Information processing method and information processing system
The use of a TEE-based neural network with random number concealment and anonymization methods safeguards machine learning parameters and inference results from unauthorized access, addressing the leakage issue in multi-layer processing environments.
Patent Information
- Application Number
- JP2023079973
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-15
- Publication Date
- 2024-11-27
AI Technical Summary
Existing technologies fail to adequately protect machine learning parameters, which are confidential information, during processing, especially when using multiple processing layers, as they cannot prevent leakage of sensitive data such as personal privacy information or trade secrets.
An information processing method employing a neural network with a Trusted Execution Environment (TEE) that isolates layer parameters and inputs, using random numbers for concealment and anonymization, ensuring secure processing and output in both TEE and normal domains.
This approach effectively protects machine learning parameters and inference results from leakage by maintaining confidentiality during processing, even in high-speed normal areas like GPUs, preventing unauthorized access and information disclosure.
Smart Images

Figure 2024164476000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to a secret information processing method and secret information processing. [Background technology]
[0002] In recent years, there has been progress in schemes for utilizing personal data through collaboration between the public and private sectors. For example, new value is being created by linking diverse information, such as collaboration between public and private services. Statistical analysis such as regression analysis and machine learning such as CNN (Convolutional Neural Network) are considered effective in creating new value by linking diverse information. A platform is also provided that allows various companies to easily run statistical analysis and machine learning using inexpensive public clouds.
[0003] On the other hand, regulations aimed at protecting privacy are becoming stronger. Therefore, there is a demand for technology that allows data holders such as hospitals and banks to use confidential information while keeping it confidential.
[0004] Against this background, the creation of new value by linking various information is being promoted, such as public services in which the public and private sectors share information, such as emergency and security companies, and private services such as pharmaceuticals, insurance, transportation, and information banks that create synergy with public services. Statistical analysis such as regression analysis and analytical processing such as machine learning are considered effective in analyzing various data to create new value. Furthermore, in order to realize these services, confidential information processing technology is required that allows public and private service providers to analyze and utilize confidential information, such as personal information, held by data holders such as hospitals and banks while protecting the privacy of individuals.
[0005] Conventional technologies have made it possible to perform specific processing on encrypted data, thereby enabling the use of confidential information in a concealed state. However, the processes that can be performed with conventional technologies are limited to basic processing such as search and order comparison, and the flexibility of the processing is limited. Therefore, there is a demand for technology that can handle advanced statistical processing and image processing using machine learning such as CNN, which require a high degree of processing flexibility.
[0006] In recent years, CPUs (Central Processing Units) are equipped with TEE (Trusted Execution Environment) functions. The TEE function is considered to be effective for processing confidential information. The TEE function provides a trusted area where information cannot be read even if the administrator privileges of the OS running on the computer are stolen. By decrypting and processing encrypted data within the trusted area, secure and advanced processing becomes possible.
[0007] A technology that utilizes the trust region of the TEE function is described in Patent Document 1. Patent Document 1 discloses a privacy-enhanced deep learning system that reduces information leakage of confidential input data in an inference pipeline.
[0008] The technology described in Patent Document 1 utilizes the hierarchical structure of a neural network to divide each deep learning model into a FrontNet that processes in the trust region and a BackNet that processes in the normal region.
[0009] In the technology described in Patent Document 1, a terminal used by an end user transmits encrypted input and encrypted FrontNet to a system. The technology described in Patent Document 1 uses the TEE function on the cloud infrastructure to perform deep learning processing in the FrontNet enclave (an isolated execution process in the TEE function) and protects the confidentiality and integrity of user input with encryption.
[0010] On the other hand, in the technology described in Patent Document 1, there is a lack of secure enclaves for BackNet inference calculations, and if the cloud machine is equipped with a deep learning acceleration chip such as a GPU (Graphical Processing Unit), the benefits of improved performance can be obtained.
[0011] The technology described in Patent Document 1 utilizes the hierarchical structure of a neural network to determine a model-specific optimal split point that balances privacy protection and performance requirements, and partitions each Deep Learning model into a FrontNet and a BackNet. In the technology described in Patent Document 1, a Deep Learning inference system measures the similarity for each set of intermediate data in each layer, and determines a split point to be used for splitting into two partitions using a selected subset of intermediate data that is most similar to the input information.
[0012] Furthermore, a technology that utilizes the trust region of the TEE function is described in Non-Patent Document 1. Non-Patent Document 1 discloses a method for concealing confidential input data that reduces information leakage of confidential input data input to a GPU in an inference pipeline that utilizes a GPU that speeds up linear arithmetic processing in deep learning. [Prior art documents] [Patent documents]
[0013] [Patent Document 1] US Patent Application Publication No. 2020 / 0082259 [Non-patent literature]
[0014] [Non-Patent Document 1] F. Tramer and D. Boneh, “Slalom: Fast, verifiable and private execution of neural networks in trusted hardware,” in International Conference on Learning Representations (ICLR), 2019. Summary of the Invention [Problem to be solved by the invention]
[0015] However, the system described in Patent Document 1 may not be able to prevent leakage of confidential information during deep learning processing of unknown input information if the unknown input information contains confidential information such as personal privacy information or trade secrets that have not been learned in the learning stage. Specifically, BackNet processes the machine learning parameters that constitute the deep learning model in the normal domain, so it cannot prevent leakage of the deep learning model if the machine learning parameters are confidential information.
[0016] In addition, the concealment method described in Non-Patent Document 1 processes the machine learning parameters used in the linear arithmetic processing, which is part of the deep learning model, on a normal GPU without concealing them, so that it cannot prevent the leakage of confidential information if the machine learning parameters are confidential information.
[0017] Therefore, one aspect of the present invention aims to protect machine learning parameters, which are confidential information, during processing such as machine learning using multiple processing layers. [Means for solving the problem]
[0018] In order to solve the above problems, one aspect of the present invention employs the following configuration: An information processing method performed by an information processing system that executes information processing using a neural network consisting of multiple i-th layers (i is a natural number between 1 and N), the information processing system having a processor and a memory, the memory being a Trusted Execution Environment (TEE) isolated from a normal area. the processor executes, for each of the i-th layers, a parameter concealment process for multiplying the i-th layer parameter by the i-th layer parameter protection random number to generate an i-th layer concealment parameter by concealing the i-th layer parameter in the isolation area, and outputs the i-th layer concealment parameter to the normal area; and the i-layer input data protection random number to the i-layer input data to generate i-layer concealed input data by concealing the i-layer input data, outputting the i-layer concealed input data to the normal area as the i-layer input data, acquiring from the normal area a result of the i-layer processing executed on the i-layer concealed input data based on the i-layer concealment parameter in the normal area and executing a parameter concealment influence removal process to divide the execution result by the i-layer parameter protection random number, executing an input concealment influence removal process to subtract the i-layer input data concealment removal parameter from the execution result of the parameter concealment influence removal process, and setting the execution result of the input concealment influence removal process as i-layer output data output from the i-layer. Effect of the Invention
[0019] According to one aspect of the present invention, machine learning parameters, which are confidential information, can be protected during processing such as machine learning using multiple processing layers.
