Privacy-preserving online data inference methods and systems
By using encrypted data and data authorization to perform advertising data inference in a trusted execution environment, the problem of user privacy information exposure is solved, and data security and real-time performance are achieved. It is applicable to fields such as advertising, finance, healthcare and retail.
Patent Information
- Application Number
- CN202611155981.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-11-17
AI Technical Summary
During the advertising process, user privacy information is easily exposed due to data integration and analysis, and existing technologies are insufficient to effectively protect user privacy.
Data inference is performed using encrypted data and data authorization methods, and model training and result verification are carried out using a Trusted Execution Environment (TEE). This ensures that data is transmitted and processed in an encrypted state, including using Intel SGX technology to form a secure zone to protect user privacy.
It effectively reduces the risk of user privacy information being exposed, improves data security, meets the real-time requirements of advertising, enhances industry cooperation, complies with data privacy policies and regulations, and is applicable to fields such as advertising, finance, healthcare, and retail.
Smart Images

Figure CN123122875A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of software engineering technology, and specifically to a privacy-preserving online data reasoning method and system. Background Technology
[0002] To improve the accuracy and effectiveness of advertising, telecommunications operators and internet companies rely on the user information they possess to integrate and analyze it for precise ad targeting. However, this process of integrating and analyzing user information often involves directly processing user data, which includes sensitive information about users' privacy. This makes the user data highly sensitive, and the analysis process can easily lead to the exposure of users' private information. Summary of the Invention
[0003] To address the shortcomings of existing technologies, this invention provides a privacy-preserving online data inference method that can effectively reduce the risk of user privacy information being exposed.
[0004] This application provides a privacy-preserving online data reasoning method, the reasoning method comprising:
[0005] The computing party obtains encrypted data from at least one data party and a data authorization based on the encrypted data according to the data inference request;
[0006] The computing party receives the encrypted model provided by the model provider and decrypts the encrypted model to obtain the training model.
[0007] The computing party verifies the data authorization in a trusted execution environment. If the verification is successful, it outputs the inference result based on the training model and the encrypted data.
[0008] The inference result is encrypted, and the encrypted inference result is fed back to the data inference requester.
[0009] In one aspect, the data party is a first data party and a second data party;
[0010] The steps of obtaining encrypted data provided by at least one data provider and data authorization formed based on the encrypted data include:
[0011] Obtain the first encrypted data provided by the first data provider, and the first data authorization formed based on the first encrypted data;
[0012] Obtain the second encrypted data provided by the second data provider, and the second data authorization formed based on the second encrypted data;
[0013] The computational party verifies the data authorization in a trusted execution environment, and if the verification is successful, outputs the inference result based on the trained model and the encrypted data, including the following steps:
[0014] If both the first data authorization and the second data authorization are verified, the first encrypted data and the second encrypted data are merged together to form merged data;
[0015] The training model is trained based on the fused data, and the inference results are output.
[0016] In one aspect, the step of merging the first encrypted data and the second encrypted data includes:
[0017] The contents of the first encrypted data and the second encrypted data are loaded together in sequence, wherein the sequence can be either forward or reverse.
[0018] In one aspect, the step of verifying the data authorization in a trusted execution environment of the computing party includes:
[0019] In the trusted execution environment of the computing party, the signature information of the data authorization is compared with the pre-generated expected information;
[0020] The authorization verification is successful when the signature information matches the expected information.
[0021] If the signature information differs from the expected information, the authorization is deemed invalid, and the reasoning process terminates.
[0022] In one aspect, the step of decrypting the encrypted model to obtain the training model includes:
[0023] Verify the model authorization of the encryption model;
[0024] When the model authorization matches the expected information, the model authorization verification passes.
[0025] If the authorized model differs from the expected information, the authorization is deemed invalid, and the model decryption process terminates.
[0026] In one aspect, the requesting party is one of the data party, the model party, and a third party, wherein the third party is independent of the computing party, the data party, and the model party;
[0027] And / or, the model is a data source.
