Methods, apparatus and computer programs related to vertical federated learning
By leveraging VFL client profile management and data alignment technology on user devices, the privacy protection issue of sample alignment in vertical federated learning is resolved, enabling effective data alignment and model training, and improving training efficiency.
Patent Information
- Application Number
- CN202610631419.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2025-05-09
- Filing Date
- 2026-05-09
- Publication Date
- 2026-11-10
AI Technical Summary
In vertical federated learning, existing technologies struggle to effectively address the challenge of selecting user devices with different characteristics of the same samples for training sample alignment without compromising participant privacy, especially in cellular or mobile telecommunications systems.
Data alignment between user devices is achieved through VFL client profile management on user devices, including feature availability requests and responses, hash value exchange, and model updates, combined with confidence matrix and integrity protection.
It enables effective data alignment and model training between user devices without compromising privacy, improving training efficiency and privacy protection.
Smart Images

Figure CN123074945A_ABST
Abstract
Description
Technical Field
[0001] Various examples of this disclosure relate to methods, apparatuses, and computer programs, particularly but not limited to methods, apparatuses, and computer programs related to vertical federated learning in telecommunications networks. Background Technology
[0002] A communication network can be viewed as a facility that enables communication between two or more communication devices, or provides communication devices with access to a data network. Mobile or wireless communication networks are an example of communication networks. Communication devices can be provided with services through application servers.
[0003] Communication networks can operate according to standards such as those issued by 3GPP (3rd Generation Partnership Project) or ETSI (European Telecommunications Standards Institute). Examples of standards provided by 3GPP are the so-called 3GPP standards for different generations of cellular technology, such as the 3GPP standards for 4G technology and the 3GPP standards for 5G technology. Communication networks operating according to 3GPP standards can be called 3GPP networks. Summary of the Invention
[0004] Some examples of this disclosure will be described with respect to certain aspects. These aspects are not intended to indicate key or essential features of the various examples of this disclosure, nor are they intended to be used to limit the scope of this disclosure. In view of this disclosure, other features, aspects, and elements will be apparent to those skilled in the art. For example, it should be understood that additional aspects may be provided by any combination of two or more of the aspects described below.
[0005] According to a first aspect, a method is provided, comprising: providing a vertical federated learning (VFL) client profile for an AI machine learning enabling client of a user device, the VFL client profile including information about one or more features of a data sample of the user device, wherein the user device is a candidate user device for participating in a vertical federated learning machine learning task; receiving from a first server a request to change the availability of one or more features of the data sample of the user device, the request including identifiers of one or more features of the data sample of the user device; changing the availability of one or more features according to the request; and sending a response to the request to the first server, the response indicating that the availability of one or more features has been changed.
[0006] The method may include: receiving from a first server a request for one or more hash values of one or more feature values of one or more data samples, and sending a response to the first server to the request for one or more hash values of one or more hash values of one or more feature values of one or more data samples.
[0007] When a user device is selected to participate in a vertical federated learning task, the method may include: providing information to a first server related to a model trained in the vertical federated learning task; and receiving updated information from the first server that updates the model trained in the vertical federated learning task.
[0008] According to a second aspect, a method is provided, comprising: for multiple user devices serving as candidate user devices for participating in a vertical federated learning machine learning task; receiving a vertical federated learning (VFL) client profile of an AI machine learning enablement client for a respective user device, the VFL client profile including information about one or more features of a data sample of the respective user device; sending a request to the respective user device to change the availability of one or more features of the data sample of the respective user device, the request including identifiers of one or more features of the data sample of the respective user device; and receiving a response from the respective user device to the request, the response indicating that the state of one or more features has been changed.
[0009] The method may include: determining one or more common features among one or more features provided by different user devices among a plurality of user devices, and selecting one or more user devices among the plurality of user devices as candidates for a vertical federated learning task based on one or more common features among one or more features provided by different user devices among a plurality of user devices.
[0010] The method may include: sending a corresponding request to a candidate user device for a vertical federated learning task based on one or more common features, the corresponding request being for one or more hash values of one or more feature values of one or more data samples, and receiving a response to the request from a corresponding user device, the request being for one or more hash values of one or more feature values of one or more data samples.
[0011] The method may include: using hash values from different user devices to determine a measurement of the alignment of data from different user devices.
[0012] Measurements of alignment of data from different user devices can include a confidence matrix.
[0013] One or more of the following may be used together with one or both of the methods of the first aspect and the second aspect.
[0014] Information about one or more features of a data sample may include information about the availability of the respective feature among the one or more features.
[0015] Information about the availability of a corresponding feature for one or more features can indicate whether the corresponding feature is inactive or deactivated.
[0016] Information about the availability of one or more features can indicate whether the corresponding feature is in a state of either available or unavailable.
[0017] Information about one or more features of the data sample may include information about the expiration of availability for the corresponding feature among the one or more features.