[0020] Problems, configurations, and effects other than those described above will become apparent from the following description of the embodiments. [Brief description of the drawings]
[0021] [Figure 1] 1 is a block diagram showing an example of a system configuration of a secret information processing system. [Diagram 2] 1 is a block diagram showing an example of a hardware configuration of a secret information processing system. [Diagram 3] 1 is an explanatory diagram showing an example of a functional configuration and a data processing flow for confidential inference processing; [Figure 4] FIG. 13 is a sequence diagram showing an example of a data processing flow of a confidential inference preparation process. [Diagram 5] FIG. 13 is a sequence diagram showing an example of a data processing flow of a confidential inference execution process. [Figure 6] 13 is a flowchart illustrating an example of a parameter concealment process. [Figure 7] 13 is a flowchart illustrating an example of an input concealment effect removal parameter generation process. [Figure 8] 13 is a flowchart illustrating an example of a concealed input data generation process. [Figure 9] 13 is a flowchart illustrating an example of a parameter concealment effect removal process. [Figure 10] 13 is a flowchart illustrating an example of an input concealment influence removal process. [Figure 11] FIG. 13 is a dialog diagram showing an example of a machine learning parameter anonymization setting information input screen. [Figure 12] 13 is a table illustrating an example of a layer processing concealment setting information table. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0022] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. In this embodiment, the same components are generally designated by the same reference numerals, and repeated explanations are omitted. Note that this embodiment is merely an example for realizing the present invention, and does not limit the technical scope of the present invention.
[0023] (1) Example of configuration of confidential information processing system 1 1 is a block diagram showing an example of a system configuration of a confidential information processing system 1. The confidential information processing system 1 is entrusted with encrypted machine learning parameters and the like from a parameter holder that holds confidential information that is a trade secret, such as machine learning parameters. The confidential information processing system 1 is also entrusted with encrypted data from a data holder that holds confidential information, such as personal information. The confidential information processing system 1 processes the entrusted machine learning parameters and the like and data in response to a processing request from a processing result user while keeping the confidential information confidential, and provides the processing result to the processing result user.
[0024] The secret information processing system 1 has, for example, a parameter holder terminal 50 used by a parameter holder to encrypt and deposit machine learning parameters, etc. The secret information processing system 1 also has a data holder terminal 10 used by a data holder to encrypt and deposit data. The secret information processing system 1 also has a processing result user terminal 20 where a processing result user generates and transmits a processing request and refers to the processing result. The secret information processing system 1 also has a secret information processing server 30 that executes the processing of the processing request received from the processing result user on the data deposited from the data holder while keeping it confidential based on the machine learning parameters, etc. deposited from the parameter holder, and transmits the processing result. The parameter holder terminal 50, the data holder terminal 10, the processing result user terminal 20, and the secret information processing server 30 are connected to each other by a network 40 such as the Internet.
[0025] The secret information processing system 1 may include a plurality of parameter holder terminals 50, a plurality of data holder terminals 10, a plurality of processing result user terminals 20, or a plurality of secret information processing servers. The parameter holder terminal 50 and the data holder terminal 10 may be the same terminal, or the data holder terminal 10 and the processing result user terminal 20 may be the same terminal.
[0026] 2 is a block diagram showing an example of a hardware configuration of the secret information processing system 1. The data holder terminal 10 is, for example, a computer such as a personal computer, a smartphone, or a server device, or a virtual computer. The data holder terminal 10 is configured by a computer including a CPU (Control Processing Unit) 11, a memory 12, an auxiliary storage device 13, a network interface 14, a display device 15, and an input device 16, which are connected to each other by, for example, an internal communication line.
[0027] The CPU 11 includes a processor and executes programs stored in the memory 12. The memory 12 includes a ROM (Read Only Memory), which is a non-volatile storage element, and a RAM (Random Access Memory), which is a volatile storage element. The ROM stores immutable programs (e.g., a Basic Input / Output System (BIOS)) and the like. The RAM is a high-speed, volatile storage element such as a DRAM (Dynamic Random Access Memory), and temporarily stores programs executed by the CPU 11 and data used when the programs are executed.
[0028] The CPU 11 is a TEE (Trusted Execution Environment)-enabled CPU having a TEE function, and the memory 12 has a TEE trusted area 17 which is a hardware trusted area secured on the memory by the TEE-enabled CPU 21 and is isolated from other areas on the memory. In the first embodiment, the CPU 11 does not need to have the TEE function, and the memory 12 does not need to have the TEE trusted area 17. The memory amount that can be handled in the TEE trusted area 17 may be limited to be smaller than the memory amount in the normal area (for example, about 128 MB). In this case, when the CPU 11 executes a process on the TEE trusted area 17, the processing speed decreases compared to when the CPU 11 executes a process on the normal area.
[0029] The auxiliary storage device 13 is a large-capacity non-volatile storage device such as a magnetic storage device (HDD (Hard Disk Drive)) or a flash memory (SSD (Solid State Drive)), and stores the programs executed by the CPU 11 and data used when the programs are executed. That is, the programs are read from the auxiliary storage device 13, loaded into the memory 12, and executed by the CPU 11.
[0030] The display device 15 is a device such as a display or printer that outputs the results of program execution in a format that can be viewed by an operator. The input device 16 is a device such as a keyboard or mouse that receives input from an operator. The network interface 14 controls communication with other devices according to a predetermined protocol. The network interface 14 may also include a serial interface such as a USB.
[0031] The program executed by the CPU 11 may be stored in advance in a non-volatile auxiliary storage device 13, which is a computer-readable non-transient storage medium, or may be a removable medium (CD-ROM, flash memory, etc.), which is a non-transient storage medium. Alternatively, the program may be provided to the data holder terminal 10 from a non-transient storage device of another device via a network and stored in the auxiliary storage device 13. For this reason, it is preferable that the data holder terminal 10 has an interface for reading data from removable media. This also applies to the processing result user terminal 20 and the confidential information processing server 30.
[0032] In this embodiment, some or all of the functions of the functional units executed by the CPU and GPU may be realized by hardware such as an Application Specific Integrated Circuit (ASIC) or a Field Programmable Gate Array (FPGA).
[0033] In this embodiment, the information used by the confidential information processing system 1 may be expressed in any data structure regardless of the data structure, and the information may be stored in a data structure appropriately selected from, for example, a list, a table, a database, or a queue.
[0034] The parameter holder terminal 50 is, for example, a computer such as a personal computer, a smartphone, or a server device, or a virtual computer. The parameter holder terminal 50 is configured by a computer including a CPU 51, a memory 52, an auxiliary storage device 53, a network interface 54, a display device 55, and an input device 56, which are connected to each other by an internal communication line.
[0035] The CPU 51, memory 52, auxiliary storage device 53, network interface 54, display device 55, and input device 56 are the same hardware as the CPU 11, memory 12, auxiliary storage device 13, network interface 14, display device 15, and input device 16. However, the CPU 51 does not need to have a TEE function, and the memory 52 does not need to have a trusted region.
[0036] The processing result user terminal 20 is, for example, a computer such as a personal computer, a smartphone, or a server device, or a virtual computer. The processing result user terminal 20 is configured by a computer including a CPU 21, a memory 22, an auxiliary storage device 23, a network interface 24, a display device 25, and an input device 26, which are connected to each other by an internal communication line, for example.
