JP2018190239A - Deep learning automatic learning system, client device and server device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To securely provide a latest learning result from a server device to a client device via the Internet with no need to hold the communication between the client device and the server device until deep layer learning is completed.SOLUTION: A deep learning automatic learning system 1 according to an embodiment includes a client device 100 and a server device 200 connected via the Internet; the server device 200 has a push notification unit 214 that push-notifies encrypted learning version data that version management information is encrypted to the client device 100; and the client device 100 has a version management information acquisition unit 113 that decrypts the encrypted learning version data and acquires the version management information, and a data request unit 114 that determines whether or not learning result data corresponding to the version management information has already been acquired and requests transmission of the learning result data to the server device 200 when not acquired.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a deep learning automatic learning system, a client device, and a server device, and more specifically, performs deep learning based on learning data transmitted from the client device to the server device, and the latest learning result is transmitted from the server device to the client device. The present invention relates to a deep learning automatic learning system provided securely over the Internet, and a client device and a server device provided in the deep learning automatic learning system.

  Systems that perform deep learning using learning data obtained from external devices such as sensors are being actively studied. Deep learning is a kind of machine learning using a neural network having a multilayer structure. A structure adaptive deep learning method is known as a method of deep learning (see Non-Patent Document 1). This structure-adaptive deep learning method is a method for automatically obtaining the optimal number of hidden neurons and the number of layers during learning in a DBN (Deep Brief Network).

  Japanese Patent Application Laid-Open No. 2004-228561 describes a method of generating recommended data by analyzing diagnostic data generated by a client computer using a vendor computer system and transmitting the recommended data to the client.

JP 2004-348730 A

Makoto Kamada, Taku Ichimura, Akira Hara, “Structurally Adaptive Restricted Boltzmann Machine Using Neuron Generation / Extinction Algorithm”, Society of Instrument and Control Engineers, 8th Computational Intelligence Study Group, pp. 90-96 (2015)

  An information processing system for using deep learning is a system including a client device and a server device connected via the Internet, wherein the client device transmits learning data to the server device, and deep learning based on the learning data After the above is performed, a system in which the server device transmits the learning result to the client device can be considered. Although depending on the performance of a learning computer that performs deep learning, generally deep learning requires a relatively long time (for example, several hours).

  However, it is difficult to maintain communication (session) between the client device and the server device until the deep learning is completed because the session times out. Also, it is not desirable from the viewpoint of maintaining security to make the session timeout time longer than the time required for deep learning.

  Therefore, the present invention does not need to hold communication between the client device and the server device until the deep learning is completed, and can automatically provide the latest learning result from the server device to the client device through the Internet. An object is to provide a learning system, a client device, and a server device.

The deep learning automatic learning system according to the present invention is
A deep learning automatic learning system comprising a client device and a server device connected via the Internet,
The client device is
A key generation unit that generates a public key and a secret key based on identification information of the client device;
A new learning data transmission unit for transmitting new learning data to the server device,
The server device
A learning result data encryption unit that encrypts learning result data based on the new learning data and generates encrypted learning result data using the public key;
A version management information encryption unit that encrypts version management information for identifying the learning result data using the public key, and generates encrypted learning version data;
A push notification unit for performing push notification of the encrypted learning version data to the client device,
The client device is
A version management information acquisition unit that acquires the version management information by decrypting the encrypted learning version data that has been pushed from the server device using the secret key;
It is determined whether or not learning result data corresponding to the version management information has already been acquired, and if not acquired, a data request unit that requests the server device to transmit the learning result data;
A learning result data acquisition unit that acquires the learning result data by decrypting the encrypted learning result data transmitted from the server device to the client device in response to a request from the data request unit, using the secret key. And further comprising.

A client device according to the present invention provides:
A client device connected to a server device via the Internet,
A key generation unit that generates a public key and a secret key based on identification information of the client device;
A new learning data transmission unit for transmitting new learning data to the server device;
A version management information acquisition unit that obtains version management information for identifying learning result data by decrypting the encrypted learning version data that has been pushed from the server device using the secret key;
It is determined whether or not learning result data corresponding to the version management information has already been acquired, and if not acquired, a data request unit that requests the server device to transmit the learning result data;
A learning result data acquisition unit for acquiring the learning result data by decrypting the encrypted learning result data transmitted from the server device to the client device in response to the request of the data request unit, using the secret key; ,
It is characterized by providing.

