Model reasoning method, server, NEF, client and computer program product

Abstract

The invention relates to a model reasoning method, a server, an NEF, a client and a computer program product, and relates to the technical field of wireless communication. The model reasoning method is executed by a server side and comprises the steps that related information of longitudinal federated learning VFL reasoning is determined for each client side in a plurality of client sides, and the related information of the VFL reasoning corresponding to each client side comprises at least one of a VFL association identifier corresponding to each client side, first sample indication information and first feature indication information; first request information is sent to each client side so that each client side can conduct reasoning, and the first request information comprises relevant information of VFL reasoning corresponding to each client side.

Description

Technical Field

[0001] This disclosure relates to the field of wireless communication technology, and in particular to a model inference method, server, NEF, client, computer-readable storage medium, and computer program product. Background Technology

[0002] With the development of federated learning technology, the concept of vertical federated learning has been introduced, which is different from the existing horizontal federated learning technology.

[0003] Vertical federated learning consists of two processes: training and inference. The inference process occurs after model training. Due to data isolation, vertical federated learning involves data samples with the same data but different sample features. That is, data samples with different features are distributed across different network elements. These original data cannot be exchanged during the inference process. Therefore, compared with the traditional inference process, the inference process of vertical federated learning requires the joint participation of all parties, i.e., distributed inference. Summary of the Invention

[0004] The inventors of this disclosure have discovered the following problem in the above-mentioned related technologies: how to improve the feasibility of distributed reasoning in the longitudinal federated learning reasoning process.

[0005] In view of this, this disclosure proposes a model inference method that can determine different relevant information for vertical federated learning inference for different clients, and send a first request message including the relevant information for vertical federated learning inference corresponding to each client to each client, so that each client can perform inference, thereby improving the feasibility of distributed inference in the vertical federated learning inference process.

[0006] According to some embodiments of the first aspect of this disclosure, a model inference method is provided, executed by a server, comprising: determining relevant information for longitudinal federated learning VFL inference for each of a plurality of clients, wherein the relevant information for VFL inference for each client includes at least one of a VFL association identifier, a first sample indication information, and a first feature indication information for each client; and sending a first request information to each client so that each client can perform inference, wherein the first request information includes relevant information for VFL inference for each client.

[0007] In some embodiments, the VFL association identifier is used to associate the VFL process and / or the VFL model; the first sample indication information is used to indicate the data sample; and / or the first feature indication information is used to indicate the data feature.

[0008] In some embodiments, the relevant information for VFL inference for each client further includes at least one of the following: an interoperability indicator, an analysis identifier, filter information, and a VFL indicator for each client. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of the following: single network slice selection assistance information, application identifier, region of interest, and data network access identifier.

[0009] In some embodiments, determining relevant information for longitudinal federated learning VFL inference for each of the multiple clients includes:

[0010] According to preset rules, at least one of the sample indication information, feature indication information, and filter information corresponding to multiple clients is divided to determine the relevant information for VFL inference for each client. The preset rules include at least one of the following: the internal configuration of the server, the internal logic of the server, the network element type of each client, the relevant information of each client obtained from other network elements during the client search process, and the information associated with each client during the training process.

[0011] In some embodiments, according to preset rules, at least one of the sample indication information, feature indication information, and filter information corresponding to multiple clients is divided to determine the relevant information for VFL inference for each client, including at least one of the following: According to preset rules, the set of user equipment identifiers in the sample indication information corresponding to multiple clients is divided to determine the set of user equipment identifiers corresponding to each client in the first sample indication information for each client; According to preset rules, the sample identifiers in the sample indication information corresponding to multiple clients are divided to determine the sample identifiers corresponding to each client in the first sample indication information for each client; According to preset rules, the feature identifiers in the feature indication information corresponding to multiple clients are divided to determine the feature identifiers corresponding to each client in the first feature indication information for each client; According to preset rules, the experience quality index information in the feature indication information corresponding to multiple clients is divided to determine the experience quality index information corresponding to each client in the first feature indication information for each client; According to preset rules, the region of interest in the filter information corresponding to multiple clients is divided to determine the region of interest corresponding to each client in the filter information for each client; According to preset rules, the experience quality index information in the filter information corresponding to multiple clients is divided to determine the experience quality index information corresponding to each client in the filter information for each client.

[0012] In some embodiments, the inference method further includes: performing local VFL inference based on data in the server to obtain intermediate results of local VFL inference.

[0013] In some embodiments, the inference method further includes: receiving first response information sent by each client, wherein the first response information corresponding to each client includes at least one of the following: intermediate result of VFL inference corresponding to each client, second sample indication information, second feature indication information, and VFL association identifier; and generating an inference result based on the first response information sent by each client and / or the intermediate result of local VFL inference.

[0014] In some embodiments, the inference method further includes: receiving first response information sent by the Network Open Function (NEF), wherein the first response information is obtained by aggregating at least one of the intermediate results of VFL inference for each client, second sample indication information, second feature indication information, and VFL association identifier received by the NEF; and generating an inference result based on the first response information and / or the intermediate results of local VFL inference.

[0015] In some embodiments, the inference method further includes: receiving a second request message sent by a consumer network function (NF), wherein the second request message is used to request analysis results, the second request message includes at least one of an analysis identifier and a VFL indicator, the analysis identifier is used to identify that the requested analysis and / or the analysis results are generated using a VFL method, and the VFL indicator is used to indicate the model trained by the VFL, wherein sending a first request message to each client includes: sending a first request message to each client according to the second request message.

[0016] In some embodiments, the second request information further includes filter information, wherein the filter information includes at least one of a single network slice selection assistance information, an application identifier, a region of interest, and a data network access identifier.

[0017] In some embodiments, the inference method further includes: sending a second response information to the consuming NF, wherein the second response information includes at least one of an analysis result and second validity time information, the analysis result being generated based on the inference result, and the second validity time information being used to indicate the validity time of the inference result.

[0018] In some embodiments, the first response information further includes: first validity time information, used to indicate the validity time of intermediate results of VFL inference.

[0019] In some embodiments, the first sample indication information is used to indicate a data sample; the second sample indication information is used to indicate a data sample; the first sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier; and / or the second sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier.

[0020] In some embodiments, the first feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the second feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the first feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier, and / or the second feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier.

[0021] In some embodiments, the inference method further includes: collecting data samples from the corresponding network element based on the first request information.

[0022] In some embodiments, the analysis result is in the form of an analysis ID output, including at least one of the data in the analysis ID output, or including at least one of the experience quality index information, latency information, and 5G service quality index.

[0023] In some embodiments, the consumer NF is the analysis logic function AnLF, and the server is the application function AF.

[0024] In some embodiments, the server is a Network Data Analysis Function (NWDAF), and each client is an Application Function (AF); or, the server and each client are different NWDAFs; or, the server is an AF and each client is an NWDAF.

[0025] In some embodiments, when the AF is an untrusted AF, sending the first request information to each client includes: sending the first request information to each client via the Network Open Function (NEF), wherein, when the server is an untrusted AF, the VFL association identifier in the first request information is an external VFL association identifier, the first sample indication information includes first external sample indication information, and / or the first feature indication information includes first external feature indication information, and the NEF maps the external VFL association identifier in the first request information to an internal VFL association identifier, maps the first external sample indication information to first internal sample indication information, and / or maps the first VFL association identifier ... VFL association identifier, and / or maps the first external sample indication information to first internal VFL association identifier, maps the first external sample indication information to first internal VFL association identifier, and / or maps the first external sample indication information to first internal VFL association identifier, An external feature indication is mapped to a first internal feature indication. In the case of multiple clients having untrusted AFs, for each client that is an untrusted AF, the VFL association identifier in the first request information is an internal VFL association identifier, the first sample indication information includes the first internal sample indication information, and / or the first feature indication information includes the first internal feature indication information. NEF maps the first internal VFL association identifier in the first request information to a first external VFL association identifier, maps the first internal sample indication information to the first external sample indication information, and / or maps the first internal feature indication information to the first external feature indication information.

[0026] In some embodiments, the inference method further includes: receiving first response information sent by each client via the Network Development Function (NEF), wherein the first response information corresponding to each client includes at least one of the following: intermediate result of VFL inference for each client, second sample indication information, second feature indication information, and VFL association identifier; and generating an inference result based on the first response information sent by each client and / or the intermediate result of local VFL inference.

[0027] In some embodiments, the inference method further includes: when multiple clients have untrusted AFs, for each client that is an untrusted AF, the intermediate result of VFL inference in the first response information is an external intermediate result, the VFL association identifier is an external VFL association identifier, the second sample indication information includes second external sample indication information, and / or the second feature indication information includes second external feature indication information; the NEF maps the external intermediate result in the first response information to an internal intermediate result, maps the external VFL association identifier to an internal VFL association identifier, maps the second external sample indication information to second internal sample indication information, and / or maps the second external feature indication information to second internal sample indication information. The indication information is mapped to the second internal feature indication information. In the case that the server is an untrusted AF, the intermediate result of VFL inference in the first response information is an internal intermediate result, the VFL association identifier is an internal VFL association identifier, the second sample indication information includes the second internal sample indication information, and / or the second feature indication information includes the second internal feature indication information. NEF maps the internal intermediate result in the first response information to the external intermediate result, maps the internal VFL association identifier to the external VFL association identifier, maps the second internal sample indication information to the second external sample indication information, and / or maps the second internal feature indication information to the second external feature indication information.

[0028] In some embodiments, generating inference results based on the first response information sent by each client and / or the intermediate results of local VFL inference includes: aggregating the first response information sent by each client and the intermediate results of local VFL inference to generate inference results.

[0029] According to some embodiments of the second aspect of this disclosure, a model inference method is provided, executed by NEF, comprising: receiving third request information sent by a server, wherein the third request information includes at least one of VFL association identifier, sample indication information corresponding to multiple clients, and feature indication information; determining relevant information for longitudinal federated learning VFL inference for each of the multiple clients according to the third request information, wherein the relevant information for VFL inference corresponding to each client includes at least one of VFL association identifier, third sample indication information, and third feature indication information corresponding to each client; and sending third sub-request information to each client so that each client can perform inference, wherein the third sub-request information includes relevant information for VFL inference corresponding to each client.

[0030] In some embodiments, VFL association identifiers are used to associate VFLs with processes and / or VFL models; sample indication information is used to indicate data samples; and / or feature indication information is used to indicate data features.

[0031] In some embodiments, the relevant information for VFL inference for each client further includes at least one of the following: an interoperability indicator, an analysis identifier, filter information, and a VFL indicator for each client. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of the following: single network slice selection assistance information, application identifier, region of interest, and data network access identifier.

[0032] In some embodiments, determining the relevant information for longitudinal federated learning (VFL) inference for each of the multiple clients based on the third request information includes: dividing at least one of the sample indication information, feature indication information, and filter information corresponding to the multiple clients according to preset rules, and determining the relevant information for VFL inference for each client. The preset rules include at least one of the following: the internal configuration of the server, the internal logic of the server, the network element type of each client, the relevant information of each client obtained from other network elements during the client search process, and the information associated with each client during the training process.

[0033] In some embodiments, according to preset rules, at least one of the sample indication information, feature indication information, and filter information corresponding to multiple clients is divided, and the relevant information for VFL inference for each client is determined, including at least one of the following: According to preset rules, the set of user equipment identifiers in the sample indication information corresponding to multiple clients is divided, and a set of user equipment identifiers corresponding to each client in the third sample indication information is determined for each client; According to preset rules, the sample identifiers in the sample indication information corresponding to multiple clients are divided, and a sample identifier corresponding to each client in the third sample indication information is determined for each client; According to preset rules, the feature identifiers in the feature indication information corresponding to multiple clients are divided, and a feature identifier corresponding to each client in the third feature indication information is determined for each client; According to preset rules, the experience quality index information in the feature indication information corresponding to multiple clients is divided, and experience quality index information corresponding to each client in the third feature indication information is determined for each client; According to preset rules, the region of interest in the filter information corresponding to multiple clients is divided, and a region of interest corresponding to each client in the filter information is determined for each client; According to preset rules, the experience quality index information in the filter information corresponding to multiple clients is divided, and experience quality index information corresponding to each client in the filter information is determined for each client.

[0034] In some embodiments, the inference method further includes: receiving third sub-response information sent by each client, wherein the third sub-response information includes at least one of the following: intermediate results of VFL inference for each client, fourth feature indication information, fourth sample indication information, and VFL association identifier; if NEF has aggregation capability, aggregating the third sub-response information to obtain third response information; and sending the third response information to the server so that the server can generate inference results based on the third response information and / or the intermediate results of local VFL inference.

[0035] In some embodiments, the third response information further includes: third validity time information, used to indicate the validity time of intermediate results of VFL inference.

[0036] In some embodiments, the third sample indication information is used to indicate a data sample; the fourth sample indication information is used to indicate a data sample; the third sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier; and / or the fourth sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier.

[0037] In some embodiments, the third feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the fourth feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the third feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier, and / or the fourth feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier.

[0038] In some embodiments, the server is a Network Data Analysis Function (NWDAF), and each client is an Application Function (AF); or, the server and each client are different NWDAFs; or, the server is an AF and each client is an NWDAF.