[0037] The CPU 21, the memory 22, the auxiliary storage device 23, the network interface 24, the display device 25, and the input device 26 are the same hardware as the CPU 11, the memory 12, the auxiliary storage device 13, the network interface 14, the display device 15, and the input device 16. However, the CPU 21 does not need to have a TEE function, and the memory 22 does not need to have a trusted region.
[0038] The confidential information processing server 30 is, for example, a computer such as a personal computer, a smartphone, or a server device, or a virtual computer. The confidential information processing server 30 is configured by a computer including a CPU 31, a memory 32, an auxiliary storage device 33, a network interface 34, a display device 35, an input device 36, and a GPU (Graphics Processing Unit) 39, which are connected to each other by an internal communication line such as a bus.
[0039] The CPU 31, the memory 32, the auxiliary storage device 33, the network interface 34, the display device 35, and the input device 36 are the same hardware as the CPU 11, the memory 12, the auxiliary storage device 13, the network interface 14, the display device 15, and the input device 16. However, the CPU 31 of the secret information processing server 30 is a TEE-compatible CPU having a TEE function. The memory 32 of the secret information processing server 30 has a TEE trusted area 38, which is a hardware trusted area secured on the memory by the TEE-compatible CPU 31 and is an execution area isolated from other areas on the memory. The TEE trusted area 38 is also called an isolated area isolated from the normal area. The GPU 39 includes a processor and executes a program stored in the memory 32 at high speed using an area (normal area) outside the TEE trusted area 38 of the memory 32. Note that the secret information processing server 30 does not need to have the GPU 39.
[0040] In this embodiment, the parameter holder terminal 50, the data holder terminal 10, and the confidential information processing server 30 have a TEE function, but a method other than the TEE function may be adopted that enables calculations to be performed in a secure execution area isolated from the normal area on the memory.
[0041] Some or all of the processes executed by the CPU 11, CPU 21, CPU 31, CPU 51, and GPU 39 may be executed by hardware such as an Application Specific Integrated Circuit (ASIC) or a Field Programmable Gate Array (FPGA).
[0042] The parameter holder terminal 50, the data holder terminal 10, the processing result user terminal 20, and the secret information processing server 30 are all computer systems that are physically configured on one computer or on multiple computers that are logically or physically configured. However, the parameter holder terminal 50, the data holder terminal 10, the processing result user terminal 20, and the secret information processing server 30 may run in separate threads on the same computer, or may run on virtual computers built on multiple physical computer resources.
[0043] The network 40 is a communication network such as the Internet, an intranet such as a company network, or a satellite line, using wired or wireless communication.
[0044] (2) Example of Functional Configuration of Secret Information Processing System 1 and Example of Data Processing in Secret Inference Processing FIG. 3 is an explanatory diagram illustrating an example of a functional configuration and a data processing flow for confidential inference processing.
[0045] First, the entire confidential inference process executed by the confidential information processing system 1 according to this embodiment will be described. In the confidential information processing system 1, the parameter holder terminal 50 first encrypts machine learning parameter information including confidential information such as each layer processing concealment setting information, and transmits the information to the confidential information processing server 30. The each layer processing concealment setting information is information that specifies an area in which the machine learning parameters of a neural network such as a CNN (Convolutional Neural Network) and the processing of each layer are executed while being concealed.
[0046] Next, the confidential information processing server 30 decrypts the machine learning parameter information in the TEE trust region 38, and stores the parameter concealment setting information in the TEE trust region 38. Then, the parameter concealment processing and the input concealment influence removal parameter generation processing are executed in the TEE trust region 38, the concealed machine learning parameter information is output to and stored in the normal region, and the input concealment removal parameter is stored in the TEE trust region 38.
[0047] Next, in the secret information processing system 1, the data holder terminal 10 encrypts the input data including confidential information such as personal information and transmits the encrypted data to the secret information processing server 30.
[0048] Next, the confidential information processing server 30 decrypts the input data in the TEE trusted domain 38, and executes the processing of each layer of the Deep Learning inference processing including multiple processing layers in the TEE trusted domain 38 or the normal domain depending on the result of the anonymization setting confirmation processing based on the parameter anonymization setting information, and encrypts the inference result and transmits it to the processing result user terminal 20. Then, the processing result user terminal 20 decrypts the received encrypted inference result and acquires the inference result.
[0049] Hereinafter, an example of a functional configuration and an example of a data processing flow will be described in detail with reference to FIG.
[0050] The confidential information processing server 30 has, for example, a normal CPU and memory, and a normal area 310 on the memory 32 for performing high-speed processing using an acceleration chip such as a GPU. The confidential information processing server 30 also has a TEE trusted area 38 on the memory 32 which is a highly secure and isolated processing execution area provided by the TEE function of the CPU or the like.
[0051] The normal region 310 includes, for example, an i-th layer processing unit 312. The i-th layer processing unit 312 executes processing of each layer in a multi-layered machine learning such as CNN (hereinafter, the description will be given taking CNN as an example). Note that CNN is merely an example, and this embodiment can also be applied to other neural networks such as RNN (Recurrent Neural Network).
[0052] The i-th layer processing unit 312 performs processing of a layer of the anonymized input data 516 based on the anonymized machine learning parameters 515 to generate anonymized intermediate data 517 and inputs it to the parameter anonymization influence removal processing unit 533 of the TEE trust region 38.
[0053] The TEE trusted domain 38 includes, for example, an i-th layer processing unit 326, a decryption processing unit 522, a decryption processing unit 322, a parameter concealment processing unit 534, an input concealment influence elimination parameter generation processing unit 535, and an i-th layer concealment setting confirmation processing unit 536. The TEE trusted domain 38 also includes a concealed input data generation processing unit 525, a parameter concealment influence elimination processing unit 533, an input concealment influence elimination processing unit 538, a random number generation unit 539, and an encryption processing unit 329.
[0054] The i-th layer processing unit 326 executes processing of the CNN layer. The random number generation unit 539 executes generation of random numbers used for various concealment processes. The decryption processing unit 522 decrypts the encrypted machine learning parameter information 521 (including machine learning parameters and each layer processing concealment setting information) using the encryption key 532 stored in the TEE trust region 38. The decryption processing unit 522 stores the decrypted machine learning parameter information 521 in the TEE trust region 38 as the machine learning parameters 523 and each layer processing concealment setting information 524. The decryption processing unit 322 decrypts the encrypted data 103 using the encryption key 331 stored in the TEE trust region 38, and inputs it to the i-th layer concealment setting confirmation processing unit 536.
[0055] Parameter concealment processing unit 534 receives machine learning parameters 523 of each layer included in machine learning parameters 523 and parameter protection random numbers 528 that protect different parameters for each layer generated by random number generation unit 539, and generates concealed machine learning parameters 515. Parameter concealment processing unit 534 stores generated concealed machine learning parameters 515 in normal area 310.
[0056] The input concealment influence elimination parameter generation processing unit 535 receives as input the machine learning parameters 523 of each layer and the input data protection random numbers 526 that protect the input data of each layer, which differ for each layer and are generated by the random number generation unit 539, and generates an input concealment influence elimination parameter 529. The input concealment influence elimination parameter generation processing unit 535 stores the generated input concealment influence elimination parameter 529 in the TEE trust region 38.