The server device according to the present invention includes:
A server device connected to a client device via the Internet,
A learning result data encryption unit that encrypts learning result data and generates encrypted learning result data using a public key received from the client device;
A version management information encryption unit that encrypts version management information for identifying the learning result data and generates encrypted learning version data using the public key;
A push notification unit for pushing notification of the encrypted learning version data to the client device;
It is characterized by providing.

  In the deep learning automatic learning system according to the present invention, the server device has completed deep learning by pushing notification of encrypted learning version data obtained by encrypting version management information for identifying learning result data to the client device. And urges the client device to download new learning result data. The client device that has received the push notification determines whether or not learning result data corresponding to the version management information has already been acquired. If not, the client device requests the server device to transmit learning result data. The server device that has received the request transmits encrypted learning result data obtained by encrypting the learning result data to the client device. By doing in this way, the client apparatus can acquire the latest learning result data without holding the communication with the server apparatus until the deep learning is completed.

  Furthermore, in the deep learning automatic learning system according to the present invention, the version management information and the learning result data are encrypted using the public key and then transmitted to the client device, so the data is securely transmitted from the server device to the client device. Can be delivered.

  Therefore, according to the present invention, it is not necessary to hold communication between the client device and the server device until deep learning is completed, and the latest learning result can be securely provided from the server device to the client device via the Internet.

It is a schematic block diagram of the deep learning automatic learning system which concerns on embodiment. It is a schematic block diagram of the client apparatus which concerns on embodiment. It is a schematic block diagram of the server apparatus which concerns on embodiment. It is a figure which shows an example of the database for terminal management which concerns on embodiment. It is a schematic block diagram of the computer for learning concerning an embodiment. It is a flowchart between the client apparatus and server apparatus which concern on embodiment. It is a flowchart between the server apparatus which concerns on embodiment, and the computer for learning. It is a flowchart between the server apparatus and client apparatus which concern on embodiment. FIG. 9 is a flowchart between the server device and the client device according to the embodiment, following FIG. 8. 10 is a flowchart between the server device and the client device according to the embodiment, following FIG. 9. It is an example of the setting screen of the push notification which concerns on embodiment. It is an example of a screen of a push notification concerning an embodiment.

  Hereinafter, embodiments according to the present invention will be described with reference to the drawings.

<Deep learning automatic learning system>
First, a schematic configuration of a deep learning automatic learning system 1 according to an embodiment of the present invention will be described with reference to FIG. Note that the internal configuration and processing of each component of the deep learning automatic learning system 1 will be described with reference to FIG. 2 and subsequent drawings.

  The deep learning automatic learning system 1 includes a client device 100, a server device 200, and a learning computer 300. The client device 100 and the server device 200 are connected via the Internet. In FIG. 1, only one client device 100 is connected to the server device 200, but a plurality of client devices 100 may be connected to the server device 200.

  The client device 100 inputs new learning data from an external device such as an IOT device having various sensors. Here, the new learning data is data that is used for deep learning and is not yet used for deep learning. The client device 100 is specifically a personal computer, a tablet terminal, an embedded board, or the like.

  The client device 100 has identification information for identifying the client device 100. This identification information is, for example, a terminal unique number such as the MAC address or serial number of the client device 100.

  The server device 200 is also called a Web data server or a cloud server, and has a PUSH server and a data server as shown in FIG. The server device 200 is communicably connected to the learning computer 300, transmits learning data to the learning computer 300, and receives learning result data from the learning computer 300. Here, the learning result data is data obtained by deep learning, and is necessary for reconfiguring the neural network in the client device. Specifically, the learning result data is the shape, weight, and other parameter values of the neural network. Note that the learning result data may include knowledge accumulated by deep learning (such as IF-THEN rules).

  The learning computer 300 performs learning using the learning data received from the server device 200. For example, the learning computer 300 performs deep learning by the structure adaptive deep learning method described above. Note that the learning computer 300 may also be referred to as a GPU (Graphics Processing Units) computer.