[0039] In some embodiments, the inference method further includes: when the server is an untrusted AF, the VFL association identifier in the third sub-request information is an external VFL association identifier, the third sample indication information includes third external sample indication information, and / or the third feature indication information includes third external feature indication information; mapping the external VFL association identifier in the third sub-request information to an internal VFL association identifier, mapping the third external sample indication information to third internal sample indication information, and / or mapping the third external feature indication information to third internal feature indication information; when multiple clients have untrusted AFs, for each client that is an untrusted AF, the VFL association identifier in the third sub-request information is an internal VFL association identifier, the third sample indication information includes third internal sample indication information, and / or the third feature indication information includes third internal feature indication information; mapping the internal VFL association identifier in the third sub-request information to an external VFL association identifier, mapping the third internal sample indication information to third external sample indication information, and / or mapping the third internal feature indication information to third external feature indication information.

[0040] In some embodiments, the inference method further includes: sending third sub-response information to the server when NEF does not have aggregation capabilities, so that the server can generate inference results based on the third sub-response information and / or intermediate results of local VFL inference.

[0041] In some embodiments, where multiple clients have untrusted AFs, for each client that is an untrusted AF, the intermediate result of VFL inference in the third sub-response information is an external intermediate result, the VFL association identifier is an external VFL association identifier, the fourth sample indication information includes fourth external sample indication information, and / or the fourth feature indication information includes fourth external feature indication information. The external intermediate result in the third sub-response information is mapped to an internal intermediate result, the external VFL association identifier is mapped to an internal VFL association identifier, the fourth external sample indication information is mapped to a fourth internal sample indication information, and / or the fourth external feature indication information is mapped to an internal VFL association identifier. The information is mapped to the fourth internal feature indication information. In the case that the server is an untrusted AF, the intermediate result of VFL inference in the third sub-response information is an internal intermediate result, the VFL association identifier is an internal VFL association identifier, the fourth sample indication information includes the fourth internal sample indication information, and / or the fourth feature indication information includes the fourth internal feature indication information. The internal intermediate result in the third sub-response information is mapped to the external intermediate result, the internal VFL association identifier is mapped to the external VFL association identifier, the fourth internal sample indication information is mapped to the fourth external sample indication information, and / or the fourth internal feature indication information is mapped to the fourth external feature indication information.

[0042] According to some embodiments of the third aspect of this disclosure, a model inference method is provided, executed by a client, comprising: receiving first request information sent by a server, wherein the first request information includes relevant information for VFL inference corresponding to the client, the relevant information for VFL inference corresponding to the client includes at least one of VFL association identifier, first sample indication information and first feature indication information corresponding to the client, and the relevant information for VFL inference corresponding to the client is divided by the server for the client from the relevant information for VFL inference corresponding to multiple clients; and performing VFL inference according to the first request information.

[0043] In some embodiments, the VFL association identifier is used to associate the VFL process and / or the VFL model; the first sample indication information is used to indicate the data sample; and / or the first feature indication information is used to indicate the data feature.

[0044] In some embodiments, the inference method further includes: sending a first response information to the server, wherein the first response information includes at least one of the following: intermediate result of VFL inference corresponding to the client, second sample indication information, second feature indication information, and VFL association identifier.

[0045] In some embodiments, the VFL inference information corresponding to the client further includes at least one of the following: interoperability indicator, analysis identifier, filter information and VFL indicator corresponding to the client; the analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL; the VFL indicator is used to indicate the model trained by VFL; and the filter information includes at least one of single network slice selection assistance information, application identifier, region of interest and data network access identifier.

[0046] In some embodiments, the first response information further includes: first validity time information, used to indicate the validity time of intermediate results of VFL inference.

[0047] In some embodiments, the first sample indication information is used to indicate a data sample; the second sample indication information is used to indicate a data sample; the first sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier; and / or the second sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier.

[0048] In some embodiments, the first feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the second feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the first feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier, and / or the second feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier.

[0049] In some embodiments, the inference method further includes: collecting data samples from the corresponding network element based on the first request information.

[0050] In some embodiments, the server is a Network Data Analysis Function (NWDAF) and the client is an Application Function (AF); or, the server and the client are different NWDAFs; or, the server is an AF and the client is an NWDAF.

[0051] According to some embodiments of the fourth aspect of this disclosure, a model inference method is provided, executed by a client, comprising: receiving third sub-request information sent by the Network Open Function (NEF), wherein the third sub-request information includes relevant information on VFL inference corresponding to the client, the relevant information on VFL inference corresponding to the client includes at least one of VFL association identifier, third sample indication information and third feature indication information corresponding to the client, and the relevant information on VFL inference corresponding to the client is divided for the client from multiple relevant information on VFL inference corresponding to the client obtained by NEF from the server; and performing inference based on the third sub-request information.

[0052] In some embodiments, VFL association identifiers are used to associate VFLs with processes and / or VFL models; third sample indication information is used to indicate data samples; and / or third feature indication information is used to indicate data features.

[0053] In some embodiments, the inference method further includes: sending a third sub-response information to the NEF, wherein the third sub-response information includes at least one of the client's VFL inference intermediate results, fourth feature indication information, fourth sample indication information, and VFL association identifier, so that the NEF generates response information based on the third sub-response information and sends it to the server.

[0054] In some embodiments, the VFL inference information corresponding to the client further includes at least one of the following: interoperability indicator, analysis identifier, filter information and VFL indicator corresponding to the client; the analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL; the VFL indicator is used to indicate the model trained by VFL; and the filter information includes at least one of single network slice selection assistance information, application identifier, region of interest and data network access identifier.

[0055] In some embodiments, the third sample indication information is used to indicate a data sample; the fourth sample indication information is used to indicate a data sample; the third sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier; and / or the fourth sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier.

[0056] In some embodiments, the third feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the fourth feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the third feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier, and / or the fourth feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier.

[0057] In some embodiments, the server is a Network Data Analysis Function (NWDAF) and the client is an Application Function (AF); or, the server and the client are different NWDAFs; or, the server is an AF and the client is an NWDAF.

[0058] According to some embodiments of the fifth aspect of this disclosure, a server is provided, comprising: a first determining unit configured to determine relevant information for longitudinal federated learning VFL inference for each of a plurality of clients, wherein the relevant information for VFL inference for each client includes at least one of a VFL association identifier, a first sample indication information, and a first feature indication information for each client; and a first sending unit configured to send first request information to each client so that each client can perform inference, wherein the first request information includes relevant information for VFL inference for each client.

[0059] According to some embodiments of the sixth aspect of this disclosure, a NEF is provided, comprising: a first receiving unit configured to receive third request information sent by a server, wherein the third request information includes at least one of VFL association identifier, sample indication information corresponding to multiple clients, and feature indication information; a second determining unit configured to determine relevant information for longitudinal federated learning VFL inference for each of the multiple clients based on the third request information, wherein the relevant information for VFL inference corresponding to each client includes at least one of VFL association identifier, third sample indication information, and third feature indication information corresponding to each client; and a second sending unit configured to send third sub-request information to each client so that each client can perform inference, wherein the third sub-request information includes relevant information for VFL inference corresponding to each client.

[0060] According to some embodiments of the seventh aspect of this disclosure, a client is provided, comprising: a second receiving unit configured to receive first request information sent by a server, wherein the first request information includes relevant information on VFL inference corresponding to the client, the relevant information on VFL inference corresponding to the client includes at least one of VFL association identifier, first sample indication information and first feature indication information corresponding to the client, and the relevant information on VFL inference corresponding to the client is divided by the server for the client from the relevant information on VFL inference corresponding to multiple clients; and a first inference unit configured to perform VFL inference according to the first request information.

[0061] According to some embodiments of the eighth aspect of this disclosure, a client is provided, comprising: a third receiving unit configured to receive third sub-request information sent by the Network Open Function (NEF), wherein the third sub-request information includes relevant information on VFL inference corresponding to the client, the relevant information on VFL inference corresponding to the client includes at least one of VFL association identifier, third sample indication information and third feature indication information corresponding to the client, and the relevant information on VFL inference corresponding to the client is divided for the client from multiple relevant information on VFL inference corresponding to clients obtained by the NEF from the server; and a second inference unit configured to perform inference based on the third sub-request information.

[0062] According to some embodiments of the ninth aspect of this disclosure, an electronic device is provided, including: a memory and a processor coupled to the memory, the processor being configured to execute an inference method of the model in any of the above embodiments based on instructions stored in the memory.

[0063] According to some embodiments of the tenth aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the reasoning method of the model in any of the above embodiments.

[0064] According to some embodiments of the eleventh aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the reasoning method of the model in any of the above embodiments.

[0065] In the above embodiments, by determining relevant information for vertical federated learning inference corresponding to each of the multiple clients, wherein the relevant information for each client's vertical federated learning inference includes at least one of vertical federated learning association identifier, sample indication information, and feature indication information corresponding to each client. The vertical federated learning association identifier identifies the vertical federated learning process of each client, the sample indication information indicates data samples in the server, and the feature indication information indicates data features in the server and / or data features in each client. Then, a first request message including the relevant information for vertical federated learning inference corresponding to that client is sent to each client to request each client to perform vertical federated learning inference, thereby improving the feasibility of distributed inference in the vertical federated learning process. Attached Figure Description

[0066] The accompanying drawings, which form part of this specification, illustrate embodiments of this disclosure and, together with the specification, serve to explain the principles of this disclosure.

[0067] This disclosure can be more clearly understood with reference to the accompanying drawings and the following detailed description.

[0068] Figure 1 Flowcharts illustrating some embodiments of the reasoning methods of the models disclosed herein.

[0069] Figure 2 Flowcharts illustrating some other embodiments of the reasoning method of the model of this disclosure are shown.

[0070] Figure 3 Flowcharts illustrating further embodiments of the reasoning methods of the models disclosed herein.

[0071] Figure 4 Flowcharts illustrating further embodiments of the reasoning methods of the models disclosed herein.

[0072] Figure 5 Schematic diagrams illustrating some embodiments of the server side of this disclosure are shown.

[0073] Figure 6 Schematic diagrams illustrating some embodiments of the NEF disclosed herein are shown.

[0074] Figure 7 Schematic diagrams illustrating some embodiments of the client of this disclosure are shown.

[0075] Figure 8Schematic diagrams illustrating other embodiments of the client of this disclosure are shown.

[0076] Figure 9 Schematic diagrams illustrating some embodiments of the electronic devices disclosed herein.

[0077] Figure 10 Schematic diagrams illustrating some embodiments of the reasoning methods of the models disclosed herein.

[0078] Figure 11 Schematic diagrams illustrating some other embodiments of the reasoning method of the model of this disclosure.

[0079] Figure 12 Schematic diagrams illustrating further embodiments of the reasoning methods of the models disclosed herein. Detailed Implementation

[0080] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present disclosure.

[0081] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0082] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this disclosure or its application or use.

[0083] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the specification.

[0084] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0085] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0086] Vertical Federated Learning (VFL) uses the same data samples but different sample characteristics. In 5GC (5 Generation Core), for example, for user experience quality metrics, the data in AF (Application Function) and NWDAF (Network Data Analytics Function) have different data characteristics but the same data samples.

[0087] Therefore, in the inference process of vertical federated learning, all participating parties need to collaborate on the reasoning, i.e., to achieve distributed inference. Currently, the participants in vertical federated learning include two roles: server and client. Furthermore, a more complex scenario exists where the participants in vertical federated learning involve a server and multiple clients. In this case, the distributed inference process requires the joint participation of both the server and the multiple clients.

[0088] Therefore, improving the feasibility of distributed reasoning in the vertical federated learning process is a problem that needs to be solved.

[0089] To address the aforementioned issues, this disclosure proposes a model inference method. By determining relevant information for vertical federated learning inference for each of the multiple clients, generating first request information for each client, and sending the first request information to each client, a vertical federated learning inference request is initiated to multiple clients. This interaction with multiple clients facilitates the joint participation of multiple clients and the server in the vertical federated learning inference process, improving the feasibility of distributed inference in the vertical federated learning process. The specific details are as follows.

[0090] Figure 1 Flowcharts illustrating some embodiments of the reasoning methods of the models disclosed herein.

[0091] like Figure 1 As shown, the inference method of the model includes steps 110 to 120, which are executed by the server.

[0092] In step 110, relevant information for longitudinal federated learning inference is determined for each of the multiple clients.

[0093] In some embodiments, the relevant information for VFL inference for each client includes at least one of the following: VFL association identifier for each client, first sample indication information, and first feature indication information.

[0094] By defining different relevant information for vertical federated learning inference for different clients, it makes it possible for multiple clients and servers to participate in vertical federated learning together, so that subsequent clients can use the collected data samples to perform inference at various levels to obtain a better performing model.

[0095] In some embodiments, the VFL association identifier is used to associate the VFL process and / or the VFL model; the first sample indication information is used to indicate the data sample; and / or the first feature indication information is used to indicate the data feature.

[0096] The VFL association identifier can be the same for different clients. The VFL association identifier can be used to let the server and other clients participating in longitudinal federated learning know which VFL task and / or VFL model it is. In addition, the VFL association identifier can also be associated with the VFL association identifier during the longitudinal federated learning training process.

[0097] By associating VFL processes and / or VFL models with VFL association identifiers, the accuracy of the vertical federated learning inference process is ensured, and the possibility of confusion in the vertical federated learning inference process is reduced.

[0098] In some embodiments, the first sample indication information is used to indicate data samples from the server and / or client. The first sample indication information includes at least one of the following: User Equipment ID (UE ID), GPSI (Generic Public Subscription Identifier), SUPI (Subscription Permanent Identifier), location information, Sample ID, and Group ID.