[0057] The i-th layer concealment setting confirmation processing unit 536 receives input data of the layer (input data 323 or intermediate data 530) and the layer number i of the processing layer. The i-th layer concealment setting confirmation processing unit 536 refers to information on the area in which the i-th layer processing of the layer number is executed in each layer processing concealment setting information 524, and if the reference result indicates a "trusted area", it inputs the input data to the i-th layer processing unit 326. On the other hand, if the reference result indicates a "normal area", the i-th layer concealment setting confirmation processing unit 536 inputs the input data to the concealed input data generation processing unit 525.
[0058] The i-th layer processing unit 326 executes processing of a layer in a multi-layer machine learning structure such as CNN within the TEE trust region 38, and outputs the execution result as intermediate data 327 if it is not the final layer, and outputs the execution result as an inference result 328 if it is the final layer.
[0059] The concealed input data generation processor 525 receives the input data of the layer in question, reads the random number of the layer in the input data protection random number 526, and reads the temporary input data protection random number 537 generated by the random number generator 539 for each concealed input data generation process. The concealed input data generation processor 525 executes a process of concealing the input data based on these random numbers, and outputs the concealed input data 516. Note that the concealment may be performed without using the temporary input data protection random number 537 in the concealed input data generation processor 525.
[0060] The parameter concealment influence elimination processor 533 executes a parameter concealment influence elimination process for removing the influence of parameter concealment from the concealment intermediate data 517 based on the parameter protection random number 528 of each layer, for the input concealment intermediate data 517. The parameter concealment influence elimination processor 533 inputs the execution result of the parameter concealment influence elimination process to the input concealment influence elimination processor 538.
[0061] The input concealment influence removal processor 538 executes an input concealment influence removal process for removing the influence of concealment of the input data to the layer based on the input concealment influence removal parameter 529 of the layer and the temporary input data protection random number 537. The input concealment influence removal processor 538 outputs an inference result 531 if the layer is other than the final layer, and inputs the inference result 531 to the encryption processor 329 if the layer is the final layer. Note that, if concealment is performed without using the temporary input data protection random number 537 in the concealment input data generation process of the layer, the temporary input data protection random number 537 is not used in the input concealment influence removal process. Also, the temporary input data protection random number 537 may be generated in advance at the stage of concealment inference preparation process.
[0062] The encryption processing unit 229 performs encryption on the input inference result 328 or inference result 531 using the encryption key 332 stored in the TEE trust area 38, and outputs the encrypted inference result 330 via the normal area 310.
[0063] The parameter holder terminal 50 also includes an encryption processing unit 502 that encrypts the machine learning parameter information 501, etc., using an encryption key 504 stored in the memory 52, the auxiliary storage device 53, and the TEE trusted area 57. The data holder terminal 10 includes an encryption processing unit 102 that encrypts data using an encryption key 104 stored in the memory 12, the auxiliary storage device 13, or the TEE trusted area 17. The processing result user terminal 20 includes a decryption processing unit 202 that decrypts the encrypted inference result, etc., using an encryption key 204 stored in the memory 22 or the auxiliary storage device 23. Any two or all of the parameter holder terminal 50, the data holder terminal 10, and the processing result user terminal may be the same terminal, in which case any two or all of the encryption processing unit 502, the encryption processing unit 102, and the decryption processing unit 202 are provided in the same terminal.
[0064] Moreover, the GPU 39 operates according to the i-th layer processing program loaded in the normal area of the memory 32, thereby executing processing as the i-th layer processing unit 312. Moreover, the TEE-enabled CPU 51 operates according to the encryption processing program loaded in the TEE trusted area 57 of the memory 52, thereby executing processing as the encryption processing unit 502. Moreover, the TEE-enabled CPU 11 operates according to the encryption processing program loaded in the TEE trusted area 17 of the memory 12, thereby executing processing as the encryption processing unit 102. Moreover, the CPU 21 operates according to the inference result display program loaded in the memory 22, thereby executing processing as the decryption processing unit 202 by operating according to the decryption processing program loaded in the memory 22.
[0065] Hereinafter, a data processing flow of the confidential inference preparation processing according to the first embodiment will be described with reference to Fig. 3 and Fig. 4. Fig. 4 is a sequence diagram showing an example of a data processing flow of the confidential inference preparation processing. First, the CPU 51 of the parameter holder terminal 50 displays the machine learning parameter concealment setting information input screen shown in Fig. 11 on the display device 55 in accordance with the machine learning parameter concealment setting information input screen display program loaded into the memory 52. Then, the CPU 51 accepts input by the user of the name of the target machine learning parameter, the number of layers of the target machine learning parameter, whether the target machine learning parameter needs to be concealed, and each layer processing concealment setting information of the target machine learning parameter (step S401).
[0066] Next, the encryption processing unit 502 of the parameter holder terminal 50 reads out the machine learning parameters with the names based on the user's input shown in FIG. 11 from the auxiliary storage device 53 and encrypts them (step S402). Next, the encryption processing unit 502 reads out the layer processing concealment setting information based on the user's input from the memory 52 and encrypts it (step S403). Next, the encryption processing unit 502 registers the machine learning parameters and each layer processing concealment setting information by transmitting the encrypted machine learning parameter information to the confidential information processing server 30 (step S404). When the confidential information processing server 30 receives the encrypted data 503, it inputs it to the TEE trust area 38. The decryption processing unit 522 of the TEE trust area 38 decrypts the encrypted data 503 to generate the machine learning parameters 523 and each layer processing concealment setting information 524 having the data structure shown in FIG. 12 (step S405).
[0067] Next, the parameter concealment processing unit 534 executes the parameter concealment processing shown in FIG. 6, outputs the execution result to the general area 310 (step S406), and stores it in the general area 310 as the concealed machine learning parameter 515 (step S407).
[0068] Next, the input concealment effect elimination parameter generation processing unit 535 executes the input concealment effect elimination parameter generation processing shown in FIG. 7 (step S408), stores the execution result as the input concealment effect elimination parameter 529 in the TEE trust region 38, and transmits a registration completion response to the parameter holder terminal 50 (step S409).
[0069] Hereinafter, the data processing flow of the secret inference execution process according to the first embodiment will be described with reference to Fig. 3 and Fig. 5. Fig. 5 is a sequence diagram showing an example of the data processing flow of the secret inference execution process. First, the encryption processing unit 102 of the data holder terminal 10 generates encrypted data 103 by encrypting data 101 including confidential information (step S501), and transmits the encrypted data 103 to the secret information processing server 30 to register the data (step S502). When the secret information processing server 30 receives the encrypted data 103, it inputs it to the TEE trust area 38. The decryption processing unit 322 of the TEE trust area 38 decrypts the encrypted data 103 to generate input data 323 (step S503), and initializes the value of a variable i indicating the layer number of the process to 1 (step S504).
