explanation of symbols]

[0322] 1. Estimation System 2.3 Learning System 100 Concealment Device 110 First Communications Department 120 First display section 130 First operation input unit 180 First memory section 190 First Control Section 191 Prompt Acquisition Section 192 Concealment Department 193 Confidentiality Prompt Issuance Department 200 Estimation device 210 Second Communications Department 220 Second display section 230 Second operation input unit 280 Second memory section 290 Second Control Section 291 Response Acquisition Unit 292 Estimation Department 293 Ideal response output section 300 Learning control device 310 Third Communications Department 320 Third display 330 Third operation input unit 380 Third Memory 390 Third Control Section 391 Training Data Acquisition Department 392 Learning control section 910 Answering Machine

Claims

1. an anonymization means for generating a second string by replacing a substring detected as an anonymization target in the first string with another substring; a prompt generating means for inputting the second character string to a device that outputs a response to the second character string; A concealment device comprising:

2. an estimation means for estimating a fourth character string, which is a response to a first character string in which a partial character string detected as a target of anonymization has not been replaced, based on a third character string returned in response to a second character string in which a partial character string detected as a target of anonymization has been replaced with another partial character string; An estimation device comprising:

3. A display means for displaying a second character string in which a subcharacter string corresponding to the masking target in the first character string has been replaced with another subcharacter string. A display device comprising:

4. a display means for displaying a fourth character string which is a response to the first character string in which a partial character string corresponding to the anonymization target in the first character string has not been replaced, the fourth character string being estimated based on a third character string which is a response to a second character string in which a partial character string corresponding to the anonymization target in the first character string has been replaced with another partial character string; A display device comprising:

5. an anonymization means for generating a second string by replacing a substring detected as an anonymization target in the first string with another substring; an estimation means for estimating a fourth character string that is a response to the first character string without replacement of the partial character string detected as a target of anonymization, based on a third character string that is a response to the second character string; An estimation system comprising:

6. The computer generating a second string in which a substring of the first string detected as a target for anonymization is replaced with another substring; inputting the second string into a device that outputs a response to the second string; The method for concealing information includes:

7. The computer Based on a third character string that is a response to a second character string in which a partial character string detected as a target of anonymization in a first character string is replaced with another partial character string, a fourth character string that is a response to the first character string in which the partial character string detected as a target of anonymization is not replaced is estimated. The estimation method includes:

8. The computer generating a second string in which a substring of the first string detected as a target for anonymization is replaced with another substring; estimating a fourth character string that is a response to the first character string without replacing the subcharacter string detected as the target of anonymization, based on a third character string that is a response to the second character string; The estimation method includes:

9. On the computer, generating a second string in which a substring detected as a target for anonymization in the first string is replaced with another substring; inputting the second string into a device that outputs a response to the second string; A program that executes the following.

10. On the computer, Based on a third character string that is a response to a second character string in which a partial character string detected as a target of anonymization in a first character string is replaced with another partial character string, a fourth character string that is a response to the first character string in which the partial character string detected as a target of anonymization is not replaced is estimated. A program to make that happen.

Images