[0099] The first feature indication information is used to indicate the data features used by the server and / or the data features used by the client. The first feature indication information includes at least one of the following: service experience information, quality of experience metrics (QoE metrics), voice quality information (e.g., MOS (Mean Opinion Score)), and feature ID.

[0100] By using the first sample indication information and the first feature indication information to notify the server or client of the data sample and data feature to be used, the possibility of wasting resources caused by the server and multiple clients using the same data sample and data feature is reduced.

[0101] In some embodiments, the relevant information for VFL inference for each client further includes at least one of the following: an interoperability indicator, an analysis identifier, filter information, and a VFL indicator for each client. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of the following: single network slice selection assistance information, application identifier, region of interest, and data network access identifier.

[0102] The information can be filtered as needed when collecting data samples on the server and client sides.

[0103] For example, the following method can be used to determine the relevant information for vertical federated learning inference for each of multiple clients: According to preset rules, at least one of the sample indication information, feature indication information, and filter information corresponding to multiple clients is divided, and the relevant information for VFL inference for each client is determined. The preset rules include at least one of the following: the internal configuration of the server, the internal logic of the server, the network element type of each client, the relevant information of each client obtained from other network elements during the client search process, and the information associated with each client during the training process. The internal configuration and internal logic of the server can be, for example, specified by the operator. In addition, the network element type of the client can be, for example, AF or NWDAF. The relevant information of each client can be, for example, the service range of the client. The information associated with each client during the training process can be, for example, the data samples and / or data features that the client has, as detailed below.

[0104] According to preset rules, the user equipment identifier set in the sample indication information corresponding to multiple clients is divided, and the user equipment identifier set corresponding to each client in the first sample indication information is determined for each client.

[0105] According to preset rules, the sample identifiers in the sample indication information corresponding to multiple clients are divided, and the sample identifier corresponding to each client in the first sample indication information is determined for each client.

[0106] According to preset rules, the feature identifiers in the feature indication information corresponding to multiple clients are divided, and the feature identifier corresponding to each client in the first feature indication information is determined for each client.

[0107] According to preset rules, the experience quality index information in the feature indication information corresponding to multiple clients is divided, and the experience quality index information corresponding to each client in the first feature indication information is determined.

[0108] According to preset rules, the regions of interest in the filter information corresponding to multiple clients are divided, and the region of interest corresponding to each client in the filter information is determined.

[0109] Based on preset rules, the experience quality index information in the filter information corresponding to multiple clients is divided, and the experience quality index information corresponding to each client in the filter information is determined.

[0110] In step 120, a first request message is sent to each client so that each client can perform inference.

[0111] By initiating a vertical federated learning inference request from the server to the client, the server interacts with multiple clients to jointly complete the vertical federated learning inference process, which solves the problem in 5GC that it does not support the vertical federated learning inference process involving multiple clients.

[0112] In some embodiments, the first request information includes information related to VFL inference for each client.

[0113] In some embodiments, the server sends a first request message to each client based on the second request message sent by the receiving consuming NF (Network Function).

[0114] For example, the second request information is used to request the analysis result (the final inference result). The second request information includes at least one of an analysis identifier and a VFL indicator. The analysis identifier is used to identify that the requested analysis and / or the analysis result is generated using the VFL method, and the VFL indicator is used to indicate the model trained by VFL.

[0115] In some embodiments, an Analytics ID may indicate a request for an Analytics Service Experience.

[0116] In some embodiments, the analysis result is in the form of an analysis ID output, including at least one of the data in the analysis ID output, or including at least one of the experience quality index information, latency information, and 5G service quality index.

[0117] In some embodiments, the second request information further includes filter information, wherein the filter information includes at least one of S-NSSAI (Single Network Slice Selection Assistance Information), AID (Application Identifier), AOI (Area of ​​Interest), and DNAI (Data Network Access Identifier).

[0118] In some embodiments, filter information can be used, for example, to filter information when collecting data on the server and client sides, such as collecting only single network slice selection auxiliary information.

[0119] For example, the consumer NF is AnLF (Analytic Logic Function), and the server is the application function AF.

[0120] In the above embodiments, by determining relevant information for vertical federated learning inference corresponding to each of the multiple clients, wherein the relevant information for each client's vertical federated learning inference includes at least one of vertical federated learning association identifier, sample indication information, and feature indication information corresponding to each client. The vertical federated learning association identifier identifies the vertical federated learning process of each client, the sample indication information indicates data samples in the server, and the feature indication information indicates data features in the server and / or data features in each client. Then, a first request message including the relevant information for vertical federated learning inference corresponding to that client is sent to each client to request each client to perform vertical federated learning inference, thereby improving the feasibility of distributed inference in the vertical federated learning process.

[0121] In some embodiments, the server may also perform local VFL inference based on the data in the server to obtain intermediate results of local VFL inference.

[0122] In some embodiments, the data on the server can be local data on the server, data collected by the server, or a combination of both.

[0123] For example, the server collects data samples from the corresponding network element based on the first request information, so that the server can perform local VFL inference based on the data in the server and obtain the intermediate results of local VFL inference.

[0124] In some embodiments, the data samples collected by the server from the corresponding network element may be, for example, experience quality index information, location information, RAT (Radio Access Technology) access information, and PDU (Protocol Data Unit) session information.

[0125] In some embodiments, the server also receives the first response information sent by each client, and then generates the inference result based on the first response information sent by each client and / or the intermediate results of local VFL inference.

[0126] For example, inference results can be generated by aggregating the first response information sent by each client and / or the intermediate results of local VFL inference.

[0127] For example, the first response information corresponding to each client includes at least one of the following: the intermediate result of VFL inference corresponding to each client, the second sample indication information, the second feature indication information, and the VFL association identifier.

[0128] The intermediate results of VFL inference for each client can be, for example, gradients, losses, or other forms of data.

[0129] The client notifies the server which VFL inference process the first response information sent by the client belongs to by sending the VFL association identifier to the server.

[0130] The second sample indication information is used for data samples on the index server and / or client, wherein the second sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier.

[0131] The second feature indication information is used to indicate the data features used by the server and / or the data features used by the client. The second feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier.

[0132] By using second sample indication information and second feature indication information to notify the server or client of the data sample and data features used, the possibility of resource waste caused by the server and multiple clients using the same data sample and data features is reduced.

[0133] In addition, the first response information also includes first validity time information, which indicates the validity time of intermediate results of VFL inference.

[0134] Informing the server and client of the validity period of the intermediate results inferred by using the first validity period information in the first response information helps to ensure the efficiency of the vertical federated learning inference process and reduces the possibility of resource waste caused by the server not responding to the first response information for too long.

[0135] In some embodiments, when the server is an untrusted AF, the interaction between the server and clients (other NFs) within the 5GC needs to be through the NEF (Network Exposure Function). This avoids exposing information inside the 5GC to the outside, improving the security of information exchange and thus enhancing the security of VFL inference. Therefore, the server can also receive the first response information sent by the NEF.

[0136] NEF enables interaction between clients and untrusted servers, allowing multiple clients to participate in inference during the vertical federated learning inference process while ensuring the security of the core network.

[0137] If NEF has aggregation capabilities, the first response information can be obtained by aggregating at least one of the following: the intermediate result of VFL inference received by NEF from each client, the second sample indication information, the second feature indication information, and the VFL association identifier. Then, the server generates the inference result based on the first response information and / or the intermediate result of local VFL inference.

[0138] For example, inference results can be generated by aggregating the first response information aggregated by NEF and / or the intermediate results of local VFL inference.

[0139] If NEF does not have aggregation capabilities, the server can also receive the first response information sent by each client through the Network Development Function (NEF). The first response information for each client includes at least one of the following: the intermediate result of VFL inference for each client, the second sample indication information, the second feature indication information, and the VFL association identifier. Then, based on the first response information sent by each client and / or the intermediate result of local VFL inference, an inference result is generated.

[0140] For example, inference results are generated by aggregating the first response information sent by each client and / or the intermediate results of local VFL inference.

[0141] In some embodiments, if NEF does not have aggregation capabilities, NEF will place the first response information sent by each client in the same instruction or the same signaling message to forward the first response information sent by each client to the server.

[0142] After receiving the first response information, the server can send a second response information to the consumer NF. The second response information includes at least one of the analysis result and the second validity time information. The analysis result is generated based on the inference result, and the second validity time information is used to indicate the validity time of the inference result.

[0143] In some embodiments, the analysis results can be considered as the inference results from the server. The analysis results may be, for example, in the form of an analysis ID output, including at least one of the data items in the analysis ID output.

[0144] In some embodiments, the server is a Network Data Analysis Function (NWDAF), and each client is an Application Function (AF); or, the server and each client are different NWDAFs; or, the server is an AF and each client is an NWDAF.

[0145] The following describes, with reference to some embodiments, how to process the information exchanged when the AF is an untrusted AF.

[0146] When the AF is a non-trusted AF, sending the first request information to each client includes: sending the first request information to each client through the Network Open Function (NEF), wherein the NEF stores the mapping relationship between the internal VFL association identifier and the external VFL association identifier, the mapping relationship between the first internal sample indication information and the first external sample indication information, and the mapping relationship between the first internal feature indication information and the first external feature indication information.

[0147] Furthermore, when the server is an untrusted AF, the VFL association identifier in the first request information is an external VFL association identifier, the first sample indication information includes the first external sample indication information, and / or the first feature indication information includes the first external feature indication information. NEF maps the external VFL association identifier in the first request information to an internal VFL association identifier, maps the first external sample indication information to the first internal sample indication information, and / or maps the first external feature indication information to the first internal feature indication information.

[0148] In the case of multiple clients having untrusted AFs, for each client that is an untrusted AF, the VFL association identifier in the first request information is an internal VFL association identifier, the first sample indication information includes the first internal sample indication information, and / or the first feature indication information includes the first internal feature indication information. NEF maps the first internal VFL association identifier in the first request information to the first external VFL association identifier, maps the first internal sample indication information to the first external sample indication information, and / or maps the first internal feature indication information to the first external feature indication information.

[0149] In some embodiments, when multiple clients have untrusted AFs, the NEF stores the mapping relationship between internal intermediate results and external intermediate results, the mapping relationship between internal VFL association identifiers and external VFL association identifiers, the mapping relationship between second internal sample indication information and second external sample indication information, and the mapping relationship between second internal feature indication information and second external feature indication information.

[0150] For each client that is a non-trusted AF, the intermediate result of VFL inference in the first response information is an external intermediate result, the VFL association identifier is an external VFL association identifier, the second sample indication information includes the second external sample indication information, and / or the second feature indication information includes the second external feature indication information. NEF maps the external intermediate result in the first response information to an internal intermediate result, maps the external VFL association identifier to an internal VFL association identifier, maps the second external sample indication information to the second internal sample indication information, and / or maps the second external feature indication information to the second internal feature indication information.

[0151] When the server is an untrusted AF, the intermediate result of VFL inference in the first response information is an internal intermediate result, the VFL association identifier is an internal VFL association identifier, the second sample indication information includes the second internal sample indication information, and / or the second feature indication information includes the second internal feature indication information. NEF maps the internal intermediate result in the first response information to the external intermediate result, maps the internal VFL association identifier to the external VFL association identifier, maps the second internal sample indication information to the second external sample indication information, and / or maps the second internal feature indication information to the second external feature indication information.

[0152] By mapping internal and external information, the security of the core network is protected, the possibility of internal information leakage in the core network is reduced, and thus the security of the VFL inference process is improved.

[0153] Figure 2 Flowcharts illustrating some other embodiments of the reasoning method of the model of this disclosure are shown.

[0154] like Figure 2 As shown, the model's inference method includes steps 210 to 230, which is executed by NEF.

[0155] In step 210, the third request information sent by the server is received.

[0156] For example, the third request information includes at least one of the following: VFL association identifier, sample indication information corresponding to multiple clients, and feature indication information.

[0157] In some embodiments, VFL association identifiers are used to associate VFLs with processes and / or VFL models; sample indication information is used to indicate data samples; and / or feature indication information is used to indicate data features.

[0158] By receiving third-party request information from the server via NEF, direct interaction between the server and the client is avoided, ensuring the security of the core network and reducing the risk of information leakage within the core network.

[0159] In step 220, based on the third request information, relevant information for longitudinal federated learning inference is determined for each of the multiple clients.

[0160] By defining different relevant information for vertical federated learning inference for different clients, it makes it possible for multiple clients and servers to participate in vertical federated learning together, so that subsequent clients can use the collected data samples to perform inference at various levels to obtain a better performing model.

[0161] In some embodiments, the relevant information for VFL inference for each client includes at least one of the following: VFL association identifier, third sample indication information, and third feature indication information for each client.

[0162] The third sample indication information is used to indicate data samples of the server and / or client, wherein the third sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier.

[0163] The fourth sample indication information is used to indicate data samples of the server and / or client, wherein the fourth sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier.

[0164] The third feature indication information is used to indicate the data features used by the server and / or the data features used by the client. The third feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier.

[0165] The fourth feature indication information is used to indicate the data features used by the server and / or the data features used by the client. The fourth feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier.

[0166] In some embodiments, the relevant information for VFL inference for each client further includes at least one of the following: an interoperability indicator, an analysis identifier, filter information, and a VFL indicator for each client. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of the following: single network slice selection assistance information, application identifier, region of interest, and data network access identifier.