[0070] The i-th layer concealment setting confirmation processing unit 536 reads out information on the region in which the i-th layer processing is to be performed in the each layer processing concealment setting information 524 (step S505). After that, the i-th layer concealment setting confirmation processing unit 536 transitions to step S506, and if the readout result indicates a "trusted region" (step S506 No), inputs the input data to the i-th layer processing unit 326 to read out the machine learning parameter information of the i-th layer (step S507). Then, the i-th layer concealment setting confirmation processing unit 536 executes the i-th layer processing within the TEE trusted region 38 (step S508).
[0071] On the other hand, if the read result indicates the "normal area" (Yes in step S506), the i-th layer anonymization setting confirmation processing unit 536 inputs the input data to the anonymized input data generation processing unit 525 and executes the anonymized input data generation processing shown in Fig. 8 (step S509). Then, the i-th layer anonymization setting confirmation processing unit 536 outputs the i-th layer anonymized input data resulting from the execution to the normal area 310 (step S510).
[0072] Furthermore, the i-th layer concealment setting confirmation processing unit 536 transitions to step S517 after step S508, and if the i-th layer is not the last processing layer (step S517 No), it increments i by using the output data of the i-th layer processing as the input data of the next layer, and transitions to step S505. On the other hand, if the i-th layer is the last processing layer (step S517 Yes), the encryption processing unit 329 encrypts the output data of the i-th layer processing as an inference result, and transmits the processing result to the user terminal 20 (step S518).
[0073] Next, the i-th layer processing unit 312 of the normal region 310 reads the i-th layer concealed input data 516 as an input (step S511). Next, the i-th layer processing unit 312 executes the i-th layer processing, and inputs the execution result as concealed intermediate data 517 to the parameter concealment influence removal processing unit 533 of the TEE trust region 38 (step S512).
[0074] For example, the i-th layer processing is performed using formula (1). Formula (1) represents an operation for multiplying the i-th layer anonymized input data x_i′ matrix and the i-th layer anonymized machine learning parameter W_i′ matrix to calculate the execution result y_i′.
[0075]
number
[0076] Next, the parameter concealment influence removal processing unit 533 executes the parameter concealment influence removal processing shown in FIG. 9, and inputs the parameter concealment influence removal intermediate data of the i-th layer, which is the execution result, to the input concealment influence removal processing unit 538 (step S513).
[0077] Next, the input concealment influence elimination processing unit 538 executes the input concealment influence elimination processing shown in Fig. 10 (step S514), and transitions to step S515. If the i-th layer is not the last processing layer (step S515 No), the input concealment influence elimination processing unit 538 increments i by taking the output data of the i-th layer, which is the execution result, as the input data of the next layer, and transitions to step S505. On the other hand, if the i-th layer is the last processing layer (step S515 Yes), the input concealment influence elimination processing unit 538 inputs the output data of the i-th layer, which is the execution result, to the encryption processing unit 329 as the inference result. Then, the input concealment influence elimination processing unit 538 transmits the encrypted inference result to the processing result user terminal 20 (step S518).
[0078] 6 is a flowchart showing an example of the parameter concealment process. The parameter concealment process unit 534 of the TEE trust region 38 first initializes the value of i to 1 (step S601).
[0079] Next, the parameter concealment processing unit 534 reads information on the region in which the i-th layer processing is to be performed from the each layer processing concealment setting information 524 (step S602), and transitions to step S603. If the read result indicates a "trusted region" (step 603 No), the parameter concealment processing unit 534 transitions to step S607. On the other hand, if the read result indicates a "normal region" (step 603 Yes), the parameter concealment processing unit 534 transitions to step S604.
[0080] Next, the parameter concealment processing unit 534 reads the i-th layer machine learning parameters from the machine learning parameters 523 (step S604). Then, the parameter concealment processing unit 534 acquires the random numbers for protecting the i-th layer parameters generated by the random number generation unit 539, and adds and stores the random numbers to the parameter protection random numbers 528 (step S605).
[0081] For example, the machine learning parameters of the i-th layer are expressed using formula (2). W_i in formula (2) represents a matrix with p rows and q columns. Also, for example, the i-th layer parameter protection random number for protecting the parameters of the i-th layer is expressed using formula (3). β_i in formula (3) is a scalar random number.
[0082]
number
[0083]
number
[0084] Next, the parameter concealment processing unit 534 multiplies the machine learning parameter of the i-th layer by the machine learning parameter of the i-th layer to calculate the concealed machine learning parameter of the i-th layer (step S606).
[0085] For example, the concealed machine learning parameters are calculated using formula (4). W_i' in formula (4) is an example of an i-th layer concealment parameter that is an anonymized i-th layer machine learning parameter, and represents a matrix with p rows and q columns.
[0086]
number
[0087] Next, if the i-th layer is not the last processing layer (step S607 No), the parameter anonymization processor 534 increments i (step S608) and transitions to step S602. On the other hand, if the i-th layer is the last processing layer (step S607 Yes), the parameter anonymization processor 534 outputs the calculated anonymized machine learning parameters to the general area 310 and stores them as anonymized machine learning parameters 515, and ends the parameter anonymization process.
[0088] 7 is a flowchart showing an example of an input concealment influence elimination parameter generation process. The input concealment influence elimination parameter generation processor 535 of the TEE trust region 38 first initializes the value of i to 1 (step S701).
[0089] Next, the input concealment influence removal parameter generation processing unit 535 reads information on the region where the i-th layer processing is to be executed from the each layer processing concealment setting information 524 (step S702), and transitions to step S703. If the read result indicates a "trusted region" (step 703 No), the input concealment influence removal parameter generation processing unit 535 transitions to step S707. On the other hand, if the read result indicates a "normal region" (step 703 Yes), the input concealment influence removal parameter generation processing unit 535 transitions to step S704.
[0090] Next, the input concealment influence elimination parameter generation processing unit 535 acquires the input data protection random number for protecting input data of the i-th layer generated by the random number generation unit 539, and adds and stores the input data protection random number 526 of the TEE trust region 38 (step S704). Next, the input concealment influence elimination parameter generation processing unit 535 reads the machine learning parameters of the i-th layer from the machine learning parameters 523 (step S705).
[0091] For example, the input data protection random number of the i-th layer is expressed using formula (5). r_i in formula (5) is an example of the input data protection random number of the i-th layer, and represents a matrix of m rows and n columns, and each element of the matrix is a different random number.
[0092]
number
[0093] Furthermore, for example, the input data protection random number of the i-th layer is expressed using formula (6). r_i^(j) in formula (6) is an example of the input data protection random number of the i-th layer, and represents one matrix group with m rows and n columns, and each element of the matrix is a different random number.
[0094]
number
[0095] Next, the input concealment influence elimination parameter generation processing unit 535 multiplies the input data protection random number of the i-th layer by the machine learning parameter of the i-th layer to calculate the input concealment influence elimination parameter of the i-th layer (step S706).
[0096] For example, the input concealment effect removal parameter of the i-th layer is calculated using formula (7). u_i in formula (7) is an example of an i-layer input data concealment removal parameter that removes the effect of concealment from the input data of the i-th layer, and is calculated by the product of the input data protection random number r_i matrix of the i-th layer and the machine learning parameter W_i matrix of the i-th layer.