[0028] Furthermore, to address the aforementioned issues, this application also provides a privacy-preserving online data inference method, the inference method comprising:
[0029] The data provider encrypts local test data based on the data inference request to form encrypted data, and then forms data authorization based on the encrypted data.
[0030] The data provider transmits the encrypted data and the data authorization to the computing provider. The computing provider receives the encrypted model provided by the model provider and decrypts the encrypted model to obtain the training model. The computing provider verifies the data authorization in a trusted execution environment. If the verification is successful, the computing provider infers and outputs the inference result based on the training model and the encrypted data. The computing provider encrypts the inference result and feeds the encrypted inference result back to the data inference requester.
[0031] Furthermore, to address the aforementioned issues, this application also provides a privacy-preserving online data inference method, the inference method comprising:
[0032] The model provider encrypts the training model based on the data inference request, generating an encrypted model;
[0033] The model provider transmits the encrypted model to the computation provider. The computation provider receives the encrypted model provided by the model provider and decrypts it to obtain a training model. The computation provider obtains at least one encrypted data provided by a data provider and a data authorization formed based on the encrypted data according to the data inference request. The computation provider verifies the data authorization in a trusted execution environment. If the verification is successful, the computation provider outputs an inference result based on the training model and the encrypted data, encrypts the inference result, and feeds back the encrypted inference result to the data inference requester.
[0034] Furthermore, to address the aforementioned issues, this application also provides a privacy-preserving online data inference method, the inference method comprising:
[0035] The requester inputs a data inference instruction, and initiates a data inference request based on the data inference instruction;
[0036] The requesting party sends the data inference request to the data party, the model party, and the computing party respectively. The computing party obtains encrypted data provided by at least one data party and a data authorization formed based on the encrypted data according to the data inference request. The computing party receives an encrypted model provided by the model party and decrypts the encrypted model to obtain a training model. The computing party verifies the data authorization in a trusted execution environment. If the verification is successful, the computing party outputs an inference result based on the training model and the encrypted data, encrypts the inference result, and feeds back the encrypted inference result to the requesting party of the data inference.
[0037] In addition, to address the aforementioned issues, this application also provides a privacy-preserving online data inference system, the inference system comprising a memory and a processor, the memory storing a computer program, and the processor executing the inference method described above when invoking the computer program in the memory.
[0038] The beneficial effects of this invention are reflected in the following: the data obtained by the computing party from the data provider is encrypted, i.e., encrypted data, and authorization for the encrypted data is also obtained. The training model obtained by the computing party is also encrypted. During data inference, the encrypted model is decrypted, and the data authorization is verified. If the verification is successful, the encrypted data is trained on the training model to obtain the inference result. Therefore, it can be seen that the technical solution of this application is encrypted throughout the transmission process and cannot be used without decryption or verification, thus improving data security. Moreover, when training the training model with encrypted data, it is done within the computing party's trusted execution environment. The trusted execution environment offers a higher level of security and further reduces the risk of data being publicly exposed. Therefore, the technical solution of this application can effectively reduce the risk of user privacy information being exposed. Attached Figure Description
[0039] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.
[0040] Figure 1 This is a schematic diagram illustrating the process steps of applying the reasoning method in the calculation method of this application;
[0041] Figure 2 This is a schematic diagram of the process steps for forming the first data authorization and the second data authorization in the reasoning method of this application;
[0042] Figure 3 This is a schematic diagram illustrating the process steps in the reasoning method of this application to derive the reasoning result based on the fusion of data and the training model;
[0043] Figure 4 This is a schematic diagram of the data fusion process in the reasoning method of this application;
[0044] Figure 5 This is a schematic diagram of the data authorization process in the reasoning method of this application;
[0045] Figure 6 This is a schematic diagram of the steps involved in the verification model authorization process in the reasoning method of this application;
[0046] Figure 7 This is a schematic diagram illustrating the process steps of applying the reasoning method to the data side in this application;
[0047] Figure 8 This is a schematic diagram illustrating the process steps of applying the reasoning method in the model in this application;
[0048] Figure 9 This is a schematic diagram illustrating the process steps of applying the reasoning method in this application to the requesting party.