[0167] For example, based on preset rules, at least one of the sample indication information, feature indication information, and filter information corresponding to multiple clients can be divided to determine the relevant information for VFL inference for each client. The preset rules include at least one of the following: the server's internal configuration, the server's internal logic, the network element type of each client, the relevant information of each client obtained from other network elements during the client lookup process, and the information associated with each client during training. The specific division method is as described in the inference method of the server-side model and will not be repeated here.

[0168] In some embodiments, if NEF lacks partitioning capabilities, meaning it cannot determine the relevant information for vertical federated learning inference for each of the multiple clients based on the third request information, then the server can determine the relevant information for vertical federated learning inference for each of the multiple clients based on the third request information, thus obtaining the third sub-request information. In this case, NEF is responsible for receiving the third sub-request information sent by the server and then sending it to each client.

[0169] Furthermore, when the third request information is segmented by the server, the server will send multiple third sub-request messages to NEF, which will then forward them to each client. When the third request information is segmented by NEF, the server only needs to send one third request message to NEF.

[0170] In step 230, a third sub-request message is sent to each client so that each client can perform inference. The third sub-request message includes relevant information about VFL inference for each client.

[0171] By using NEF to enable interaction between the server and the client, the security of the core network is ensured and the risk of information leakage within the core network is reduced.

[0172] In some embodiments, NEF may also receive third sub-response information sent by each client, wherein the third sub-response information includes at least one of the following: intermediate results of VFL inference for each client, fourth feature indication information, fourth sample indication information, and VFL association identifier.

[0173] If NEF has aggregation capabilities, it will also aggregate the third sub-response information to obtain the third response information, and then send the third response information to the server so that the server can generate the inference result based on the third response information and / or the intermediate results of local VFL inference.

[0174] For example, the server can generate inference results by aggregating third-party response information and / or intermediate results of local VFL inference.

[0175] In some embodiments, the third response information further includes: third validity time information, used to indicate the validity time of intermediate results of VFL inference.

[0176] If NEF does not have aggregation capabilities, it will directly send the third sub-response information to the server so that the server can generate inference results based on the third sub-response information and / or the intermediate results of local VFL inference.

[0177] For example, NEF can send multiple third-party response messages to the server in the same instruction or signaling message.

[0178] For example, the server can aggregate third-party response information and / or intermediate results of local VFL inference to generate inference results.

[0179] In the vertical federated learning process, the server and client can be the following combinations of network elements: the server is a network data analysis function (NWDAF) and each client is an application function (AF); or the server and each client are different NWDAFs; or the server is an AF and each client is an NWDAF.

[0180] When the server is an untrusted AF or multiple clients have untrusted AFs, NEF will map the information included in the third sub-request information to ensure the security of the core network and reduce the possibility of information leakage. Specifically, NEF stores the mapping relationship between internal VFL association identifiers and external VFL association identifiers, the mapping relationship between third internal sample indication information and third external sample indication information, and the mapping relationship between third internal feature indication information and third external feature indication information.

[0181] How NEF maps the third sub-request information can be found in the reasoning method of the server-side model, which maps the first request information. It will not be elaborated here.

[0182] When the server is an untrusted AF or multiple clients have untrusted AFs, NEF will map the information included in the third response information to ensure the security of the core network and reduce the possibility of information leakage. NEF stores the mapping relationship between internal and external intermediate results, the mapping relationship between internal VFL association identifiers and external VFL association identifiers, the mapping relationship between third internal sample indication information and third external sample indication information, and the mapping relationship between third internal feature indication information and third external feature indication information.

[0183] How NEF maps the third response information can be found in the reasoning method of the server-side model, which maps the first response information. It will not be elaborated here.

[0184] When the server is an untrusted AF or multiple clients have untrusted AFs, NEF will map the information included in the third sub-response information to ensure the security of the core network and reduce the possibility of information leakage. Specifically, NEF stores the mapping relationship between internal and external intermediate results, the mapping relationship between internal VFL association identifiers and external VFL association identifiers, the mapping relationship between third internal sample indication information and third external sample indication information, and the mapping relationship between third internal feature indication information and third external feature indication information.

[0185] How NEF maps the third sub-response information can be found in the reasoning method of the server-side model, which maps the first response information. It will not be elaborated here.

[0186] By mapping internal and external information, the security of the core network is protected, and the possibility of leakage of internal information of the core network is reduced.

[0187] In the above embodiments, by determining relevant information for vertical federated learning inference corresponding to each of the multiple clients, wherein the relevant information for each client's vertical federated learning inference includes at least one of vertical federated learning association identifier, sample indication information, and feature indication information corresponding to each client. The vertical federated learning association identifier identifies the vertical federated learning process of each client, the sample indication information indicates data samples in the server, and the feature indication information indicates data features in the server and / or data features in each client. Then, a first request message including the relevant information for vertical federated learning inference corresponding to that client is sent to each client to request each client to perform vertical federated learning inference, thereby improving the feasibility of distributed inference in the vertical federated learning process.

[0188] Figure 3Flowcharts illustrating further embodiments of the reasoning methods of the models disclosed herein.

[0189] like Figure 3 As shown, the model's inference method includes steps 310 to 320, and the model's inference method is executed by the client.

[0190] In step 310, a first request message sent by the server is received. This first request message includes information related to the VFL inference for the client. This information includes at least one of the following: the client-specific VFL association identifier, first sample indication information, and first feature indication information. The relevant information for the client-specific VFL inference is partitioned by the server from the relevant information for multiple clients' VFL inferences.

[0191] In some embodiments, the VFL association identifier is used to associate the VFL process and / or the VFL model; the first sample indication information is used to indicate the data sample; and / or the first feature indication information is used to indicate the data feature.

[0192] In addition, the first sample indication information is used to indicate the data sample, wherein the first sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier.

[0193] The second sample indication information is used to indicate data samples, wherein the second sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier.

[0194] The first feature indication information is used to indicate the data features used by the server and / or the data features used by the client, wherein the first feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier.

[0195] The second feature indication information is used to indicate the data features used by the server and / or the data features used by the client. The second feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier.

[0196] For example, the relevant information for VFL inference corresponding to the client also includes at least one of the following: interoperability indicator, analysis identifier, filter information and VFL indicator corresponding to the client. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of the following: single network slice selection auxiliary information, application identifier, region of interest and data network access identifier.

[0197] When either the server or the client is an untrusted AF (Automatic Request), meaning direct interaction between the server and client is impossible, the client can receive the first request information sent by the server through NEF (Non-Trusted Request). If NEF has segmentation capabilities, it can segment the complete first request information sent by the server and then forward the segmented information to the client. If NEF does not have segmentation capabilities, it can directly receive the segmented first request information from the server and forward it to the client.

[0198] In step 320, VFL inference is performed based on the first request information.

[0199] In some embodiments, VFL inference based on the first request information also requires the client to collect data samples from the corresponding network element based on the first request information.

[0200] After the client performs VFL inference based on the first request information, it can obtain the first response information. At this time, the client can send the first response information to the server. The first response information includes at least one of the following: the intermediate result of the client's VFL inference, the second sample indication information, the second feature indication information, and the VFL association identifier.

[0201] In some embodiments, the first response information further includes: first validity time information, used to indicate the validity time of intermediate results of VFL inference.

[0202] Sending a first response message to the server helps the server generate inference results based on the first response message and / or intermediate results of local VFL inference.

[0203] In some embodiments, the server is a Network Data Analysis Function (NWDAF) and the client is an Application Function (AF); or, the server and the client are different NWDAFs; or, the server is an AF and the client is an NWDAF.

[0204] When either the server or the client is an untrusted AF (Automatic AF), meaning direct interaction between the server and client is impossible, the client can send its first response information to the NEF (Neural Frame Provider), which will then forward it to the server. If the NEF has aggregation capabilities, it can aggregate the first response information sent by the client and then forward the aggregated information to the server. If the NEF lacks aggregation capabilities, it can place multiple first response messages in the same instruction or signaling message and forward them to the server.

[0205] In addition, when the server is an untrusted AF or the client has an untrusted AF, that is, when the server and the client cannot interact directly, NEF can be used to map the first request information and the first response information to ensure the security of the core network.

[0206] First, NEF stores the mapping relationships between internal VFL association identifiers and external VFL association identifiers, the mapping relationships between first internal sample indication information and first external sample indication information, the mapping relationships between first internal feature indication information and first external feature indication information, and the mapping relationships between internal intermediate results and external intermediate results.

[0207] How NEF maps the first request information and the first response information can be found in the reasoning method of the server-side model, which will not be elaborated here.

[0208] In the above embodiments, by determining relevant information for vertical federated learning inference corresponding to each of the multiple clients, wherein the relevant information for each client's vertical federated learning inference includes at least one of vertical federated learning association identifier, sample indication information, and feature indication information corresponding to each client. The vertical federated learning association identifier identifies the vertical federated learning process of each client, the sample indication information indicates data samples in the server, and the feature indication information indicates data features in the server and / or data features in each client. Then, a first request message including the relevant information for vertical federated learning inference corresponding to that client is sent to each client to request each client to perform vertical federated learning inference, thereby improving the feasibility of distributed inference in the vertical federated learning process.

[0209] Figure 4 Flowcharts illustrating further embodiments of the reasoning methods of the models disclosed herein.

[0210] like Figure 4 As shown, the inference method of the model includes steps 410 to 420, and the inference method of the model is executed by the client.

[0211] In step 410, the third sub-request information sent by the Network Open Function (NEF) is received. The third sub-request information includes relevant information about the VFL inference corresponding to the client. The relevant information about the VFL inference corresponding to the client includes at least one of the VFL association identifier, third sample indication information, and third feature indication information corresponding to the client. The relevant information about the VFL inference corresponding to the client is divided for the client from the multiple relevant information about the VFL inference corresponding to the client obtained by NEF from the server.

[0212] In some embodiments, VFL association identifiers are used to associate VFLs with processes and / or VFL models; third sample indication information is used to indicate data samples; and / or third feature indication information is used to indicate data features.

[0213] In addition, the third sample indication information is used to indicate the data sample, wherein the third sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier.

[0214] The fourth sample indication information is used to indicate data samples, wherein the fourth sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier.

[0215] The third feature indication information is used to indicate the data features used by the server and / or the data features used by the client. The third feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier.

[0216] The fourth feature indication information is used to indicate the data features used by the server and / or the data features used by the client. The fourth feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier.

[0217] For example, the relevant information for VFL inference corresponding to the client also includes at least one of the following: interoperability indicator, analysis identifier, filter information and VFL indicator corresponding to the client. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of the following: single network slice selection auxiliary information, application identifier, region of interest and data network access identifier.

[0218] When NEF has the ability to partition, it can partition the complete third request information sent by the server and then forward the partitioned third sub-request information to the client. When NEF does not have the ability to partition, it can directly receive the partitioned third sub-request information from the server and forward it to the client.

[0219] In step 420, VFL inference is performed based on the third sub-request information.

[0220] In some embodiments, VFL inference based on third sub-request information also requires the client to obtain mobile data samples from the corresponding network element based on the third sub-request information.

[0221] After the client performs VFL inference based on the third sub-request information, it can obtain the third sub-response information. At this time, the client can send the third sub-response information to the server. The third sub-response information includes at least one of the client's intermediate VFL inference results, fourth feature indication information, fourth sample indication information, and VFL association identifier, so that NEF can generate response information based on the third sub-response information and send it to the server.

[0222] In some embodiments, the third sub-response information further includes: third validity time information, used to indicate the validity time of intermediate results of VFL inference.

[0223] Sending a third sub-response message to the server helps the server generate inference results based on the third sub-response message and / or intermediate results of local VFL inference.

[0224] In some embodiments, the server is a Network Data Analysis Function (NWDAF) and the client is an Application Function (AF); or, the server and the client are different NWDAFs; or, the server is an AF and the client is an NWDAF.

[0225] When NEF has aggregation capabilities, it can aggregate the third sub-response information sent by the client and then forward the aggregated third sub-response information to the server. When NEF does not have aggregation capabilities, it can put multiple third sub-response information in the same instruction or the same signaling message and forward it to the server.

[0226] In addition, when the server is an untrusted AF or the client has an untrusted AF, that is, when the server and the client cannot interact directly, the NEF can be used to map the third request information and the third sub-response information to ensure the security of the core network, as follows.

[0227] First, NEF stores the mapping relationships between internal VFL association identifiers and external VFL association identifiers, the mapping relationships between third internal sample indication information and third external sample indication information, the mapping relationships between third internal feature indication information and third external feature indication information, and the mapping relationships between internal intermediate results and external intermediate results.

[0228] How NEF maps the third sub-request information can be found in the reasoning method of the server-side model, which maps the first request information. It will not be elaborated here.

[0229] How NEF maps the third sub-response information and / or the third response information can be found in the reasoning method of the server-side model, which maps the first response information. It will not be elaborated here.

[0230] In the above embodiments, by determining relevant information for vertical federated learning inference corresponding to each of the multiple clients, wherein the relevant information for each client's vertical federated learning inference includes at least one of vertical federated learning association identifier, sample indication information, and feature indication information corresponding to each client. The vertical federated learning association identifier identifies the vertical federated learning process of each client, the sample indication information indicates data samples in the server, and the feature indication information indicates data features in the server and / or data features in each client. Then, a first request message including the relevant information for vertical federated learning inference corresponding to that client is sent to each client to request each client to perform vertical federated learning inference, thereby improving the feasibility of distributed inference in the vertical federated learning process.

[0231] Figure 5 Schematic diagrams illustrating some embodiments of the server side of this disclosure are shown.