  In the deep learning automatic learning system 1, as will be described in detail later, new learning data is uploaded from the client device 100 to the server device 200, and learning result data is downloaded from the server device 200 to the client device 100.

  The deep learning automatic learning system according to the present invention is not limited to the configuration shown in FIG. For example, the PUSH server and the data server of the server device 200 may be configured as separate server devices connected via a communication device. The server device 200 may be configured to include a learning computer 300. That is, the server device 200 and the learning computer 300 may be configured as one information processing device.

  Next, each of the client device 100, the server device 200, and the learning computer 300 will be described in more detail. First, details of the client device 100 will be described.

<Client device>
As illustrated in FIG. 2, the client device 100 includes a control unit 110, a storage unit 120, and a communication unit 130.

  The control unit 110 is realized by a processor in the client device 100 executing a predetermined program. The control unit 110 includes a key generation unit 111, a new learning data transmission unit 112, a version management information acquisition unit 113, a data request unit 114, a learning result data acquisition unit 115, a data deletion request unit 116, an inference Part 117 and output part 118. Details of each part will be described later.

  The storage unit 120 has a new learning data DB 121. The new learning data DB 121 stores new learning data input from an external device such as an IOT device. Note that the storage unit 120 includes, for example, a hard disk or a semiconductor memory (SSD or the like).

  The communication unit 130 is an interface for the client device 100 to transmit / receive information to / from the server device 200 via the Internet.

  Here, the detail of each part of the control part 110 is demonstrated.

  The key generation unit 111 generates a public key and a secret key based on the identification information of the client device 100. The public key is a key used when data is encrypted, and the secret key is a key used when data is decrypted. The key generation method is not particularly limited, and a known key generation method can be used.

  The new learning data transmission unit 112 transmits new learning data stored in the new learning data DB 121 to the server device 200. In the present embodiment, new learning data is securely transmitted to the server apparatus 200 using https.

  The version management information acquisition unit 113 decrypts the encrypted learning version data notified by the push from the server device 200 using the secret key generated by the key generation unit 111, and acquires version management information. Here, the version management information is information for identifying learning result data. The encrypted learning version data is data obtained by encrypting at least version management information with a public key. In this embodiment, the encrypted learning version data is data obtained by encrypting version management information and a registration ID (Registration ID) with a public key.

  The data request unit 114 determines whether or not learning result data corresponding to the version management information has already been acquired. If not acquired, the data request unit 114 requests the server device 200 to transmit learning result data. Note that the data request unit 114 may transmit the version of the version management information to the server device 200 when requesting transmission of learning result data. Thereby, the server apparatus 200 can grasp learning result data to be transmitted to the client apparatus 100.

  The learning result data acquisition unit 115 decrypts the encrypted learning result data transmitted from the server device 200 to the client device 100 in response to a request from the data request unit 114, using the secret key generated by the key generation unit 111. Learning result data. Note that the learning result data may include learning data used for deep learning. That is, the learning result data acquisition unit 115 may acquire a learning data set as learning result data. Here, the “learning data set” is a combination (data set) of learning data used for deep learning and learning result data.

  The encrypted learning result data is data obtained by encrypting at least the learning result data with a public key. In the present embodiment, the encrypted learning result data is data obtained by encrypting the learning result data and the registration ID with a public key.

  After the learning result data is acquired by the learning result data acquisition unit 115, the data deletion request unit 116 requests the server device 200 to delete the encrypted learning result data of the server device 200. Thereby, the encrypted learning result data downloaded to the client apparatus 100 is deleted from the server apparatus 200, and the memory of the server apparatus 200 can be saved.

  The inference unit 117 performs inference using the learning result data acquired by the learning result data acquisition unit 115. In this embodiment, inference is performed by a structure adaptive deep learning method.

  The inference unit 117 uses the learning result data when the registration ID obtained by decrypting the encrypted learning version data matches the registration ID obtained by decrypting the encrypted learning result data. It is preferable to perform inference by a structure adaptive deep learning method. As a result, it is possible to prevent an unintended situation from being inferred using another version of the learning result data.

  The output unit 118 outputs the result of inference performed by the inference unit 117. The output destination of the inference result is a display of the client device 100, a printer (not shown), the storage unit 120, or another device connected to the client device 100.

<Server device>
Next, details of the server apparatus 200 will be described.