[0049] Figure 10 This is a schematic diagram illustrating the data transmission principle and process steps of the reasoning method in this application. Detailed Implementation
[0050] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.
[0051] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0052] like Figure 1 and Figure 10 As shown, this application provides a privacy-preserving online data inference method. This inference method can be applied to advertising placement. Based on a trained model, by training on user information data, it can infer precise targeting of users. The inference method in this embodiment is mainly applied in the computation unit 100, and the inference method includes:
[0053] In step S10, the computing party 100 obtains encrypted data from at least one data party and a data authorization based on the encrypted data according to the data inference request; the computing party 100 mainly completes model training on the encrypted data. Through the settings of the computing party 100, encrypted data from multiple data parties can be integrated together. To improve data confidentiality during this process, data is transmitted separately in encrypted form. Furthermore, the protected data authorization transmitted along with the encrypted data allows the computing party 100 to use the encrypted data after the data authorization is approved.
[0054] In step S20, the computing party 100 receives the encrypted model provided by the model provider and decrypts it to obtain the training model; the encrypted model can be understood as a large model. The training model is also transmitted in an encrypted manner, i.e., it is transmitted as an encrypted model.
[0055] In step S30, the computational method 100 verifies the data authorization within a Trusted Execution Environment (TEE). If the verification passes, it outputs the inference result based on the trained model and encrypted data. Regarding TEE technology, Intel proposed Software Guard Extensions (SGX), a practical user privacy protection measure that uses a hardware-assisted enclave (also known as a secure zone) to encapsulate sensitive data and critical program code, preventing their leakage. The technical solution of this application can employ SGX technology to form a Trusted Execution Environment; that is, the inference method of this application is based on SGX.
[0056] It should be noted that the storage area of the computing unit 100 is divided into a trusted execution environment and an untrusted execution environment. The transmission of encrypted data and encryption models generally passes through the untrusted execution environment first and then enters the trusted execution environment.
[0057] Step S40: Encrypt the inference result and send the encrypted inference result back to the requester of the data inference. After the calculation of the inference result is completed, the inference result is also encrypted. This enhances the protection of the inference result during the feedback process and reduces the risk of the inference result being exposed.
[0058] In this embodiment, the data obtained by the computing party 100 from the data provider is encrypted, i.e., encrypted data, and the computing party 100 also obtains authorization for the encrypted data. The training model obtained by the computing party 100 is also encrypted. During data inference, the encrypted model is decrypted, and the data authorization is verified. If the verification is successful, the encrypted data is trained on the training model to obtain the inference result. Therefore, it can be seen that the technical solution of this application is encrypted throughout the transmission process and cannot be used without decryption or verification, thus improving data security. Moreover, when training the training model with encrypted data, it is done within the trusted execution environment of the computing party 100. The trusted execution environment has a higher security level and better security, further reducing the risk of data being publicly exposed. Therefore, the technical solution of this application can effectively reduce the risk of user privacy information being exposed.
[0059] Furthermore, with increasingly stringent policies regarding data security and privacy, publicly sharing user information data for analysis and processing is no longer appropriate. This application reduces data sharing by sending this user information data to the computing party 100. Moreover, encryption is used to minimize the public disclosure of user information data.
[0060] like Figure 2As shown in one embodiment of this application, there can be one, two, or more data providers. The example given is two data providers: a first data provider 200 and a second data provider 300. The user data provided by the first data provider 200 and the second data provider 300 are different, and their data can complement each other. For example, the first data provider 200 could be a telecommunications operator, and the second data provider 300 could be an internet company.