[0232] like Figure 5 As shown, the server 50 includes a first determining unit 51 and a first sending unit 52.

[0233] The first determining unit 51 is configured to determine relevant information for longitudinal federated learning VFL inference for each of the multiple clients, wherein the relevant information for VFL inference for each client includes at least one of VFL association identifier, first sample indication information and first feature indication information for each client.

[0234] In some embodiments, the VFL association identifier is used to associate the VFL process and / or the VFL model; the first sample indication information is used to indicate the data sample; and / or the first feature indication information is used to indicate the data feature.

[0235] In some embodiments, the first sample indication information is used to indicate a data sample; the second sample indication information is used to indicate a data sample; the first sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier; and / or the second sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier.

[0236] In some embodiments, the first feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the second feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the first feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier, and / or the second feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier.

[0237] In some embodiments, the relevant information for VFL inference for each client further includes at least one of the following: an interoperability indicator, an analysis identifier, filter information, and a VFL indicator for each client. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of the following: single network slice selection assistance information, application identifier, region of interest, and data network access identifier.

[0238] In some embodiments, the first determining unit 51 is further configured to divide at least one of the sample indication information, feature indication information, and filter information corresponding to multiple clients according to preset rules, and determine the relevant information for VFL inference for each client. The preset rules include at least one of the following: the internal configuration of the server, the internal logic of the server, the network element type of each client, the relevant information of each client obtained from other network elements during the client search process, and the information associated with each client during the training process.

[0239] In some embodiments, the first determining unit 51 is further configured to perform at least one of the following: dividing the set of user equipment identifiers in the sample indication information corresponding to multiple clients according to a preset rule, and determining the set of user equipment identifiers in the first sample indication information corresponding to each client for each client; dividing the sample identifiers in the sample indication information corresponding to multiple clients according to a preset rule, and determining the sample identifiers in the first sample indication information corresponding to each client for each client; dividing the feature identifiers in the feature indication information corresponding to multiple clients according to a preset rule, and determining the feature identifiers in the first feature indication information corresponding to each client for each client; dividing the experience quality index information in the feature indication information corresponding to multiple clients according to a preset rule, and determining the experience quality index information in the first feature indication information corresponding to each client for each client; dividing the region of interest in the filter information corresponding to multiple clients according to a preset rule, and determining the region of interest in the filter information corresponding to each client for each client; dividing the experience quality index information in the filter information corresponding to multiple clients according to a preset rule, and determining the experience quality index information in the filter information corresponding to each client for each client.

[0240] The first sending unit 52 is configured to send a first request message to each client so that each client can perform inference, wherein the first request message includes information related to VFL inference for each client. In some embodiments, the server is a Network Data Analysis Function (NWDAF) and each client is an Application Function (AF); or, the server and each client are different NWDAFs; or, the server is an AF and each client is an NWDAF.

[0241] In some embodiments, when the AF is an untrusted AF, the first sending unit 52 is further configured to send a first request message to each client via the Network Open Function (NEF). In the case where the server is an untrusted AF, the VFL association identifier in the first request message is an external VFL association identifier, the first sample indication information includes first external sample indication information, and / or the first feature indication information includes first external feature indication information. The NEF maps the external VFL association identifier in the first request message to an internal VFL association identifier, maps the first external sample indication information to first internal sample indication information, and / or maps the first external feature indication information to first internal feature indication information. The feature indication information is mapped to the first internal feature indication information. In the case of multiple clients having untrusted AFs, for each client that is an untrusted AF, the VFL association identifier in the first request information is an internal VFL association identifier, the first sample indication information includes the first internal sample indication information, and / or the first feature indication information includes the first internal feature indication information. NEF maps the first internal VFL association identifier in the first request information to the first external VFL association identifier, maps the first internal sample indication information to the first external sample indication information, and / or maps the first internal feature indication information to the first external feature indication information.

[0242] In some embodiments, the server 50 further includes a first collection unit, wherein the first collection unit is configured to collect data samples from the corresponding network element according to the first request information.

[0243] In some embodiments, the server 50 further includes a third inference unit, wherein the third inference unit is configured to perform local VFL inference based on data in the server to obtain intermediate results of local VFL inference.

[0244] In some embodiments, the server 50 further includes a fourth receiving unit, wherein the fourth receiving unit is configured to receive first response information sent by each client, wherein the first response information corresponding to each client includes at least one of the intermediate result of VFL inference corresponding to each client, second sample indication information, second feature indication information, and VFL association identifier; and generate inference results based on the first response information sent by each client and / or the intermediate result of local VFL inference.

[0245] In some embodiments, the first response information further includes: first validity time information, used to indicate the validity time of intermediate results of VFL inference.

[0246] In some embodiments, the fourth receiving unit is further configured to receive first response information sent by the Network Open Function (NEF), wherein the first response information is obtained by aggregating at least one of the intermediate results of VFL inference for each client, second sample indication information, second feature indication information, and VFL association identifier received by the NEF; and to generate inference results based on the first response information and / or the intermediate results of local VFL inference.

[0247] In some embodiments, the fourth receiving unit is further configured to receive first response information sent by each client via the Network Development Function (NEF), wherein the first response information corresponding to each client includes at least one of the following: intermediate result of VFL inference for each client, second sample indication information, second feature indication information, and VFL association identifier; and to generate inference results based on the first response information sent by each client and / or the intermediate result of local VFL inference.

[0248] For example, server 50 also includes a first aggregation unit, which is configured to aggregate the first response information sent by each client and the intermediate results of local VFL inference to generate inference results.

[0249] In some embodiments, the fourth receiving unit is further configured to receive a second request message sent by a consumer network function (NF), wherein the second request message is used to request analysis results, and the second request message includes at least one of an analysis identifier and a VFL indicator, wherein the analysis identifier is used to identify that the requested analysis and / or the analysis results are generated using a VFL method, and the VFL indicator is used to indicate the model trained by the VFL, wherein sending the first request message to each client includes: sending the first request message to each client according to the second request message.

[0250] In some embodiments, the analysis result is in the form of an analysis ID output, including at least one of the data in the analysis ID output, or including at least one of the experience quality index information, latency information, and 5G service quality index.

[0251] In some embodiments, the consumer NF is the analysis logic function AnLF, and the server is the application function AF.

[0252] In some embodiments, the second request information further includes filter information, wherein the filter information includes at least one of a single network slice selection assistance information, an application identifier, a region of interest, and a data network access identifier.

[0253] In some embodiments, the first sending unit 52 is further configured to send second response information to the consuming NF, wherein the second response information includes at least one of analysis result and second validity time information, the analysis result being generated based on the inference result, and the second validity time information being used to indicate the validity time of the inference result.

[0254] In some embodiments, the server 50 further includes a first mapping unit, wherein the first mapping unit is configured to, in the case where multiple clients have untrusted AFs, for each client that is an untrusted AF, map the intermediate result of VFL inference in the first response information to an external intermediate result, the VFL association identifier to an external VFL association identifier, the second sample indication information to include second external sample indication information, and / or the second feature indication information to include second external feature indication information, and map the external intermediate result in the first response information to an internal intermediate result, the external VFL association identifier to an internal VFL association identifier, and the second external sample indication information to second internal sample indication information through NEF. / Or, the second external feature indication information is mapped to the second internal feature indication information. In the case that the server is an untrusted AF, the intermediate result of VFL inference in the first response information is an internal intermediate result, the VFL association identifier is an internal VFL association identifier, the second sample indication information includes the second internal sample indication information, and / or the second feature indication information includes the second internal feature indication information. The internal intermediate result in the first response information is mapped to the external intermediate result, the internal VFL association identifier is mapped to the external VFL association identifier, the second internal sample indication information is mapped to the second external sample indication information, and / or the second internal feature indication information is mapped to the second external feature indication information through NEF.

[0255] In the above embodiments, by determining relevant information for vertical federated learning inference corresponding to each of the multiple clients, wherein the relevant information for each client's vertical federated learning inference includes at least one of vertical federated learning association identifier, sample indication information, and feature indication information corresponding to each client. The vertical federated learning association identifier identifies the vertical federated learning process of each client, the sample indication information indicates data samples in the server, and the feature indication information indicates data features in the server and / or data features in each client. Then, a first request message including the relevant information for vertical federated learning inference corresponding to that client is sent to each client to request each client to perform vertical federated learning inference, thereby improving the feasibility of distributed inference in the vertical federated learning process.

[0256] Figure 6 Schematic diagrams illustrating some embodiments of the NEF disclosed herein are shown.

[0257] like Figure 6 As shown, the NEF60 includes a first receiving unit 61, a second determining unit 62, and a second transmitting unit 63.

[0258] The first receiving unit 61 is configured to receive third request information sent by the server, wherein the third request information includes at least one of VFL association identifier, sample indication information corresponding to multiple clients, and feature indication information.

[0259] In some embodiments, VFL association identifiers are used to associate VFLs with processes and / or VFL models; sample indication information is used to indicate data samples; and / or feature indication information is used to indicate data features.

[0260] In some embodiments, the relevant information for VFL inference for each client further includes at least one of the following: an interoperability indicator, an analysis identifier, filter information, and a VFL indicator for each client. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of the following: single network slice selection assistance information, application identifier, region of interest, and data network access identifier.

[0261] The second determining unit 62 is configured to determine relevant information for longitudinal federated learning VFL inference for each of the multiple clients based on the third request information, wherein the relevant information for VFL inference for each client includes at least one of VFL association identifier, third sample indication information and third feature indication information for each client.

[0262] In some embodiments, the third sample indication information is used to indicate a data sample; the fourth sample indication information is used to indicate a data sample; the third sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier; and / or the fourth sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier.

[0263] In some embodiments, the third feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the fourth feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the third feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier, and / or the fourth feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier.

[0264] In some embodiments, the second determining unit 62 is further configured to divide at least one of the sample indication information, feature indication information, and filter information corresponding to multiple clients according to preset rules, and determine the relevant information for VFL inference for each client. The preset rules include at least one of the following: the internal configuration of the server, the internal logic of the server, the network element type of each client, the relevant information of each client obtained from other network elements during the client search process, and the information associated with each client during the training process.

[0265] In some embodiments, the second determining unit 62 is further configured to perform at least one of the following: dividing the set of user equipment identifiers in the sample indication information corresponding to multiple clients according to a preset rule, and determining the set of user equipment identifiers in the third sample indication information corresponding to each client for each client; dividing the sample identifiers in the sample indication information corresponding to multiple clients according to a preset rule, and determining the sample identifiers in the third sample indication information corresponding to each client for each client; dividing the feature identifiers in the feature indication information corresponding to multiple clients according to a preset rule, and determining the feature identifiers in the third feature indication information corresponding to each client for each client; dividing the experience quality index information in the feature indication information corresponding to multiple clients according to a preset rule, and determining the experience quality index information in the third feature indication information corresponding to each client for each client; dividing the region of interest in the filter information corresponding to multiple clients according to a preset rule, and determining the region of interest in the filter information corresponding to each client for each client; dividing the experience quality index information in the filter information corresponding to multiple clients according to a preset rule, and determining the experience quality index information in the filter information corresponding to each client for each client.

[0266] The second sending unit 63 is configured to send a third sub-request message to each client so that each client can perform inference, wherein the third sub-request message includes relevant information about VFL inference for each client.

[0267] In some embodiments, the first receiving unit 61 is further configured to receive third sub-response information sent by each client, wherein the third sub-response information includes at least one of the following: intermediate results of VFL inference for each client, fourth feature indication information, fourth sample indication information, and VFL association identifier; if NEF has aggregation capability, the third sub-response information is aggregated to obtain third response information; and the third response information is sent to the server so that the server generates inference results based on the third response information and / or the intermediate results of local VFL inference.

[0268] In some embodiments, the third response information further includes: third validity time information, used to indicate the validity time of intermediate results of VFL inference.

[0269] In some embodiments, the second sending unit 63 is further configured to send the third sub-response information to the server when NEF does not have aggregation capability, so that the server can generate inference results based on the third sub-response information and / or the intermediate results of local VFL inference.

[0270] In some embodiments, the server is a Network Data Analysis Function (NWDAF), and each client is an Application Function (AF); or, the server and each client are different NWDAFs; or, the server is an AF and each client is an NWDAF.

[0271] In some embodiments, NEF further includes a second mapping unit, wherein the second mapping unit is configured to, when the server is an untrusted AF, map the external VFL association identifier in the third sub-request information to an internal VFL association identifier, the third external sample indication information to a third internal sample indication information, and / or the third feature indication information to a third external feature indication information, and / or map the third external feature indication information to a third internal feature indication information, wherein, when multiple clients are untrusted AFs, for each client that is an untrusted AF, the internal VFL association identifier in the third sub-request information is mapped to an external VFL association identifier, the third internal sample indication information is mapped to a third external sample indication information, and / or the third internal feature indication information is mapped to a third external feature indication information, and / or the third feature indication information is mapped to a third external feature indication information, wherein, when multiple clients are untrusted AFs, for each client that is an untrusted AF, the internal VFL association identifier in the third sub-request information is mapped to an internal VFL association identifier, the third sample indication information is mapped to a third internal sample indication information, and / or the third internal feature indication information is mapped to a third external feature indication information, wherein, in the case of an untrusted AF, the internal VFL association identifier in the third sub-request information is mapped to an external VFL association identifier, the third internal sample indication information is mapped to a third external sample indication information, and / or the third internal feature indication information is mapped to a third external feature indication information, wherein, in the case of an untrusted AF, the third sample indication information is mapped to an internal ...