[0097]
number
[0098] Furthermore, for example, an input concealment effect removal parameter for the i-th layer is calculated using formula (8). u_i^(j) in formula (8) is an example of an input concealment effect removal parameter group for the i-th layer, and is calculated by multiplying the input data protection random number r_i^(j) matrix for the i-th layer and the machine learning parameter W_i matrix for the i-th layer.
[0099]
number
[0100] Next, in step S707, if the i-th layer is not the last processing layer (step S707 No), the input concealment influence elimination parameter generation processing unit 535 increments i (step S708) and transitions to step S702. On the other hand, if the i-th layer is the last processing layer (step S707 Yes), the input concealment influence elimination parameter generation processing unit 535 saves the calculated input concealment influence elimination parameter group as the input concealment influence elimination parameters 529 of the TEE trust region 38 (step S707). When step S707 ends, the input concealment influence elimination parameter generation processing unit 535 ends the input concealment influence elimination parameter generation processing.
[0101] 8 is a flowchart showing an example of the concealed input data generation process. The concealed input data generation unit 525 of the TEE trust region 38 first reads out input data of the i-th layer, which is the relevant layer (step S801).
[0102] For example, the input data of the i-th layer is expressed using formula (9). x_i in formula (9) is an example of the input data of the i-th layer, and represents a matrix with m rows and n columns, and each element of the matrix is a numerical value.
[0103]
number
[0104] Next, the concealed input data generation processing unit 525 reads the input data protection random number of the i-th layer from the input data protection random number 526 (step S802), obtains a new random number from the random number generation unit 539, and stores it as a temporary input data protection random number 537 (step S803).
[0105] For example, the temporary input data protection random number of the i-th layer is expressed by using the formula (10). α_i in the formula (10) is an example of the temporary input data protection random number of the i-th layer, and represents a single random number.
[0106]
number
[0107] Furthermore, for example, the temporary input data protection random numbers of the i-th layer are expressed using equation (11). α_i^(j) in equation (11) is l different random numbers.
[0108]
number
[0109] Next, the concealed input data generation processor 525 multiplies the temporary input data protection random number 537 by the input data protection random number of the i-th layer to calculate the temporary input data protection random number of the i-th layer (step S804).
[0110] Next, the anonymized input data generation processing unit 525 adds the input data of the i-th layer and the temporary input data protection random number of the i-th layer to calculate the anonymized input data of the i-th layer, outputs it to the normal area 310 and stores it as anonymized input data 516, and terminates the processing (step S805).
[0111] For example, the concealed input data of the i-th layer is calculated using equation (12). x_i′ in equation (12) is an example of concealed input data, and represents a matrix with m rows and n columns.
[0112]
number
[0113] Furthermore, for example, the concealed input data of the i-th layer is calculated using equation (13). x_i′ in equation (13) is an example of concealed input data, and represents a matrix with m rows and n columns.
[0114]
number
[0115] 9 is a flowchart showing an example of the parameter concealment influence removal process. The parameter concealment influence removal process unit 533 of the TEE trust region 38 first reads the concealment intermediate data 517 of the i-th layer, which is the layer in question (step S901). Next, the parameter concealment influence removal process unit 533 reads the parameter protection random number of the i-th layer, which is the layer in question, from the parameter protection random number 528 (step S902). Next, the parameter concealment influence removal process unit 533 divides the concealment intermediate data 517 of the i-th layer by the parameter protection random number of the i-th layer to calculate the parameter concealment influence removal intermediate data of the i-th layer and store it in the TEE trust region 38 (step S903).
[0116] For example, the parameter concealment effect removed intermediate data of the i-th layer is calculated using formula (14). y_î in formula (14) is an example of the parameter concealment effect removed intermediate data concealment input data of the i-th layer, and represents a matrix.
[0117]
number
[0118] 10 is a flowchart showing an example of the input concealment influence removal process. The input concealment influence removal processing unit 538 of the TEE trust region 38 first reads out the parameter concealment influence removal intermediate data of the i-th layer, which is the layer in question (step S1001). Next, the input concealment influence removal processing unit 538 reads out the temporary input data protection random number 537 (step S1002). Next, the input concealment influence removal processing unit 538 reads out the input concealment influence removal parameters of the i-th layer, which is the layer in question, from the input concealment influence removal parameters 529 (step S1003).
[0119] Next, the input concealment influence elimination processing unit 538 multiplies the temporary input data protection random number 537 by the input concealment influence elimination parameter of the i-th layer to calculate the temporary input concealment influence elimination parameter of the i-th layer (step S1004).
[0120] Next, the input concealment influence removal processing unit 538 outputs the result obtained by subtracting the temporary input concealment influence removal parameter of the i-th layer from the concealment parameter influence removal intermediate data of the i-th layer as output data of the i-th layer, and terminates the processing (step S1005).
[0121] For example, the output data of the i-th layer is calculated using formula (15). Formula (15) is an example of the i-th layer output data that is the execution result of the input concealment influence removal process, and y_i is expressed by the matrix of formula (16).
[0122]
number
[0123]
number
[0124] Furthermore, for example, the output data of the i-th layer is calculated using equation (17). Equation (17) is an example of the output data of the i-th layer, and y_i is expressed by the matrix of equation (16).
[0125]
number
[0126] 11 is a dialog diagram showing an example of a machine learning parameter concealment setting information input screen. The CPU 51 of the parameter holder terminal 50 displays a machine learning parameter concealment setting information input screen 1100 on the display device 55 in accordance with a machine learning parameter concealment setting information input screen display program loaded into the memory 52.
[0127] The target parameter input section 1101 of the machine learning parameter anonymization setting information input screen 1100 is an input form in which the user inputs identification information (such as a file name) of the machine learning parameters to be transmitted to the confidential information processing server 30. The total processing layer number input section 1102 is an input form in which the user inputs the total number of target machine learning parameters, or in which the CPU 51 reads and inputs the number of layers from the machine learning parameters. The parameter anonymization necessity input section 1103 is an input form in which the user specifies whether or not the target machine learning parameters need to be anonymized for the confidential information processing server 30, or in which the CPU 51 reads and inputs the anonymization requirements from the target machine learning parameters.
[0128] The layer processing type input sections 1111-1114 of the machine learning parameter concealment setting information input screen 1100 are input forms in which the user inputs the type of operation of each processing layer, or the CPU 51 reads and inputs the type of operation of each layer from the target machine learning parameters. The trust region concealment processing availability input sections 1121-1124 are input forms in which the user inputs whether or not the operation of the layer can be processed in the TEE trust region 38, or the CPU 51 inputs based on the type of operation of each layer from the target machine learning parameters. The normal region concealment processing availability input sections 1131-1134 are input forms in which the user inputs whether or not the operation of the layer can be calculated while being concealed in the normal region 310, or the CPU 51 inputs based on the type of operation of each layer from the target machine learning parameters. The i-th layer processing execution area input sections 1141 to 1144 are input forms in which the user inputs whether the calculation for that layer is to be executed in the normal area 310 or in the TEE trust area 38, or in which the CPU 51 inputs the target machine learning parameters based on the type of calculation for each layer.