[0061] The steps of obtaining encrypted data provided by at least one data provider and data authorization based on the encrypted data include:
[0062] Step S110: Obtain the first encrypted data provided by the first data provider 200, and the first data authorization formed based on the first encrypted data; when providing training data, the first data provider 200 encrypts the training data to form the first encrypted data, and forms the first data authorization for the first encrypted data.
[0063] Step S120: Obtain the second encrypted data provided by the second data provider 300, and the second data authorization formed based on the second encrypted data; when providing training data, the second data provider 300 also encrypts the training data to form the second encrypted data, and forms the second data authorization for the second encrypted data.
[0064] like Figure 3 As shown, the computational method 100 verifies data authorization in a trusted execution environment. If the verification passes, it outputs the inference result based on the trained model and encrypted data. The steps include:
[0065] Step S310: If both the first data authorization and the second data authorization are verified, the first encrypted data and the second encrypted data are merged together to form merged data; the first encrypted data and the second encrypted data are integrated together to form new data, namely merged data.
[0066] Step S320: Train the training model based on the fused data and infer the inference result. Through data integration, the training simulation is completed, thereby deriving the inference result.
[0067] As can be seen from this embodiment, encrypted data can come from one data provider or from multiple different data providers.
[0068] like Figure 4 As shown, in one embodiment of this application, the step of merging the first encrypted data and the second encrypted data together includes:
[0069] Step S311: Load the contents of the first encrypted data and the second encrypted data together in sequence, where "in sequence" can mean either forward or reverse order. That is, arrange the contents of the encrypted data according to a certain order. For example, if the encrypted data is user information, and each user information has its own identification information, it can be arranged according to the order of the identification information. More specifically, it could be ID card information, phone number information, or age, etc.
[0070] like Figure 5 As shown in one embodiment of this application, the step of verifying data authorization in the trusted execution environment of computing party 100 includes:
[0071] Step S300: In the trusted execution environment of the computing party 100, the signature information for data authorization is compared with the pre-generated expected information; while encrypting the data, a signature information for data authorization is formed. Before authorization verification, the expected information is generated in the computing party 100. Generally, the generation process of the signature information and the expected information is similar, and they can be generated using the same key or the same software tool, and the two can be mutually verified to see if they are the same.
[0072] Step S301: If the signature information matches the expected information, the authorization verification is successful. If the signature information matches the expected information, it can be understood that the expected information contains the information in the signature. For example, if the signature information contains Pkey and Hpred, and the expected information also contains Pkey and Hpred, then the authorization verification is successful, and the encrypted data can be used.
[0073] In step S302, if the signature information differs from the expected information, the authorization verification is invalid, and the inference process terminates. The difference between the signature information and the expected information can be understood as the expected information only including a portion of the signature information, or the expected information not including the signature information at all. For example, if the signature information includes Pkey and Hpred, but the expected information only includes Pkey or Hpred, the authorization verification will fail. Alternatively, if the expected information includes neither Pkey nor Hpred, the authorization verification will also fail.
[0074] like Figure 6 As shown, in one embodiment of this application, the step of decrypting the encrypted model to obtain the training model includes:
[0075] Step S210 verifies the model authorization of the encrypted model; when using the training model, it is also necessary to verify whether the use of the encrypted model is authorized. It can be understood that the model authorization information is generated when the training model is encrypted.
[0076] Step S211: If the model authorization is the same as the expected information, the model authorization verification passes; the model authorization information is the same as the data authorization information. The model authorization being the same as the expected information can be understood as the expected information containing information from the model authorization. For example, if the model authorization contains Pkey and Hpred information, and the expected information also contains Pkey and Hpred, then the authorization verification passes, and the encryption model can be used.
[0077] In step S212, if the model authorization differs from the expected information, the authorization verification is invalid, and model decryption terminates. A difference between the model authorization and the expected information can be understood as the expected information only including a portion of the model authorization, or the expected information not including the model authorization at all. For example, if the model authorization includes Pkey and Hpred, but the expected information only includes Pkey or Hpred, then the authorization verification is invalid. Alternatively, if the expected information includes neither Pkey nor Hpred, the authorization verification is also invalid.