[0272] In some embodiments, the second mapping unit is further configured to, in the case where multiple clients have untrusted AFs, map the intermediate result of VFL inference in the third response information as an external intermediate result, the VFL association identifier as an external VFL association identifier, the fourth sample indication information including fourth external sample indication information, and / or the fourth feature indication information including fourth external feature indication information, for each client of the untrusted AF, the intermediate result of VFL inference in the third response information is an external intermediate result, the external VFL association identifier is an internal VFL association identifier, the fourth external sample indication information is a fourth external sample indication information, and / or the fourth external feature indication information is a fourth external feature indication information, mapping the external intermediate result in the third response information to an internal intermediate result, mapping the external VFL association identifier to an internal VFL association identifier, mapping the fourth external sample indication information to a fourth internal sample indication information, and / or mapping the fourth external feature indication information to a fourth external feature indication information. The feature indication information is mapped to the fourth internal feature indication information. In the case that the server is an untrusted AF, the intermediate result of VFL inference in the third response information is an internal intermediate result, the VFL association identifier is an internal VFL association identifier, the fourth sample indication information includes the fourth internal sample indication information, and / or the fourth feature indication information includes the fourth internal feature indication information. The internal intermediate result in the third response information is mapped to the external intermediate result, the internal VFL association identifier is mapped to the external VFL association identifier, the fourth internal sample indication information is mapped to the fourth external sample indication information, and / or the fourth internal feature indication information is mapped to the fourth external feature indication information.

[0273] In some embodiments, the second mapping unit is further configured to, in the case where multiple clients have untrusted AFs, map the intermediate result of VFL inference in the third sub-response information as an external intermediate result, the VFL association identifier as an external VFL association identifier, the fourth sample indication information including fourth external sample indication information, and / or the fourth feature indication information including fourth external feature indication information, for each client that is an untrusted AF, the external intermediate result in the third sub-response information is mapped to an internal intermediate result, the external VFL association identifier is mapped to an internal VFL association identifier, the fourth external sample indication information is mapped to a fourth internal sample indication information, and / or the fourth external feature indication information is mapped to a fourth external feature indication information. The feature indication information is mapped to the fourth internal feature indication information. In the case that the server is an untrusted AF, the intermediate result of VFL inference in the third sub-response information is an internal intermediate result, the VFL association identifier is an internal VFL association identifier, the fourth sample indication information includes the fourth internal sample indication information, and / or the fourth feature indication information includes the fourth internal feature indication information. The internal intermediate result in the third sub-response information is mapped to the external intermediate result, the internal VFL association identifier is mapped to the external VFL association identifier, the fourth internal sample indication information is mapped to the fourth external sample indication information, and / or the fourth internal feature indication information is mapped to the fourth external feature indication information.

[0274] In the above embodiments, by determining relevant information for vertical federated learning inference corresponding to each of the multiple clients, wherein the relevant information for each client's vertical federated learning inference includes at least one of vertical federated learning association identifier, sample indication information, and feature indication information corresponding to each client. The vertical federated learning association identifier identifies the vertical federated learning process of each client, the sample indication information indicates data samples in the server, and the feature indication information indicates data features in the server and / or data features in each client. Then, a first request message including the relevant information for vertical federated learning inference corresponding to that client is sent to each client to request each client to perform vertical federated learning inference, thereby improving the feasibility of distributed inference in the vertical federated learning process.

[0275] Figure 7 Schematic diagrams illustrating some embodiments of the client of this disclosure are shown.

[0276] like Figure 7 As shown, the client 70 includes a second receiving unit 71 and a first inference unit 72.

[0277] The second receiving unit 71 is configured to receive first request information sent by the server. The first request information includes relevant information about VFL inference corresponding to the client. The relevant information about VFL inference corresponding to the client includes at least one of VFL association identifier, first sample indication information and first feature indication information corresponding to the client. The relevant information about VFL inference corresponding to the client is divided by the server for the client from the relevant information about VFL inference corresponding to multiple clients.

[0278] In some embodiments, the VFL association identifier is used to associate the VFL process and / or the VFL model; the first sample indication information is used to indicate the data sample; and / or the first feature indication information is used to indicate the data feature.

[0279] In some embodiments, the first sample indication information is used to indicate a data sample; the second sample indication information is used to indicate a data sample; the first sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier; and / or the second sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier.

[0280] In some embodiments, the first feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the second feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the first feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier, and / or the second feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier.

[0281] In some embodiments, the VFL inference information corresponding to the client further includes at least one of the following: interoperability indicator, analysis identifier, filter information and VFL indicator corresponding to the client; the analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL; the VFL indicator is used to indicate the model trained by VFL; and the filter information includes at least one of single network slice selection assistance information, application identifier, region of interest and data network access identifier.

[0282] The first inference unit 72 is configured to perform VFL inference based on the first request information.

[0283] In some embodiments, the client 70 further includes a second collection unit, wherein the second collection unit is configured to collect data samples from the corresponding network element according to the first request information.

[0284] In some embodiments, the client 70 further includes a third sending unit, wherein the third sending unit is configured to send a first response information to the server, wherein the first response information includes at least one of the following: intermediate result of VFL inference corresponding to the client, second sample indication information, second feature indication information, and VFL association identifier.

[0285] In some embodiments, the first response information further includes: first validity time information, used to indicate the validity time of intermediate results of VFL inference.

[0286] In some embodiments, the server is a Network Data Analysis Function (NWDAF) and the client is an Application Function (AF); or, the server and the client are different NWDAFs; or, the server is an AF and the client is an NWDAF.

[0287] In the above embodiments, by determining relevant information for vertical federated learning inference corresponding to each of the multiple clients, wherein the relevant information for each client's vertical federated learning inference includes at least one of vertical federated learning association identifier, sample indication information, and feature indication information corresponding to each client. The vertical federated learning association identifier identifies the vertical federated learning process of each client, the sample indication information indicates data samples in the server, and the feature indication information indicates data features in the server and / or data features in each client. Then, a first request message including the relevant information for vertical federated learning inference corresponding to that client is sent to each client to request each client to perform vertical federated learning inference, thereby improving the feasibility of distributed inference in the vertical federated learning process.

[0288] Figure 8 Schematic diagrams illustrating other embodiments of the client of this disclosure are shown.

[0289] like Figure 8 As shown, the client 80 includes a third receiving unit 81 and a second inference unit 82.

[0290] The third receiving unit 81 is configured to receive third sub-request information sent by the Network Open Function (NEF). The third sub-request information includes relevant information about the VFL inference corresponding to the client. The relevant information about the VFL inference corresponding to the client includes at least one of the VFL association identifier, third sample indication information, and third feature indication information corresponding to the client. The relevant information about the VFL inference corresponding to the client is divided for the client from multiple relevant information about the VFL inference corresponding to the client obtained by NEF from the server.

[0291] In some embodiments, VFL association identifiers are used to associate VFLs with processes and / or VFL models; third sample indication information is used to indicate data samples; and / or third feature indication information is used to indicate data features.

[0292] In some embodiments, the third sample indication information is used to indicate a data sample; the fourth sample indication information is used to indicate a data sample; the third sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier; and / or the fourth sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier.

[0293] In some embodiments, the third feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the fourth feature indication information is used to indicate the data features used by the server and / or the data features used by the client; the third feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier, and / or the fourth feature indication information includes at least one of service experience information, experience quality index information, voice quality information, and feature identifier.

[0294] In some embodiments, the VFL inference information corresponding to the client further includes at least one of the following: interoperability indicator, analysis identifier, filter information and VFL indicator corresponding to the client; the analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL; the VFL indicator is used to indicate the model trained by VFL; and the filter information includes at least one of single network slice selection assistance information, application identifier, region of interest and data network access identifier.

[0295] The second reasoning unit 82 is configured to perform reasoning based on the third sub-request information.

[0296] In some embodiments, the client 80 further includes a fourth sending unit, wherein the fourth sending unit is configured to send a third sub-response information to the NEF, wherein the third sub-response information includes at least one of the client's VFL inference intermediate results, fourth feature indication information, fourth sample indication information, and VFL association identifier, so that the NEF generates response information based on the third sub-response information and sends it to the server.

[0297] In some embodiments, the server is a Network Data Analysis Function (NWDAF) and the client is an Application Function (AF); or, the server and the client are different NWDAFs; or, the server is an AF and the client is an NWDAF.

[0298] In the above embodiments, by determining relevant information for vertical federated learning inference corresponding to each of the multiple clients, wherein the relevant information for each client's vertical federated learning inference includes at least one of vertical federated learning association identifier, sample indication information, and feature indication information corresponding to each client. The vertical federated learning association identifier identifies the vertical federated learning process of each client, the sample indication information indicates data samples in the server, and the feature indication information indicates data features in the server and / or data features in each client. Then, a first request message including the relevant information for vertical federated learning inference corresponding to that client is sent to each client to request each client to perform vertical federated learning inference, thereby improving the feasibility of distributed inference in the vertical federated learning process.

[0299] Figure 9 Schematic diagrams illustrating some embodiments of the electronic devices disclosed herein.

[0300] like Figure 9 As shown, the electronic device 90 of this embodiment includes a memory 91 and a processor 92 coupled to the memory 91. The processor 92 is configured to execute the inference method of the model in any of the foregoing embodiments based on instructions stored in the memory 91.

[0301] The memory 91 may include, for example, system memory, fixed non-volatile storage media, etc. The system memory stores, for example, the operating system, application programs, boot loader, and other programs.

[0302] Electronic device 90 may also include input / output interfaces 93, network interfaces 94, and storage interfaces 95. These interfaces 93, 94, and 95, as well as the memory 91 and processor 92, can be connected via, for example, a bus 96. Specifically, input / output interface 93 provides a connection interface for input / output devices such as monitors, mice, keyboards, touchscreens, microphones, and speakers. Network interface 94 provides a connection interface for various networked devices. Storage interface 95 provides a connection interface for external storage devices such as SD cards and USB flash drives.

[0303] Figure 10 Schematic diagrams illustrating some embodiments of the reasoning methods of the models disclosed herein.

[0304] like Figure 10 As shown, the reasoning method of the model includes steps S11 to S17.

[0305] In step S11, the consuming NF sends a second request to the server to request the vertical federated learning inference process.

[0306] In some embodiments, the second request information includes at least one of an analysis identifier and a VFL indicator, wherein the analysis identifier is used to identify that the requested analysis and / or the analysis results are generated using the VFL method, and the VFL indicator is used to indicate the model trained by VFL.

[0307] In some embodiments, the second request information further includes filter information, wherein the filter information includes at least one of a single network slice selection assistance information, an application identifier, a region of interest, and a data network access identifier.

[0308] In step S12, the server sends a first request message to each client.

[0309] The first request information in step S12 includes relevant information on vertical federated learning inference determined by the server for each of the multiple clients. The relevant information on VFL inference for each client includes at least one of VFL association identifier, first sample indication information, and first feature indication information for each client. The first request information is sent to each client so that each client can perform inference.

[0310] In step S13, the client collects data samples from the corresponding NF based on the received first request information to perform longitudinal federated learning inference.

[0311] In step S14, the server and client perform VFL inference.

[0312] In step S14a, the client performs VFL inference based on the collected data samples.

[0313] In step S14b, the server performs VFL inference based on the data in the server.

[0314] The data on the server can be, for example, local data on the server and / or data collected by the server.

[0315] In step S15, each client sends a first response information to the server, wherein the first response information corresponding to each client includes at least one of the following: the intermediate result of VFL inference corresponding to each client, the second sample indication information, the second feature indication information, and the VFL association identifier.

[0316] In step S16, the server aggregates the intermediate results of each client and / or the intermediate results of the server's inference to obtain the second response information.

[0317] The second response information includes at least one of the analysis result and the second validity time information, wherein the analysis result is generated based on the reasoning result, and the second validity time information is used to indicate the validity time of the reasoning result.

[0318] In step S17, a second response message is sent to the consumer NF.

[0319] In the above embodiments, the server initiates vertical federated learning inference requests to multiple clients, interacts with multiple clients, and completes the vertical federated learning inference process, thus solving the problem that 5GC does not support multiple clients participating in the vertical federated learning inference process.

[0320] Figure 11 Schematic diagrams illustrating some other embodiments of the reasoning method of the model of this disclosure.

[0321] like Figure 11As shown, the reasoning method of the model includes steps S21 to S28.

[0322] In step S21, the server sends the third request information or the third sub-request information to NEF.

[0323] When NEF has the capability to partition and partitions the data, the server sends a third request message to NEF, and only one third request message needs to be sent.

[0324] If NEF does not have the ability to segment, the third request information is segmented by the server to obtain multiple third sub-request information. The server sends the third sub-request information to NEF, and multiple third sub-request information needs to be sent to NEF.

[0325] In step S22, NEF maps the parameters as follows.

[0326] If the partitioning is performed by NEF, then NEF needs to partition the third request information to obtain the third sub-request information and map the third sub-request information; or map the third request information and then partition the mapped third request information.

[0327] If the partitioning is performed by the server, then NEF receives the third sub-request information, and NEF needs to map the third sub-request information.

[0328] In step S23, NEF sends the mapped third sub-request information to each client.

[0329] In step S24, the client collects data samples from the corresponding BF based on the third sub-request information.

[0330] In step S25, the server and client perform VFL inference.