[0129] An input confirmation operation section 1105 of the machine learning parameter concealment setting information input screen 1100 is an input button that the user presses using the input device 56 such as a mouse when confirming the contents of the input to each form. An input cancellation operation section 1106 is an input button that the user presses using the input device 56 such as a mouse when canceling the contents of the input to each form and canceling the input.
[0130] Fig. 12 is a table showing an example of a data structure of each layer processing concealment setting information. In Fig. 12, as an example of each layer processing concealment setting information 524 of the TEE trust region 38, each layer processing concealment setting information 524 storing information on concealment setting of processing of each layer of the machine learning parameters transmitted by the parameter holder terminal 50 is shown.
[0131] Each layer processing concealment setting information 524 in FIG. 12 is configured with columns of layer number, processing type of layer, concealment processing in the trusted region or not, concealment processing in the normal region or not, and region in which the i-th layer processing is executed. Specifically, for example, the record 5243 in the third row indicates that the processing type of the third layer of the multiple processing layers is a process that executes a linear operation called dense. The record 5243 also indicates that concealment processing is possible in the trusted region to process confidential parameter data while protecting it. The record 5243 also indicates that concealment processing is possible in the normal region to process confidential parameter data while protecting it, and that the region in which the user-specified third layer processing is executed is the normal region. It should be noted that the layer processing concealment setting information 524 available in the confidential information processing system 1 according to this embodiment is not limited to the above-mentioned setting information data, and any data structure such as a database can be adopted.
[0132] (3) Effects of the embodiment When the confidential information processing server 30 of the above embodiment receives the machine learning parameters, which are confidential information, it performs parameter concealment processing of the machine learning parameters in the TEE trusted domain before placing the machine learning parameters in a normal domain where high-speed processing such as a GPU is possible. This makes it possible to prevent information leakage of the machine learning parameters, which are confidential information, from a server administrator who can access data in the normal domain of the confidential information processing server 30 or a cyber attacker who has usurped the authority of the administrator.
[0133] Furthermore, when the confidential information processing server 30 executes an inference process consisting of multiple layers of processing such as CNN on confidential input information in the normal domain, it executes anonymized input data generation processing in the TEE trusted domain and then outputs the anonymized input data to the normal domain. Then, in the normal domain capable of high-speed processing such as GPU, it executes processing of each layer such as CNN based on the anonymized input data and the anonymized machine learning parameters to generate anonymized intermediate data. Then, it inputs the generated anonymized intermediate data into the TEE trusted domain and then executes parameter anonymization influence removal processing and input anonymization influence removal processing to calculate the inference process result. Then, it encrypts the inference process result, which is confidential information, in the TEE trusted domain and transmits the processing result to the user terminal 20 or the data holder terminal 10. Therefore, the input data, machine learning parameters, and inference results, which are confidential information, can be always handled in anonymized or encrypted state when stored or processed in the normal domain capable of high-speed processing such as GPU. This makes it possible to prevent confidential information such as input data, machine learning parameters, and inference processing results from being leaked to attackers of the confidential information processing server 30 when inference processing consisting of multiple layers of processing is executed in a normal area where high-speed processing is possible, such as a GPU.
[0134] Furthermore, when the confidential information processing server 30 executes an inference process consisting of multiple layers of processing such as CNN on confidential input information in the normal domain, it generates a temporary random number to be used only for that input information in the TEE trusted domain. Then, it executes anonymized input data generation process including a different random number for each piece of input information, and then outputs the anonymized input data to the normal domain. Therefore, when a different temporary random number is not used for each piece of input information, the server administrator of the confidential information processing server 30 or a cyber attacker intentionally sets the input information to 0 to perform the anonymized input data generation process. This makes it possible to prevent information leakage caused by identifying the input data protection random number and exposing the input information.
[0135] Furthermore, the confidential information processing server 30 can specify whether to process each layer of the machine learning parameters in the normal domain or the TEE trusted domain, based on each layer processing concealment setting information specified by the user of the parameter holder terminal 50. This makes it possible to specify the domain for executing the processing for each layer, taking into consideration the feasibility of processing by different GPUs and the presence or absence of a speed-up effect for each feature such as the type of each layer of the machine learning parameters, and thus makes it possible to improve the efficiency of inference processing consisting of processing of multiple layers such as CNN.
[0136] (4) Other embodiments (4-1) Executor of the i-th layer processing In the above embodiment, the secret information processing server 30 executes the i-th layer processing in the normal domain 310, and performs the concealment and de-concealment of the i-th layer parameters and the i-th layer input data in the TEE trusted domain 38. The i-th layer processing does not necessarily need to be executed by the secret information processing server 30, and may be executed by an external device having a normal domain connected to the secret information processing server 30 via communication. That is, the secret information processing server 30 outputs an execution instruction for the i-th layer processing to the external device. Then, in response to the execution instruction, the external device executes the i-th layer processing in the normal domain and returns the execution result to the secret information processing server 30. Then, the secret information processing server 30 may acquire the execution result of the i-th layer processing from the external device.
[0137] (4-2) Temporary random numbers for protecting input data of the i-th layer In the above embodiment, the input data of the i-th layer is concealed by using the i-th layer temporary input data protection random number α_i·r_i (equation (12)) obtained by multiplying the i-th layer input data protection random number (equation (5)) by the temporary random number α_i (equation (10)) that protects the input data of the i-th layer. However, it is not always necessary to use the temporary random number α_i of the i-th layer, and "1" may be used instead of the temporary random number α_i of the i-th layer.
[0138] (4-3) i-th layer input data protection random number In the above-described embodiment, the input data of the i-th layer is concealed by using the i-th layer temporary input data protection random number α_i·r_i (equation (12)) obtained by multiplying the i-th layer input data protection random number (equation (5)) by the temporary random number α_i (equation (10)) that protects the input data of the i-th layer. However, it is not always necessary to use the i-th layer input data protection random number, and instead of the i-th layer input data protection random number, a matrix having the same size as the i-th layer input data protection random number and all elements of which are "1" may be used.
[0139] The present invention is not limited to the above-described embodiment, and includes various modifications and equivalent configurations within the spirit of the appended claims. For example, the above-described embodiment has been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to having all of the configurations described. In addition, a part of the configuration of a certain embodiment may be added to, deleted from, or replaced with another configuration.
[0140] In addition, the above-mentioned configurations, functions, processing units, processing means, etc. may be realized in part or in whole by hardware, for example by designing them as integrated circuits, etc. In addition, the above-mentioned configurations, functions, processing units, processing means, etc. may be realized in software, by a processor interpreting and executing a program that realizes each function.
[0141] Information such as programs, tables, and files that realize each function can be stored in a storage device such as a memory, a hard disk, or an SSD (Solid State Drive), or in a recording medium such as an IC card, an SD card, or a DVD.
[0142] In addition, the control lines and information lines in the drawings are those that are considered necessary for the explanation, and do not necessarily show all the control lines and information lines that are necessary for implementation. In reality, it can be considered that almost all components are connected to each other.
[0143] In addition, the present invention can achieve the same effect when providing confidential information, such as trade secrets and other highly confidential information whose disclosure to outside the company is restricted by internal company rules, in addition to personal information, to external organizations, etc., by concealing the information.