  As illustrated in FIG. 3, the server device 200 includes a control unit 210, a storage unit 220, and a communication unit 230.

  The control unit 210 is realized by a processor in the server device 200 executing a predetermined program. The control unit 210 includes a registration ID issuing unit 211, a learning result data encryption unit 212, a version management information encryption unit 213, a push notification unit 214, a new learning data storage unit 215, and a learning data transmission unit 216. And a learning result data transmitting unit 217 and a data erasing unit 218. Details of each part will be described later.

  In this embodiment, the registration ID issuing unit 211, the version management information encryption unit 213, and the push notification unit 214 are provided in the PUSH server, and the learning result data encryption unit 212, the new learning data storage unit 215, the learning The data transmission unit 216, the learning result data transmission unit 217, and the data deletion unit 218 are provided in the data server.

  The storage unit 220 includes a terminal management DB 221, a learning data DB 222, and a learning result DB 223. The storage unit 220 includes, for example, a hard disk and a semiconductor memory (SSD or the like).

  As illustrated in FIG. 4, the terminal management DB 221 is a database that stores a terminal unique number (more generally, identification information) received from the client device 100 and a public key in association with each other. Note that the terminal management DB 221 may store a registration ID in association with a terminal unique number and a public key, as shown in FIG.

  The learning data DB 222 is a database in which new learning data received from the client device 100 is accumulated. The learning data DB 222 stores both existing learning data and new learning data. Here, the existing learning data is data already used for deep learning in the learning computer 300.

  The learning result DB 223 is a database in which learning result data received from the learning computer 300 is accumulated.

  The communication unit 230 is an interface for the server apparatus 200 to transmit / receive information to / from the client apparatus 100 via the Internet and to transmit / receive information to / from the learning computer 300.

  Here, the detail of each part of the control part 210 is demonstrated.

  The registration ID issuing unit 211 issues a registration ID (Registration ID) based on the identification information of the client device 100. The registration ID is information necessary for the push notification unit 214 to make a push notification to the client device 100, and is an ID unique to the client device 100. The registration ID includes information related to the communication path between the client device 100 and the server device 200.

  From the viewpoint of ensuring the security of communication between the client device 100 and the server device 200, the registration ID issuing unit 211 preferably issues a registration ID only when necessary, and deletes the registration ID when it is no longer needed. . Further, the registration ID issuing unit 211 may reissue (refresh) the registration ID at regular intervals.

  The learning result data encryption unit 212 uses the public key received from the client device 100 to encrypt the learning result data based on the new learning data, and generates encrypted learning result data. In the present embodiment, the learning result data encryption unit 212 encrypts the learning result data received from the learning computer 300 and the registration ID issued by the registration ID issuing unit 211, thereby encrypting learning result data. Is generated.

  Note that the learning result data encryption unit 212 may generate encrypted learning result data by encrypting the learning data set.

  The version management information encryption unit 213 encrypts the version management information of the learning result data received from the learning computer 300 using the public key received from the client device 100, and generates encrypted learning version data. In the present embodiment, the version management information encryption unit 213 encrypts the version management information and the registration ID to generate encrypted learning version data.

  The push notification unit 214 notifies the client device 100 of push of encrypted learning version data. In this embodiment, the push notification by the push notification unit 214 is performed using the registration ID. Note that the push notification may be performed using a device token other than the registration ID.

  The new learning data storage unit 215 stores the new learning data received from the client device 100 in the learning data DB 222.

  The learning data transmission unit 216 transmits the learning data to the learning computer 300 when the learning data in the learning data DB 222 satisfies a predetermined criterion. In the present embodiment, only the newly increased amount of the learning data stored in the learning data DB 222 is transmitted to the learning computer 300. That is, the difference between the learning data stored in the learning data DB 321 of the learning computer 300 and the learning data stored in the learning data DB 222 is transmitted to the learning computer 300. Not limited to this, the learning data transmission unit 216 may transmit both new learning data and existing learning data stored in the learning data DB 222 to the learning computer 300.