[0078] In one embodiment of this application, the requesting party is one of the data party, the model party, and a third party, wherein the third party is independent of the computing party 100, the data party, and the model party; and / or, the model party is a data party.
[0079] As can be seen from this embodiment, the requesting party can be in various situations. It can be the data provider, the model provider, or an independent third party other than the data provider and the model provider that initiates the data inference request. Moreover, the party providing the training model can also be the party providing the encrypted data, meaning that the encrypted data and the encrypted model can come from the same party.
[0080] like Figure 10 As shown, based on the above content, the above data reasoning process is illustrated with an example:
[0081] The data providers are designated as First Data Provider 200 and Second Data Provider 300. First Data Provider 200 also serves as the model provider, meaning it provides the encryption model and the first encrypted data. First Data Provider 200 generates an asymmetric key pair with a public key of PA and a private key of SA. Second Data Provider 300 also generates an asymmetric key pair with a public key of PB and a private key of SB. Computation Provider 100 generates an asymmetric key pair with a public key of Pkey and a private key of Skey. The enclave hash of the training model is denoted as Htrain, and the enclave hash of the inference result is denoted as Hpred.
[0082] First, the first data provider 200 sorts the test data test_a.txt in ascending order by ID (Identity document) to obtain the file o_test_a.txt; the second data provider 300 sorts the test data test_b.txt in ascending order by ID to obtain the file o_test_b.txt.
[0083] In the second step, the first data provider 200 uses PA to encrypt the data o_test_a.txt to obtain the first encrypted data and generates the first data authorization, which is to sign a piece of information using SA. The signature information includes: Pkey and Hpred. The second data provider 300 uses PB to encrypt the data o_test_b.txt to obtain the second encrypted data and generates the second data authorization, which is to sign a piece of information using SB. The signature information includes: Pkey and Hpred.
[0084] Third, the first data provider 200 transmits the first encrypted data and the first data authorization to the untrusted execution environment of the computing provider 100; the second data provider 300 transmits the second encrypted data and the second data authorization to the untrusted execution environment of the computing provider 100.
[0085] The fourth step involves loading the first and second encrypted data, along with the first and second data authorizations, into the trusted environment of the computing party 100, and verifying the legality of the data authorizations, i.e., whether they are the expected information Pkey and Hpred. If the verification fails, it is considered an illegal authorization, which will immediately halt the entire process.
[0086] The fifth step involves decrypting the encrypted data from the first data provider (200) and the second data provider (300) after successful data authorization verification, and then merging the data based on their IDs. Specifically, the contents of the first and second encrypted data are arranged together in ascending order to form the merged data file, merged_test.txt.
[0087] In the sixth step, the first data provider 200 loads the generated encrypted model and model authorization into the trusted execution environment of the computing provider 100, and verifies the legality of the model authorization, i.e., whether it is the expected Pkey and Hpred. An illegal authorization will cause the entire process to stop immediately.
[0088] Step 7: After successful authorization verification, the model is decrypted to obtain the original model, which is then used for result prediction and inference with the fused data. The fused data is loaded, and the model is loaded to obtain the inference results.
[0089] Step 8: After the prediction inference is completed, the prediction result will be obtained. In a trusted execution environment, the result is encrypted using PA.
[0090] In the ninth step, the encryption result will be output to the untrusted execution environment of the computing party 100 and finally transmitted to the first data party 200.
[0091] Step 10: After receiving the encrypted result, the first data party 200 verifies the signature. If the signature information is Pkey and Hpred, the encrypted result is decrypted to obtain the inference result. The first data party 200, i.e., the requesting party, can then use the inference result to target advertisements to users.
[0092] As described above, the technical solution of this application can protect user privacy and improve data security. The entire process employs encryption technology, ensuring that the privacy and security of user data and models are always protected throughout the inference process. The entire inference process is conducted in an encrypted environment, greatly reducing the risk of data leakage.