[0331] In step S25a, the client performs VFL inference based on the collected data samples.

[0332] In step S25b, the server performs VFL inference based on the server's data. This data can be, for example, local server data and / or data collected by the server.

[0333] In step S26, the client sends the intermediate results of its inference to NEF via a third sub-response message.

[0334] In step S27, NEF maps the parameters as follows.

[0335] If aggregation is performed through NEF, NEF sends a third response message to the server. NEF can first map the third sub-response message and then aggregate the mapped third sub-response message to obtain the third response message; or NEF can first aggregate the third sub-response message and then map the aggregated third sub-response message to obtain the third response message.

[0336] If aggregation is performed through the server, NEF only maps the third sub-response information to obtain the mapped third sub-response information, and then NEF sends the mapped third sub-response information to the server.

[0337] In step S28, NEF sends the third response information or the mapped third sub-response information to the server. Then, the server aggregates the third response information or the mapped third sub-response information with the intermediate results of the server's inference to obtain the final inference result.

[0338] In the above embodiments, NEF is used to realize the interaction between the server and the client. The server initiates vertical federated learning inference requests to multiple clients through NEF, interacts with multiple clients, and completes the vertical federated learning inference process, thus solving the problem that 5GC does not support multiple clients participating in the vertical federated learning inference process. In addition, when the server is an untrusted AF, the security of the core network is also guaranteed, reducing the risk of information leakage within the core network.

[0339] Figure 12 Schematic diagrams illustrating further embodiments of the reasoning methods of the models disclosed herein.

[0340] like Figure 12 As shown, the reasoning method of the model includes steps S310 to S319.

[0341] like Figure 12 As shown, the server is NWDAF, and the clients include NWDAF and AF. AF is an untrusted AF, meaning that the client (AF) and the server cannot interact directly, as detailed below.

[0342] In step S310, the server sends a first request message or a first sub-request message to NEF.

[0343] If NEF performs partitioning, the server sends the first request information to NEF. If the server performs partitioning, the server sends the partitioned first request information to NEF, i.e., the first sub-request information.

[0344] In step S311, NEF performs parameter mapping, as follows.

[0345] If the partitioning is performed by NEF, then NEF needs to partition the first request information to obtain the first sub-request information and map the first sub-request information; or map the first request information and then partition the mapped first request information.

[0346] If the partitioning is performed by the server, then NEF receives the first sub-request information, and NEF needs to map the first sub-request information.

[0347] In step S312, NEF sends the mapped first sub-request information to the client (AF).

[0348] In step S313, the server directly sends the first sub-request information to the client (NWDAF), wherein the first sub-request information is the information divided by the server according to the first request information.

[0349] In step S314, the client (NWDAF) collects data samples from the corresponding NF based on the first sub-request information.

[0350] In step S315, the client (NWDAF), the client (AF), and the server perform VFL inference.

[0351] In step S315a, the client (NWDAF) performs VFL inference based on the collected data samples.

[0352] In step S315b, the client (AF) performs VFL inference based on the client (AF) data.

[0353] In step S315c, the server performs VFL inference based on the server's data. This data can be, for example, local server data and / or data collected by the server.

[0354] In step S316, the client (AF) sends the intermediate results of its inference to the NEF via the first sub-response information.

[0355] In step S317, NEF maps the parameters as follows.

[0356] If aggregation is performed through NEF, NEF sends the first response information to the server. NEF can first map the first sub-response information and then aggregate the mapped first sub-response information to obtain the first response information; or NEF can first aggregate the first sub-response information and then map the aggregated first sub-response information to obtain the first response information.

[0357] If aggregation is performed through the server, NEF only maps the first sub-response information to obtain the mapped first sub-response information, and then NEF sends the mapped first sub-response information to the server.

[0358] In step S318, NEF sends the first response information or the mapped first sub-response information to the server.

[0359] In step S319, the client (NWDAF) sends its intermediate inference results to the server via the first sub-response information. Then, the server aggregates the intermediate inference results from the client (NWDAF), the client (AF), and the server to obtain the final inference result.

[0360] In the above embodiments, NEF is used to realize the interaction between the server and the client. The server initiates vertical federated learning inference requests to multiple clients through NEF, interacts with multiple clients, and completes the vertical federated learning inference process, thus solving the problem that 5GC does not support multiple clients participating in the vertical federated learning inference process. In addition, when there is an untrusted AF on the client, the security of the core network is also guaranteed, reducing the risk of information leakage within the core network.

[0361] In some embodiments, a computer program product is protected, comprising a computer program or instructions that, when executed by a processor, implement the registration selection method described above. The computer program product includes a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a CPU, it performs the functions defined in the methods of the embodiments of this disclosure.

[0362] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable non-transitory storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0363] The inference method, server, NEF, client, computer-readable storage medium, and computer program product of the model disclosed herein have been described in detail. To avoid obscuring the concept of this disclosure, some details known in the art have not been described. Those skilled in the art will fully understand how to implement the technical solutions disclosed herein based on the above description.

[0364] The methods and systems of this disclosure may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the methods is for illustrative purposes only, and the steps of the methods of this disclosure are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, this disclosure may also be implemented as a program recorded on a recording medium, the program including machine-readable instructions for implementing the methods according to this disclosure. Thus, this disclosure also covers recording media storing programs for performing the methods according to this disclosure.

[0365] While specific embodiments of this disclosure have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of this disclosure. Those skilled in the art should understand that modifications can be made to the above embodiments without departing from the scope and spirit of this disclosure. The scope of this disclosure is defined by the appended claims.

Claims

1. A reasoning method for a model, executed by a server, comprising: For each of the multiple clients, determine the relevant information for vertical federated learning VFL inference, wherein the relevant information for the VFL inference corresponding to each client includes at least one of the VFL association identifier, first sample indication information and first feature indication information corresponding to each client; A first request message is sent to each client so that each client can perform inference, wherein the first request message includes relevant information about the VFL inference corresponding to each client.

2. The reasoning method according to claim 1, wherein, The VFL association identifier is used to associate the VFL process and / or the VFL model; The first sample indication information is used to indicate the data sample; and / or The first feature indication information is used to indicate data features.

3. The reasoning method according to claim 1, wherein, The relevant information for VFL inference corresponding to each client also includes at least one of the following: interoperability indicator, analysis identifier, filter information, and VFL indicator corresponding to each client. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of single network slice selection assistance information, application identifier, region of interest, and data network access identifier.

4. The reasoning method according to claim 3, wherein, The process of determining relevant information for longitudinal federated learning VFL inference for each of the multiple clients includes: According to preset rules, at least one of the sample indication information, feature indication information, and filter information corresponding to the multiple clients is divided to determine the relevant information of the VFL inference for each client. The preset rules include at least one of the following: the internal configuration of the server, the internal logic of the server, the network element type of each client, the relevant information of each client obtained from other network elements during the client search process, and the information associated with each client during the training process.

5. The reasoning method according to claim 4, wherein, According to preset rules, at least one of the sample indication information, feature indication information, and filter information corresponding to the multiple clients is divided, and the relevant information for the VFL inference of each client is determined, including at least one of the following: According to the preset rules, the user equipment identifier set in the sample indication information corresponding to the multiple clients is divided, and the user equipment identifier set in the first sample indication information corresponding to each client is determined for each client; According to the preset rules, the sample identifiers in the sample indication information corresponding to the multiple clients are divided, and a sample identifier in the first sample indication information corresponding to each client is determined for each client. According to the preset rules, the feature identifiers in the feature indication information corresponding to the multiple clients are divided, and the feature identifier in the first feature indication information corresponding to each client is determined for each client; According to the preset rules, the experience quality index information in the feature indication information corresponding to the multiple clients is divided, and the experience quality index information in the first feature indication information corresponding to each client is determined for each client. According to the preset rules, the regions of interest in the filter information corresponding to the multiple clients are divided, and the region of interest in the filter information corresponding to each client is determined for each client. According to the preset rules, the experience quality index information in the filter information corresponding to the multiple clients is divided, and the experience quality index information corresponding to each client in the filter information is determined for each client.

6. The reasoning method according to claim 1 further includes: VFL inference is performed based on the data in the server to obtain the intermediate results of the local VFL inference.

7. The reasoning method according to claim 1 further includes: Receive first response information sent by each client, wherein the first response information corresponding to each client includes at least one of the following: intermediate result of VFL inference corresponding to each client, second sample indication information, second feature indication information, and VFL association identifier; Inference results are generated based on the first response information sent by each client and / or the intermediate results of local VFL inference.

8. The reasoning method according to claim 1 further includes: The system receives a first response message sent by the Network Open Function (NEF), wherein the first response message is obtained by aggregating at least one of the following: the intermediate result of the VFL inference corresponding to each client, the second sample indication information, the second feature indication information, and the VFL association identifier. Based on the first response information and / or the intermediate results of local VFL inference, an inference result is generated.

9. The reasoning method according to claim 1 further includes: The system receives a second request message sent by the Consumer Network Function (NF), wherein the second request message is used to request analysis results, and the second request message includes at least one of an analysis identifier and a VFL indicator, wherein the analysis identifier is used to identify that the requested analysis and / or the analysis results are generated using the VFL method, and the VFL indicator is used to indicate the model trained by the VFL. The step of sending the first request information to each client includes: Based on the second request information, the first request information is sent to each client.

10. The reasoning method according to claim 9, wherein, The second request information also includes filter information, wherein the filter information includes at least one of a single network slice selection assistance information, an application identifier, a region of interest, and a data network access identifier.

11. The reasoning method according to claim 9, further comprising: Send a second response message to the consuming NF, wherein the second response message includes at least one of the analysis result and a second validity period information, wherein the analysis result is generated based on the inference result, and the validity period information is used to indicate the validity period of the inference result.

12. The reasoning method according to claim 7, wherein, The first response information further includes: first validity time information, used to indicate the validity time of the intermediate results of the VFL inference.

13. The reasoning method according to claim 7, wherein, The first sample indication information is used to indicate the data sample; The second sample indication information is used to indicate data samples; The first sample indication information includes at least one of the following: user equipment identifier, general public subscriber identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier, and / or The second sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier.

14. The reasoning method according to claim 7, wherein, The first feature indication information is used to indicate the data features used by the server and / or the data features used by the client; The second feature indication information is used to indicate the data features used by the server and / or the data features used by the client; The first feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier, and / or The second feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier.

15. The reasoning method according to claim 1, further comprising: Data samples are collected from the corresponding network element based on the first request information.

16. The reasoning method according to claim 9, wherein, The analysis results are in the form of analysis ID output, including at least one of the data in the analysis ID output, or including at least one of the experience quality index information, latency information, and 5G service quality index.

17. The reasoning method according to claim 9, wherein, The consumer NF is the analysis logic function AnLF, and the server is the application function AF.

18. The reasoning method according to any one of claims 1 to 17, wherein, The server is a Network Data Analysis Function (NWDAF), and each client is an Application Function (AF). Alternatively, the server and each client are different NWDAFs, or the server is an AF and each client is an NWDAF.

19. The reasoning method according to claim 18, wherein, In the case that the AF is a non-trusted AF, sending the first request information to each client includes: The first request information is sent to each client via the Network Openness Function (NEF). Wherein, in the case that the server is the non-trusted AF, the VFL association identifier in the first request information is an external VFL association identifier, the first sample indication information includes first external sample indication information, and / or the first feature indication information includes first external feature indication information. The NEF maps the external VFL association identifier in the first request information to an internal VFL association identifier, maps the first external sample indication information to first internal sample indication information, and / or maps the first external feature indication information to first internal feature indication information. In the case where the multiple clients have the non-trusted AF, for each client that is a non-trusted AF, the VFL association identifier in the first request information is an internal VFL association identifier, the first sample indication information includes first internal sample indication information, and / or the first feature indication information includes first internal feature indication information. The NEF maps the first internal VFL association identifier in the first request information to a first external VFL association identifier, maps the first internal sample indication information to a first external sample indication information, and / or maps the first internal feature indication information to a first external feature indication information.

20. The reasoning method according to claim 18, further comprising: Receive first response information sent by each client via the Network Development Function (NEF), wherein the first response information corresponding to each client includes at least one of the following: intermediate result of VFL inference corresponding to each client, second sample indication information, second feature indication information, and VFL association identifier; Inference results are generated based on the first response information sent by each client and / or the intermediate results of local VFL inference.

21. The reasoning method according to claim 20, further comprising: In the case where multiple clients have the untrusted AF, for each client that is an untrusted AF, the intermediate result of the VFL inference in the first response information is an external intermediate result, the VFL association identifier is an external VFL association identifier, the second sample indication information includes second external sample indication information, and / or the second feature indication information includes second external feature indication information. The NEF maps the external intermediate result in the first response information to an internal intermediate result, maps the external VFL association identifier to an internal VFL association identifier, maps the second external sample indication information to a second internal sample indication information, and / or maps the second external feature indication information to a second internal feature indication information. When the server is the untrusted AF, the intermediate result of the VFL inference in the first response information is an internal intermediate result, the VFL association identifier is an internal VFL association identifier, the second sample indication information includes second internal sample indication information, and / or the second feature indication information includes second internal feature indication information. The NEF maps the internal intermediate result in the first response information to the external intermediate result, maps the internal VFL association identifier to the external VFL association identifier, maps the second internal sample indication information to the second external sample indication information, and / or maps the second internal feature indication information to the second external feature indication information.