  “When the predetermined criterion is satisfied” means, for example, when a predetermined amount of new learning data is accumulated in the learning data DB 222 and / or when an error between the new learning data and the existing learning data exceeds a threshold value. is there. “When the error exceeds the threshold” means, for example, that the difference between the statistic of the existing learning data and the statistic of the new learning data is larger than a predetermined value, or the statistic of the existing learning data This is a case where the difference between the statistics of the data (existing learning data and new learning data) is larger than a predetermined value. For example, an average value is used as the statistic, but other statistic may be used.

  Upon receiving a learning result data transmission request (transmission request) from the client device 100, the learning result data transmission unit 217 transmits the encrypted learning result data generated by the learning result data encryption unit 212 to the client device 100.

  The data erasure unit 218 erases the encrypted learning result data when requested by the client device 100 to erase the encrypted learning result data of the server device 200 (that is, when an erasure request is received).

<Learning computer>
Next, details of the learning computer 300 will be described.

  As shown in FIG. 5, the learning computer 300 includes a control unit 310, a storage unit 320, and a communication unit 330.

  The control unit 310 is realized by a processor in the learning computer 300 executing a predetermined program. The control unit 310 includes a deep learning execution unit 311.

  The deep learning execution unit 311 performs deep learning using the learning data stored in the learning data DB 321. The deep learning execution unit 311 performs deep learning by, for example, a structure adaptive deep learning method.

  The storage unit 320 includes a learning data DB 321 and a learning result DB 322. The learning data DB 321 is a database in which learning data (existing learning data and new learning data) received from the server device 200 is accumulated. The learning result DB 322 is a database in which the output (learning result data) of the deep learning execution unit 311 is accumulated. That is, in the learning result DB 322, knowledge (IF-THEN rules, etc.) accumulated by deep learning is accumulated in addition to the shape, weight, and other parameter values of the neural network obtained by learning.

  The storage unit 320 includes, for example, a hard disk or a semiconductor memory (SSD or the like).

  The communication unit 330 is an interface for the learning computer 300 to transmit / receive information to / from the server device 200.

<Operation of deep learning automatic learning system>
Next, an example of the processing operation of the deep learning automatic learning system 1 having the above configuration will be described with reference to FIGS.

  First, with reference to FIG. 6, steps such as registration of a public key to a server device and uploading new learning data to the server device will be described.

  The key generation unit 111 of the client device 100 generates a public key and a secret key based on its own terminal number (that is, the client device 100) (step S11). This step is executed, for example, by operating the toggle button on the screen shown in FIG. 11 and turning on “Enable push notification”. When the push notification validation process is completed, registration ID information is displayed in the registration information window W.

  After the public key and the private key are generated, the client device 100 receives the terminal unique number and the public key via the communication unit 130 in order to have the server device 200 issue a registration ID necessary for push notification. It transmits to the apparatus 200 (PUSH server) (step S12).

  When the server device 200 receives the terminal unique number and the public key from the client device 100, the server device 200 associates and stores the terminal unique number and the public key in the terminal management DB 221 (step S21). Thereafter, the registration ID issuing unit 211 issues a registration ID based on the terminal unique number, and stores the issued registration ID in association with the terminal unique number in the terminal management DB 221 (step S22).

  The new learning data transmitting unit 112 of the client device 100 uploads new learning data input from an external device and stored in the new learning data DB 121 to the server device 200 (data server) (step S13). This step is performed, for example, after notification that the registration ID has been issued from the server device 200 (that is, the push notification validation process has been completed).

  When receiving the new learning data from the client device 100, the new learning data storage unit 215 of the server device 200 stores the received new learning data in the learning data DB 222 (step S23).

  Next, with reference to FIG. 7, processes such as learning data accumulation and automatic learning will be described.

  The learning data transmission unit 216 of the server device 200 determines whether or not the learning data stored in the learning data DB 222 satisfies a predetermined criterion (step S31). When the learning data satisfies a predetermined criterion (S31: Yes), the learning data transmission unit 216 transmits learning data for the newly increased amount to the learning computer 300 (step S32). When the learning computer 300 receives the learning data from the server device 200, the learning computer 300 stores the received learning data in the learning data DB 321.

  The learning computer 300 determines whether or not the learning data stored in the learning data DB 321 satisfies a predetermined criterion (step S41). The reference used in this step is, for example, a case where the number of learning data accumulated in the learning data DB 321 reaches a predetermined number. In addition, the reference | standard in this step may be the same reference | standard as the reference | standard in step S31, and a different reference | standard may be sufficient as it.