[0144] Furthermore, the present invention can improve the processing efficiency of advanced deep learning and other processes while keeping personal information and trade secrets confidential in multiple organizations that handle personal information such as customer information and medical information, and confidential information such as trade secrets, and can therefore be widely applied to information processing systems for public and private services that link a variety of confidential information. [Explanation of symbols]
[0145] 1: confidential information processing system, 30: confidential information processing server, 31: CPU, 38: TEE trusted region, 310: normal region, 312: i-th layer processing unit.
Claims
1. An information processing method performed by an information processing system that executes information processing using a neural network consisting of multiple i-th layers (i is a natural number between 1 and N), comprising the steps of: The information processing system includes a processor and a memory. The memory includes an isolated area that is a Trusted Execution Environment (TEE) isolated from a normal area, The isolation region includes: an i-th layer parameter used in processing the i-th layer; an i-th layer parameter protection random number that protects the i-th layer parameter; i-th layer input data input to the i-th layer; a random number for protecting the i-th layer input data; and The processor, in the isolated area, performing a parameter concealment process for each of the i-th layers, the parameter concealment process generating an i-th layer concealed parameter by multiplying the i-th layer parameter by the i-th layer parameter protection random number, and outputting the i-th layer concealed parameter to the normal area; performing, for each of the i-th layers, an input concealment influence removal parameter generation process for generating an i-th layer input data concealment removal parameter that removes the concealment applied to the i-th layer input data by multiplying the i-th layer parameter by the i-th layer input data protection random number, and storing the generated input concealment influence removal parameter in the isolated area; executing a concealed input data generation process for generating i-layer concealed input data by adding the i-layer input data protection random number to the i-layer input data and concealing the i-layer input data, and outputting the i-layer concealed input data to the normal area as the i-layer input data; performing a parameter concealment influence removal process in which an execution result of the i-th layer processing executed on the i-th layer concealed input data based on the i-th layer concealment parameter in the normal domain is obtained from the normal domain, and the execution result is divided by the i-th layer parameter protecting random number; executing an input concealment influence removal process in which the i-th layer input data concealment removal parameter is subtracted from the execution result of the parameter concealment influence removal process, and the execution result of the input concealment influence removal process is set as i-th layer output data output from the i-th layer. An information processing method comprising each of the processes.
2. 2. The information processing method according to claim 1, the memory includes the normal area, The processor, in the normal area, performing a process of the i-th layer on the i-th layer concealed input data based on the i-th layer concealment parameters, and outputting a result of the process of the i-th layer to the isolated area; 23. An information processing method comprising the steps of:
3. 2. The information processing method according to claim 1, The processor, in the isolated area, generating a temporary random number for each process of the i-layer for each piece of the i-layer input data, multiplying the i-layer input data protection random number by the temporary random number, and adding the result to the i-layer input data to generate the i-layer concealed input data, and outputting the i-layer concealed input data to the normal area as the i-layer input data; executing the input concealment influence removal process, in which a result of multiplying the i-layer input data concealment removal parameter and the temporary random number is subtracted from a result of execution of the parameter concealment influence removal process, and the result of execution of the input concealment influence removal process is set as i-layer output data output from the i-layer.
23. An information processing method comprising the steps of:
4. 4. The information processing method according to claim 3, The processor, Instead of the i-th layer input data protection random number, a matrix with all elements of 1 having the same size as the i-th layer input data protection random number is used.
23. An information processing method comprising:
5. 2. The information processing method according to claim 1, The isolation region includes: Hold each layer processing concealment setting information, which is information indicating whether the processing of the i-th layer is to be executed in the normal area or the isolated area; The processor, The processing of the i-th layer is executed in either the normal area or the isolated area according to the layer processing concealment setting information.
23. An information processing method comprising:
6. 2. The information processing method according to claim 1, The neural network is a convolutional neural network (CNN), The i-th layer parameters are machine learning parameters of the CNN.
23. An information processing method comprising:
7. An information processing system that executes information processing using a neural network consisting of multiple i-th layers (i is a natural number between 1 and N), A processor and a memory, The memory includes an isolated area that is a Trusted Execution Environment (TEE) isolated from a normal area, The isolation region includes: an i-th layer parameter used in processing the i-th layer; an i-th layer parameter protection random number that protects the i-th layer parameter; i-th layer input data input to the i-th layer; a random number for protecting the i-th layer input data; and The processor, in the isolated area, performing a parameter concealment process for each of the i-th layers, the parameter concealment process generating an i-th layer concealed parameter by multiplying the i-th layer parameter by the i-th layer parameter protection random number, and outputting the i-th layer concealed parameter to the normal area; performing, for each of the i-th layers, an input concealment influence removal parameter generation process for generating an i-th layer input data concealment removal parameter that removes the concealment applied to the i-th layer input data by multiplying the i-th layer parameter by the i-th layer input data protection random number, and storing the generated input concealment influence removal parameter in the isolated area; executing a concealed input data generation process for generating i-layer concealed input data by adding the i-layer input data protection random number to the i-layer input data and concealing the i-layer input data, and outputting the i-layer concealed input data to the normal area as the i-layer input data; performing a parameter concealment influence removal process in which an execution result of the i-th layer processing executed on the i-th layer concealed input data based on the i-th layer concealment parameter in the normal domain is obtained from the normal domain, and the execution result is divided by the i-th layer parameter protecting random number; executing an input concealment influence removal process in which the i-th layer input data concealment removal parameter is subtracted from the execution result of the parameter concealment influence removal process, and the execution result of the input concealment influence removal process is set as i-th layer output data output from the i-th layer. An information processing system comprising:
8. The information processing system according to claim 7, the memory includes the normal area, The processor, in the normal area, performing a process of the i-th layer on the i-th layer concealed input data based on the i-th layer concealment parameters, and outputting a result of the process of the i-th layer to the isolated area; An information processing system comprising:
9. The information processing system according to claim 7, The processor, in the isolated area, generating a temporary random number for each process of the i-layer for each piece of the i-layer input data, multiplying the i-layer input data protection random number by the temporary random number, and adding the result to the i-layer input data to generate the i-layer concealed input data, and outputting the i-layer concealed input data to the normal area as the i-layer input data; executing the input concealment influence removal process, in which a result of multiplying the i-layer input data concealment removal parameter and the temporary random number is subtracted from a result of execution of the parameter concealment influence removal process, and the result of execution of the input concealment influence removal process is set as i-layer output data output from the i-layer. An information processing system comprising:
10. 10. The information processing system according to claim 9, The processor, Instead of the i-th layer input data protection random number, a matrix with all elements of 1 having the same size as the i-th layer input data protection random number is used. An information processing system comprising:
11. The information processing system according to claim 7, The isolation region includes: Hold each layer processing concealment setting information, which is information indicating whether the processing of the i-th layer is to be executed in the normal area or the isolated area; The processor, The processing of the i-th layer is executed in either the normal area or the isolated area according to the layer processing concealment setting information. An information processing system comprising:
12. The information processing system according to claim 7, The neural network is a convolutional neural network (CNN), The i-th layer parameters are machine learning parameters of the CNN. An information processing system comprising:
Citation Information
Patent Citations
System for Measuring Information Leakage of Deep Learning Models
US20200082259A1