22. The reasoning method according to claim 7, wherein, The step of generating inference results based on the first response information sent by each client and / or the intermediate results of local VFL inference includes: The inference result is generated by aggregating the first response information sent by each client and the intermediate results of the local VFL inference.

23. A reasoning method for a model, performed by the Network Open Function (NEF), comprising: Receive third request information sent by the server, wherein the third request information includes at least one of VFL association identifier, sample indication information corresponding to multiple clients, and feature indication information; Based on the third request information, relevant information for longitudinal federated learning VFL inference is determined for each of the plurality of clients, wherein the relevant information for VFL inference corresponding to each client includes at least one of VFL association identifier, third sample indication information and third feature indication information corresponding to each client; A third sub-request message is sent to each client so that each client can perform inference, wherein the third sub-request message includes relevant information about the VFL inference corresponding to each client.

24. The reasoning method according to claim 23, wherein, The VFL association identifier is used to associate the VFL process and / or the VFL model; The sample indication information is used to indicate data samples; and / or The feature indication information is used to indicate data features.

25. The reasoning method according to claim 23, wherein, The relevant information for the VFL inference corresponding to each client also includes at least one of the following: interoperability indicator, analysis identifier, filter information, and VFL indicator corresponding to each client. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of the following: single network slice selection assistance information, application identifier, region of interest, and data network access identifier.

26. The reasoning method according to claim 25, wherein, The step of determining the relevant information for longitudinal federated learning VFL inference for each of the plurality of clients based on the third request information includes: According to preset rules, at least one of the sample indication information, feature indication information, and filter information corresponding to the multiple clients is divided to determine the relevant information of the VFL inference for each client. The preset rules include at least one of the following: the internal configuration of the server, the internal logic of the server, the network element type of each client, the relevant information of each client obtained from other network elements during the client search process, and the information associated with each client during the training process.

27. The reasoning method according to claim 26, wherein, According to preset rules, at least one of the sample indication information, feature indication information, and filter information corresponding to the multiple clients is divided, and the relevant information for the VFL inference of each client is determined, including at least one of the following: According to the preset rules, the user equipment identifier set in the sample indication information corresponding to the multiple clients is divided, and the user equipment identifier set in the third sample indication information corresponding to each client is determined for each client; According to the preset rules, the sample identifiers in the sample indication information corresponding to the multiple clients are divided, and the sample identifier in the third sample indication information corresponding to each client is determined for each client; According to the preset rules, the feature identifiers in the feature indication information corresponding to the multiple clients are divided, and the feature identifier in the third feature indication information corresponding to each client is determined for each client; According to the preset rules, the experience quality index information in the feature indication information corresponding to the multiple clients is divided, and the experience quality index information in the third feature indication information corresponding to each client is determined for each client. According to the preset rules, the regions of interest in the filter information corresponding to the multiple clients are divided, and the region of interest in the filter information corresponding to each client is determined for each client. According to the preset rules, the experience quality index information in the filter information corresponding to the multiple clients is divided, and the experience quality index information corresponding to each client in the filter information is determined for each client.

28. The reasoning method according to claim 23 further includes: Receive a third sub-response information sent by each client, wherein the third sub-response information includes at least one of the following for each client: intermediate result of VFL inference, fourth feature indication information, fourth sample indication information, and VFL association identifier; If the NEF has aggregation capabilities, the third sub-response information is aggregated to obtain the third response information; The third response information is sent to the server so that the server can generate inference results based on the third response information and / or the intermediate results of local VFL inference.

29. The reasoning method according to claim 28, wherein, The third response information also includes: third validity time information, used to indicate the validity time of the intermediate results of the VFL inference.

30. The reasoning method according to claim 23, wherein, The third sample indication information is used to indicate data samples; The fourth sample indication information is used to indicate data samples; The third sample indication information includes at least one of the following: user equipment identifier, general public subscriber identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier, and / or The fourth sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier.

31. The reasoning method according to claim 23, wherein, The third feature indication information is used to indicate the data features used by the server and / or the data features used by the client; The fourth feature indication information is used to indicate the data features used by the server and / or the data features used by the client; The third feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier, and / or The fourth feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier.

32. The reasoning method according to any one of claims 23 to 31, wherein, The server is a Network Data Analysis Function (NWDAF), and each client is an Application Function (AF). Alternatively, the server and each client are different NWDAFs, or the server is an AF and each client is an NWDAF.

33. The reasoning method according to claim 32 further includes: In the case where the server is the untrusted AF, the VFL association identifier in the third sub-request information is an external VFL association identifier, the third sample indication information includes third external sample indication information, and / or the third feature indication information includes third external feature indication information. The external VFL association identifier in the third sub-request information is mapped to an internal VFL association identifier, the third external sample indication information is mapped to third internal sample indication information, and / or the third external feature indication information is mapped to third internal feature indication information. In the case where the multiple clients have the untrusted AF, for each client that is an untrusted AF, the VFL association identifier in the third sub-request information is an internal VFL association identifier, the third sample indication information includes third internal sample indication information, and / or the third feature indication information includes third internal feature indication information. The internal VFL association identifier in the third sub-request information is mapped to an external VFL association identifier, the third internal sample indication information is mapped to a third external sample indication information, and / or the third internal feature indication information is mapped to a third external feature indication information.

34. The reasoning method according to claim 28 further includes: If the NEF does not have aggregation capabilities, the third sub-response information is sent to the server so that the server can generate inference results based on the third sub-response information and / or the intermediate results of local VFL inference.

35. The reasoning method according to claim 28, further comprising: In the case where multiple clients have the untrusted AF, for each client that is an untrusted AF, the intermediate result of the VFL inference in the third sub-response information is an external intermediate result, the VFL association identifier is an external VFL association identifier, the fourth sample indication information includes fourth external sample indication information, and / or the fourth feature indication information includes fourth external feature indication information. The external intermediate result in the third response information is mapped to an internal intermediate result, the external VFL association identifier is mapped to an internal VFL association identifier, the fourth external sample indication information is mapped to a fourth internal sample indication information, and / or the fourth external feature indication information is mapped to a fourth internal feature indication information. In the case where the server is the untrusted AF, the intermediate result of the VFL inference in the third sub-response information is an internal intermediate result, the VFL association identifier is an internal VFL association identifier, the fourth sample indication information includes fourth internal sample indication information, and / or the fourth feature indication information includes fourth internal feature indication information. The internal intermediate result in the third response information is mapped to the external intermediate result, the internal VFL association identifier is mapped to the external VFL association identifier, the fourth internal sample indication information is mapped to the fourth external sample indication information, and / or the fourth internal feature indication information is mapped to the fourth external feature indication information.

36. A reasoning method for a model, executed by a client, comprising: The server receives a first request message, wherein the first request message includes relevant information about the VFL inference corresponding to the client, and the relevant information about the VFL inference corresponding to the client includes at least one of the VFL association identifier, first sample indication information and first feature indication information corresponding to the client, and the relevant information about the VFL inference corresponding to the client is divided by the server for the client from the relevant information about VFL inference corresponding to multiple clients; VFL inference is performed based on the first request information.

37. The reasoning method according to claim 36, wherein, The VFL association identifier is used to associate the VFL process and / or the VFL model; The first sample indication information is used to indicate the data sample; and / or The first feature indication information is used to indicate data features.

38. The reasoning method according to claim 36 further includes: Send a first response message to the server, wherein the first response message includes at least one of the following: the intermediate result of the VFL inference corresponding to the client, the second sample indication information, the second feature indication information, and the VFL association identifier.

39. The reasoning method according to claim 36, wherein, The VFL inference information corresponding to the client also includes at least one of the following: interoperability indicator, analysis identifier, filter information, and VFL indicator. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of single network slice selection assistance information, application identifier, region of interest, and data network access identifier.

40. The reasoning method according to claim 38, wherein, The first response information also includes: first validity time information, used to indicate the validity time of the intermediate results of the VFL inference.

41. The reasoning method according to claim 38, wherein, The first sample indication information is used to indicate the data sample; The second sample indication information is used to indicate data samples; The first sample indication information includes at least one of the following: user equipment identifier, general public subscriber identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier, and / or The second sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier.

42. The reasoning method according to claim 38, wherein, The first feature indication information is used to indicate the data features used by the server and / or the data features used by the client; The second feature indication information is used to indicate the data features used by the server and / or the data features used by the client; The first feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier, and / or The second feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier.

43. The reasoning method according to any one of claims 37 to 43, wherein, The server is a network data analysis function (NWDAF), and the client is an application function (AF). Alternatively, the server and the client are different NWDAFs, or the server is an AF and the client is an NWDAF.

44. A reasoning method for a model, executed by a client, comprising: The system receives a third sub-request information sent by the Network Open Function (NEF), wherein the third sub-request information includes relevant information about the VFL inference corresponding to the client, and the relevant information about the VFL inference corresponding to the client includes at least one of the VFL association identifier, third sample indication information, and third feature indication information corresponding to the client, and the relevant information about the VFL inference corresponding to the client is divided for the client from the multiple relevant information about VFL inference corresponding to the client obtained by the NEF from the server; Reasoning is performed based on the third sub-request information.

45. The reasoning method according to claim 44, wherein, The VFL association identifier is used to associate the VFL process and / or the VFL model; The third sample indication information is used to indicate data samples; and / or The third feature indication information is used to indicate data features.

46. ​​The reasoning method according to claim 44 further includes: The client sends a third sub-response information to the NEF, wherein the third sub-response information includes at least one of the following: intermediate result of the VFL inference of the client, fourth feature indication information, fourth sample indication information, and VFL association identifier, so that the NEF generates a third response information based on the third sub-response information and sends it to the server.

47. The reasoning method according to claim 44, wherein, The relevant information for the VFL inference corresponding to the client also includes at least one of the following: interoperability indicator, analysis identifier, filter information, and VFL indicator corresponding to the client. The analysis identifier is used to identify the requested analysis or to identify the requested analysis and / or the analysis result is generated using VFL. The VFL indicator is used to indicate the model trained by VFL. The filter information includes at least one of the following: single network slice selection assistance information, application identifier, region of interest, and data network access identifier.

48. The reasoning method according to claim 46, wherein, The third sample indication information is used to indicate data samples; The fourth sample indication information is used to indicate data samples; The third sample indication information includes at least one of the following: user equipment identifier, general public subscriber identifier (GPSI), subscription perpetual identifier (SUPI), location information, sample identifier, and group identifier, and / or The fourth sample indication information includes at least one of the following: user equipment identifier, general public user identifier (GPSI), subscription permanent identifier (SUPI), location information, sample identifier, and group identifier.

49. The reasoning method according to claim 46, wherein, The third feature indication information is used to indicate the data features used by the server and / or the data features used by the client; The fourth feature indication information is used to indicate the data features used by the server and / or the data features used by the client; The third feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier, and / or The fourth feature indication information includes at least one of the following: service experience information, experience quality index information, voice quality information, and feature identifier.

50. The reasoning method according to any one of claims 44 to 49, wherein, The server is a network data analysis function (NWDAF), and the client is an application function (AF). Alternatively, the server and the client are different NWDAFs, or the server is an AF and the client is an NWDAF.

51. A server, comprising: The first determining unit is configured to determine relevant information for longitudinal federated learning VFL inference for each of a plurality of clients, wherein the relevant information for VFL inference corresponding to each client includes at least one of VFL association identifier, first sample indication information and first feature indication information corresponding to each client; The first sending unit is configured to send a first request message to each client so that each client can perform inference, wherein the first request message includes relevant information about the VFL inference corresponding to each client.

52. A NEF comprising: The first receiving unit is configured to receive third request information sent by the server, wherein the third request information includes at least one of VFL association identifier, sample indication information corresponding to multiple clients, and feature indication information; The second determining unit is configured to determine relevant information for longitudinal federated learning VFL inference for each of the plurality of clients based on the third request information, wherein the relevant information for VFL inference corresponding to each client includes at least one of VFL association identifier, third sample indication information and third feature indication information corresponding to each client; The second sending unit is configured to send a third sub-request information to each client so that each client can perform inference, wherein the third sub-request information includes relevant information about the VFL inference corresponding to each client.

53. A client, comprising: The second receiving unit is configured to receive first request information sent by the server, wherein the first request information includes relevant information of the VFL inference corresponding to the client, and the relevant information of the VFL inference corresponding to the client includes at least one of the VFL association identifier, first sample indication information and first feature indication information corresponding to the client, and the relevant information of the VFL inference corresponding to the client is divided by the server for the client from the relevant information of VFL inference corresponding to multiple clients; The first inference unit is configured to perform VFL inference based on the first request information.

54. A client, comprising: The third receiving unit is configured to receive third sub-request information sent by the Network Open Function (NEF), wherein the third sub-request information includes relevant information about the VFL inference corresponding to the client, and the relevant information about the VFL inference corresponding to the client includes at least one of the VFL association identifier, third sample indication information and third feature indication information corresponding to the client, and the relevant information about the VFL inference corresponding to the client is divided for the client from the multiple VFL inference relevant information corresponding to the client obtained by the NEF from the server; The second reasoning unit is configured to perform reasoning based on the third sub-request information.

55. An electronic device, comprising: Memory; and A processor coupled to the memory, the processor being configured to execute the inference method of any one of claims 1 to 50 based on instructions stored in the memory.

56. A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the reasoning method of any one of claims 1 to 50.

57. A computer program product comprising a computer program that, when executed by a processor, implements the reasoning method of any one of claims 1 to 50.

Images