  When the learning data in the learning data DB 321 satisfies a predetermined criterion (S41: Yes), the deep learning execution unit 311 performs deep learning using the learning data accumulated in the learning data DB 321 (step S42). In this step, for example, deep learning is performed by a structure adaptive deep learning method.

  The deep learning execution part 311 preserve | saves learning result data in learning result DB322 after completion | finish of learning (step S43). Thereafter, the learning computer 300 transmits the learning result data obtained in step S42 to the server device 200 (step S44).

  When receiving the learning result data from the learning computer 300, the server device 200 stores the learning result data received from the learning computer 300 in the learning result DB 223 together with the version management information (step S33). Note that the learning result data may include learning data. In the present embodiment, the server device 200 stores the learning data set in the learning result DB 223 together with the version management information.

  Next, processes such as generation of encrypted data and push notification will be described with reference to FIG.

  When the learning result data is stored in the learning result DB 223, the learning result data encryption unit 212 of the server device 200 encrypts the registration ID and the learning data set by using the public key received from the client device 100, and performs encryption. The learning result data is generated (step S51).

  The version management information encryption unit 213 of the server device 200 encrypts the registration ID and the version management information using the public key received from the client device 100, and generates encrypted learning version data (step S52).

  The push notification unit 214 of the server device 200 notifies the client device 100 of the encrypted learning version data generated in step S52 (step S53). The push notification is performed using the registration ID issued in step S22. More specifically, a push notification is sent to the client device 100 corresponding to the registration ID issued in step S22.

  The version management information acquisition unit 113 of the client device 100 uses the private key generated in step S11 to decrypt the encrypted learning version data notified by push from the server device 200, and acquires the registration ID and version management information. (Step S61). The version management information acquired in this way indicates the version of the learning data set. FIG. 12 shows an example of the screen of the client device 100 notified of the push. In this example, the learning data set is updated in the server device 200, and the version of the updated learning data set is also shown.

  Next, with reference to FIG. 9, steps such as a learning result data transmission request and download will be described.

  The data request unit 114 of the client device 100 determines whether or not a learning data set corresponding to the version management information acquired in step S61 has been acquired (step S62). If the learning data set corresponding to the version management information has not yet been acquired (S62: Yes), the data request unit 114 requests the server device 200 to transmit the learning data set (step S63). In this step, the version management information is transmitted to the server device 200.

  When receiving the transmission request from the client apparatus 100, the learning result data transmission unit 217 of the server apparatus 200 transmits the encrypted learning result data corresponding to the version of the version management information received from the client apparatus 100 to the client apparatus 100 (Step S200). S54). The client device 100 downloads the encrypted learning result data from the server device 200.

  When the download of the encrypted learning result data is completed, the learning result data acquisition unit 115 of the client device 100 decrypts the encrypted learning result data using the secret key generated in step S11, and registers the registration ID and the learning data set. Is acquired (step S64). Note that after the completion of this step, a message indicating that a new learning data set has been acquired may be displayed on the display of the client device 100. For example, a message “a new learning data set is available” may be pushed.

  Next, processes such as erasure of encrypted data and execution of inference will be described with reference to FIG.

  When the learning data set is acquired by the client device 100, the data erasure request unit 116 requests the server device 200 to erase the encrypted learning result data (step S65).

  When receiving the erasure request from the client device 100, the data erasure unit 218 of the server device 200 erases the encrypted learning result data (step S55).

  Thereafter, the inference unit 117 of the client device 100 matches the registration ID obtained by decrypting the encrypted learning version data in step S61 and the registration ID obtained by decrypting the encrypted learning result data in step S64. It is determined whether or not to perform (step S66).

  When the two registration IDs match (S66: Yes), the inference unit 117 performs inference using the learning data set acquired in step S64 (step S67). As described above, inference is performed only when the registration IDs coincide with each other, it is possible to reliably perform inference using the latest version of the learning data set.

  When the inference is completed, the output unit 118 of the client device 100 outputs the result of the inference performed by the inference unit 117 (step S68).