[0093] Compared to traditional data processing methods, the inference method based on Intel SGX ensures real-time performance, meeting the stringent real-time requirements of ad delivery, typically taking less than ten seconds from the start of inference to ad delivery. This application not only improves ad delivery accuracy but also enables personalized recommendations based on users' actual needs, thereby increasing ad conversion rates.
[0094] The technical solution of this application can also enhance industry cooperation: due to data encryption, different telecommunications operators and Internet companies can share data more confidently, promoting cooperation between industries.
[0095] Furthermore, as regulations on data privacy and security become increasingly stringent in various countries, this invention can help companies comply with relevant regulations and reduce regulatory risks. Moreover, the technical solution of this application is not limited to advertising placement; it can be extended to other scenarios requiring data analysis and reasoning, such as finance, healthcare, and retail.
[0096] like Figure 7 As shown, this application also provides a privacy-preserving online data inference method. The inference method in this embodiment is mainly applied to the data side, and the inference method includes:
[0097] Step S11: The data provider encrypts the local test data according to the data reasoning request to form encrypted data, and then forms a data authorization based on the encrypted data.
[0098] In step S12, the data provider transmits encrypted data and data authorization to the computing provider 100. The computing provider 100 receives the encrypted model provided by the model provider and decrypts the encrypted model to obtain the training model. The data authorization is verified in a trusted execution environment. If the verification is successful, the computing provider outputs the inference result based on the training model and encrypted data, encrypts the inference result, and feeds the inference result back to the data inference requester.
[0099] like Figure 8 As shown, this application also provides a privacy-preserving online data inference method. The inference method in this embodiment is mainly applied to the model side, and the inference method includes:
[0100] Step S01: The model encrypts the training model according to the data inference request to generate an encrypted model;
[0101] In step S02, the model provider transmits the encrypted model to the computing provider 100. The computing provider 100 receives the encrypted model provided by the model provider and decrypts it to obtain the training model. The computing provider 100 obtains at least one encrypted data provided by the data provider and a data authorization formed based on the encrypted data according to the data inference request. The computing provider 100 verifies the data authorization in a trusted execution environment. If the verification is successful, the computing provider outputs the inference result based on the training model and the encrypted data, encrypts the inference result, and feeds the inference result back to the data inference requester.
[0102] like Figure 9 As shown, this application also provides a privacy-preserving online data inference method. The inference method in this embodiment is mainly applied to the requesting party, and the inference method includes:
[0103] Step S1: The requester inputs a data inference instruction and initiates a data inference request based on the instruction.
[0104] In step S2, the requesting party sends the data inference request to the data party, the model party, and the computing party 100 respectively. The computing party 100 obtains at least one encrypted data provided by the data party and a data authorization formed based on the encrypted data according to the data inference request. The computing party 100 receives the encrypted model provided by the model party and decrypts the encrypted model to obtain the training model. The computing party 100 verifies the data authorization in a trusted execution environment. If the verification is successful, the computing party outputs the inference result based on the training model and the encrypted data, encrypts the inference result, and feeds the inference result back to the requesting party of the data inference.
[0105] This application also provides a privacy-preserving online data inference system. The inference system includes a memory and a processor. The memory stores a computer program, and the processor executes the inference method described above when it calls the computer program in the memory.
[0106] Specific embodiments and beneficial effects of the reasoning system in this application are described in the above-described reasoning method and will not be repeated here.
[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.
Claims
1. A method for privacy-preserving data online inference, characterized in that, The reasoning method includes: The computing party obtains encrypted data from at least one data party and a data authorization based on the encrypted data according to the data inference request; The computing party receives the encrypted model provided by the model provider and decrypts the encrypted model to obtain the training model. The computing party verifies the data authorization in a trusted execution environment. If the verification is successful, it outputs the inference result based on the training model and the encrypted data. The inference result is encrypted, and the encrypted inference result is fed back to the data inference requester.