  The above processing flow is merely an example, and various other processing flows are assumed. For example, in the above processing flow, the encrypted learning result data is generated in step S51. However, the present invention is not limited to this, and the encrypted learning result data may be generated after receiving a transmission request from the client device 100. In the above processing flow, the transmission request is automatically transmitted. However, the present invention is not limited to this, and the transmission request may be transmitted to the server device 200 after the user of the client device 100 has been approved.

  It should be noted that the operation algorithm of the client device and software such as security software (hereinafter referred to as “software”) are automatically updated to the latest version by the same method as the distribution of the learning result data to the client device by the above processing flow. You may be made to do. Specifically, as in the above-described processing flow, the server device (data server) push-notifies the client device that the software or the like has been updated, and updates the software or the like in response to a transmission request from the client device. Necessary data is encrypted and transmitted to the client device. The client device downloads the encrypted data received from the server device and decrypts it using the secret key, thereby obtaining data necessary for updating the software or the like. In this way, the client device can securely update software and the like automatically.

  As described above, in the deep learning automatic learning system 1, the server device 200 pushes the encrypted learning version data to the client device 100, thereby obtaining new learning result data (learning data set). To be notified. The client device 100 that has received the push notification determines whether or not learning result data corresponding to the version management information has already been acquired. If not, the client device 100 transmits a transmission request to the server device 200. The server device 200 that has received the transmission request transmits the encrypted learning result data to the client device 100. By doing in this way, the client apparatus 100 can acquire the latest learning result data without holding the communication with the server apparatus 200 until the deep learning is completed.

  In the deep learning automatic learning system 1, the version management information and the learning result data are encrypted using a public key and then transmitted to the client device 100. Therefore, the data is securely distributed from the server device to the client device. be able to.

  Further, in the deep learning automatic learning system 1, the server device 200 transmits the encrypted learning version data and the encrypted learning result data only to the client device 100 to which the registration ID is issued (that is, the authenticated client device). Thus, the latest learning result can be provided from the server device to the client device more securely through the Internet.

  Therefore, according to this embodiment, it is not necessary to hold communication between the client device and the server device until deep learning is completed, and the latest learning result can be securely provided from the server device to the client device via the Internet. .

  Further, in the deep learning automatic learning system 1, the server device is configured to transmit learning data to the learning computer when the learning data in its own learning data DB 222 satisfies a predetermined criterion. When learning data in the learning data DB 321 satisfies a predetermined standard, deep learning is performed based on the learning data. For this reason, according to this embodiment, deep learning can be automatically performed according to accumulation of new learning data.

  Based on the above description, those skilled in the art may be able to conceive additional effects and various modifications of the present invention, but the aspects of the present invention are not limited to the above-described embodiments. Various additions, modifications, and partial deletions can be made without departing from the concept and spirit of the present invention derived from the contents defined in the claims and equivalents thereof.

  At least a part of the deep learning automatic learning system described in the embodiment described above may be configured by hardware or software. When configured by software, a program for realizing at least a part of the functions of the deep learning automatic learning system may be stored in a recording medium such as a flexible disk or a CD-ROM and read and executed by a computer. The recording medium is not limited to a removable medium such as a magnetic disk or an optical disk, but may be a fixed recording medium such as a hard disk device or a memory.

  A program that realizes at least a part of the deep learning automatic learning system may be distributed via a communication line (including wireless communication) such as the Internet. Further, the program may be distributed in a state where the program is encrypted, modulated or compressed, and stored in a recording medium via a wired line such as the Internet or a wireless line.

DESCRIPTION OF SYMBOLS 1 Deep learning automatic learning system 100 Client apparatus 110 Control part 111 Key generation part 112 New learning data transmission part 113 Version management information acquisition part 114 Data request part 115 Learning result data acquisition part 116 Data deletion request part 117 Inference part 118 Output part 120 Storage unit 121 New learning data DB
130, 230, 330 Communication unit 200 Server device 210 Control unit 211 Registration ID issuing unit 212 Learning result data encryption unit 213 Version management information encryption unit 214 Push notification unit 215 New learning data storage unit 216 Learning data transmission unit 217 Learning result Data transmission unit 218 Data erasure unit 220 Storage unit 221 Terminal management DB
222 Learning Data DB
223 Learning result DB
300 learning computer 310 control unit 311 deep learning execution unit 320 storage unit 321 learning data DB
322 Learning result DB