2. The inference method of claim 1, wherein, The data party is a first data party and a second data party; The steps of obtaining encrypted data provided by at least one data provider and data authorization formed based on the encrypted data include: Obtain the first encrypted data provided by the first data provider, and the first data authorization formed based on the first encrypted data; Obtain the second encrypted data provided by the second data provider, and the second data authorization formed based on the second encrypted data; The computational party verifies the data authorization in a trusted execution environment, and if the verification is successful, outputs the inference result based on the trained model and the encrypted data, including the following steps: If both the first data authorization and the second data authorization are verified, the first encrypted data and the second encrypted data are merged together to form merged data; The training model is trained based on the fused data, and the inference results are output.
3. The inference method of claim 2, wherein, The step of merging the first encrypted data and the second encrypted data together includes: The contents of the first encrypted data and the second encrypted data are loaded together in sequence, wherein the sequence can be either forward or reverse.
4. The inference method of claim 1, wherein, The step of verifying the data authorization in the trusted execution environment of the computing party includes: In the trusted execution environment of the computing party, the signature information of the data authorization is compared with the pre-generated expected information; The authorization verification is successful when the signature information matches the expected information. If the signature information differs from the expected information, the authorization is deemed invalid, and the reasoning process terminates.
5. The inference method of claim 4, wherein, The step of decrypting the encrypted model to obtain the training model includes: Verify the model authorization of the encryption model; When the model authorization matches the expected information, the model authorization verification passes. If the authorized model differs from the expected information, the authorization is deemed invalid, and the model decryption process terminates.
6. The reasoning method according to any one of claims 1 to 5, characterized in that, The requesting party is one of the data party, the model party, and a third party, wherein the third party is independent of the computing party, the data party, and the model party; And / or, the model is a data source.
7. A privacy-preserving data online inference method, characterized in that, The reasoning method includes: The data provider encrypts local test data based on the data inference request to form encrypted data, and then forms data authorization based on the encrypted data. The data provider transmits the encrypted data and the data authorization to the computing provider. The computing provider receives the encrypted model provided by the model provider and decrypts the encrypted model to obtain the training model. The computing provider verifies the data authorization in a trusted execution environment. If the verification is successful, the computing provider infers and outputs the inference result based on the training model and the encrypted data. The computing provider encrypts the inference result and feeds the encrypted inference result back to the data inference requester.
8. A privacy-preserving data online inference method, characterized in that, The reasoning method includes: The model provider encrypts the training model based on the data inference request, generating an encrypted model; The model provider transmits the encrypted model to the computation provider. The computation provider receives the encrypted model provided by the model provider and decrypts it to obtain a training model. The computation provider obtains at least one encrypted data provided by a data provider and a data authorization formed based on the encrypted data according to the data inference request. The computation provider verifies the data authorization in a trusted execution environment. If the verification is successful, the computation provider outputs an inference result based on the training model and the encrypted data, encrypts the inference result, and feeds back the encrypted inference result to the data inference requester. 9.A method for privacy-preserving data online inference, the method comprising: The reasoning method includes: The requester inputs a data inference instruction, and initiates a data inference request based on the data inference instruction; The requesting party sends the data inference request to the data party, the model party, and the computing party respectively. The computing party obtains encrypted data provided by at least one data party and a data authorization formed based on the encrypted data according to the data inference request. The computing party receives an encrypted model provided by the model party and decrypts the encrypted model to obtain a training model. The computing party verifies the data authorization in a trusted execution environment. If the verification is successful, the computing party outputs an inference result based on the training model and the encrypted data, encrypts the inference result, and feeds back the encrypted inference result to the requesting party of the data inference.
10. A privacy-preserving online data inference system, characterized in that, The reasoning system includes a memory and a processor, the memory storing a computer program, and the processor executing the reasoning method as described in any one of claims 1 to 9 when it invokes the computer program in the memory.