Claims (13)

A deep learning automatic learning system comprising a client device and a server device connected via the Internet,
The client device is
A key generation unit that generates a public key and a secret key based on identification information of the client device;
A new learning data transmission unit for transmitting new learning data to the server device,
The server device
A learning result data encryption unit that encrypts learning result data based on the new learning data and generates encrypted learning result data using the public key;
A version management information encryption unit that encrypts version management information for identifying the learning result data using the public key, and generates encrypted learning version data;
A push notification unit for performing push notification of the encrypted learning version data to the client device,
The client device is
A version management information acquisition unit that acquires the version management information by decrypting the encrypted learning version data that has been pushed from the server device using the secret key;
It is determined whether or not learning result data corresponding to the version management information has already been acquired, and if not acquired, a data request unit that requests the server device to transmit the learning result data;
A learning result data acquisition unit that acquires the learning result data by decrypting the encrypted learning result data transmitted from the server device to the client device in response to a request from the data request unit, using the secret key. And a deep learning automatic learning system characterized by further comprising: The server device further includes a registration ID issuing unit that issues a registration ID based on the identification information of the client device,
The deep learning automatic learning system according to claim 1, wherein the push notification unit pushes the encrypted learning version data to the client device using the registration ID. The learning result data encryption unit encrypts the learning result data and the registration ID to generate the encrypted learning result data,
3. The deep learning automatic learning system according to claim 2, wherein the version management information encryption unit generates the encrypted learning version data by encrypting the version management information and the registration ID.   The client device uses the learning result data when a registration ID obtained by decrypting the encrypted learning version data matches a registration ID obtained by decrypting the encrypted learning result data. The deep learning automatic learning system according to claim 3, further comprising an inference unit that performs inference by a structure adaptive deep learning method. The server device
A new learning data storage unit that stores the new learning data received from the client device in a learning data DB;
The learning data transmitting unit according to any one of claims 1 to 4, further comprising: a learning data transmitting unit that transmits the learning data to a learning computer when the learning data in the learning data DB satisfies a predetermined criterion. Deep learning automatic learning system.   6. The deep learning automatic according to claim 5, wherein the predetermined criterion is a case where a certain amount of data is accumulated or an error between the new learning data and the existing learning data exceeds a threshold value. Learning system.   The deep learning automatic learning system according to claim 5, wherein the learning computer includes a deep learning execution unit that performs learning by a structure adaptive deep learning method using the learning data. The client device is
The data erasure request unit that requests the server device to erase the encrypted learning result data after the learning result data is acquired. Deep learning automatic learning system. The client device is
An output unit for outputting a result of inference performed using the learning result data;
The deep learning automatic learning system according to claim 1, further comprising: The server device
The learning result data transmission part which transmits the said encryption learning result data to the said client apparatus, if the transmission request of the said learning result data is received from the said client apparatus, The any one of Claims 1-9 characterized by the above-mentioned. The described deep learning automatic learning system. The server device
The deep learning according to any one of claims 1 to 10, further comprising a data erasure unit that erases the encrypted learning result data when receiving an erasure request for the encrypted learning result data from the client device. Automatic learning system. A client device connected to a server device via the Internet,
A key generation unit that generates a public key and a secret key based on identification information of the client device;
A new learning data transmission unit for transmitting new learning data to the server device;
A version management information acquisition unit that obtains version management information for identifying learning result data by decrypting the encrypted learning version data that has been pushed from the server device using the secret key;
It is determined whether or not learning result data corresponding to the version management information has already been acquired, and if not acquired, a data request unit that requests the server device to transmit the learning result data;
A learning result data acquisition unit for acquiring the learning result data by decrypting the encrypted learning result data transmitted from the server device to the client device in response to the request of the data request unit, using the secret key; ,
A client device comprising: A server device connected to a client device via the Internet,
A learning result data encryption unit that encrypts learning result data and generates encrypted learning result data using a public key received from the client device;
A version management information encryption unit that encrypts version management information for identifying the learning result data and generates encrypted learning version data using the public key;
A push notification unit for pushing notification of the encrypted learning version data to the client device;
A server device comprising:

Images