[0018] The request may include information indicating the priority of a corresponding feature in a vertical federated learning machine learning task, and changes in availability for the priority of that corresponding feature are requested.
[0019] The request may include information indicating the availability of the corresponding feature and the time at which it will be applied.
[0020] The request may include information indicating the time at which a change in the state of the corresponding feature will be applied.
[0021] The request may include information indicating whether the state of the corresponding feature will be set to active or deactivated to change the state of the corresponding feature.
[0022] The request may include information indicating whether the state of the corresponding feature will be set to available or unavailable to change the state of the corresponding feature.
[0023] The response may include information indicating the extent to which the request to change the availability of one or more of the features is met.
[0024] The response includes information indicating the timing of any changes to the availability of the corresponding feature and the adjustments that will be applied.
[0025] According to a third aspect, a method is provided, comprising: receiving a first request for information regarding one or more features of a data sample of a user device; sending a response to the first request to a first server, the response including information regarding one or more features of the data sample of the user device, wherein the user device is a candidate user device for participating in a vertical federated learning task; receiving a second request from the first server for one or more integrity-protected values of one or more features of the data sample of the user device; and in response to the first request, providing the first server with one or more integrity-protected values of one or more features of the data sample of the user device.
[0026] When a user device is selected to participate in a vertical federated learning task, the method may include: providing information to a first server related to a model trained in the vertical federated learning task; and receiving updated information from the first server that updates the model trained in the vertical federated learning task.
[0027] According to a fourth aspect, a method is provided, comprising: for a plurality of user devices serving as candidate user devices for participating in a vertical federated learning machine learning task; sending a first request to a respective user device for information regarding one or more features of a data sample of the respective user device; receiving a response from the respective user device to the first request, the response including information regarding one or more features of the data sample of the respective user device; sending a second request to the respective user device for one or more integrity-protected values of one or more features of the data sample of the respective user device; and in response to the second request, receiving one or more integrity-protected values of one or more features of the data sample of the user device.
[0028] The method may include: identifying one or more common features among one or more features of different user devices among a plurality of user devices, and selecting one or more user devices among the plurality of user devices as candidates for a vertical federated learning task based on the one or more common features.
[0029] The method may include: using integrity protection values from different user equipment to determine one or more measurements of alignment of data from different user equipment.
[0030] One or more measurements of alignment of data from different user devices may include a confidence matrix.
[0031] The method may include sending a message to a vertical application layer server, the message including information about one or more measurements of alignment of data from different user devices from the user device.
[0032] The messages sent to the vertical application layer server may also include information about one or more standards for one or more measurements used to align data from different user devices based on vertical federated learning tasks.
[0033] The method may include: receiving from a vertical application layer server a response to a message sent to the vertical application layer server, the response including information about a minimum confidence threshold required to participate in a vertical federated learning task; and using the information about the minimum confidence threshold to select one or more user devices for evaluating the corresponding model training capabilities of the respective user devices.
[0034] One or more of the following may be used in one or both of the methods of the third aspect and the methods of the fourth aspect.
[0035] The value for integrity protection can include a hash value.
[0036] The first request may include information about the vertical federated learning task.
[0037] The first request includes information instructing the management process for the requested feature.
[0038] The requested feature management process may include one or more of the following: feature preparation for vertical federated learning tasks; or feature importance analysis.
[0039] The first request may indicate that information is requested regarding one or more of the following: the data type of one or more features; the dimensions of the representation of one or more features; or privacy requirements associated with one or more features.
[0040] A response to the first request may include information about one or more of the following: the data type of one or more features; the dimensions of the representation of one or more features; or privacy requirements associated with one or more features.
[0041] The methods of any one of the first to fourth aspects can be executed by the device.
[0042] The apparatus may include components for performing the method.
[0043] The device may include at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the device to perform at least any of the methods described above.
[0044] The means of the first and / or third aspects may be user equipment or provided in user equipment.
[0045] The apparatus of the second and / or fourth aspects may be a server, provided in a server, or provided by a server.
[0046] According to another aspect, a computer-readable medium is provided, including program instructions stored thereon for performing at least one of the methods described above.
[0047] According to one aspect, a non-volatile computer-readable medium is provided, including program instructions stored thereon for performing at least one of the methods described above.
[0048] According to one aspect, a non-volatile tangible storage medium is provided, including program instructions stored thereon for performing at least one of the methods described above.
[0049] According to one aspect, a program is provided, including instructions that, when executed by at least one processor, cause any of the methods described above to be performed.
[0050] Many different aspects have been described above. It should be understood that additional aspects may be provided by combination of any two or more aspects (or their corresponding portions) described above or in other parts of this document.
[0051] Various other aspects are also described in the following detailed description and claims. Attached Figure Description
[0052] Some examples will now be described only by way of non-limiting and illustrative examples with reference to one or more accompanying drawings, in which:
[0053] Figure 1 An example is shown that the examples disclosed in this paper can be applied to its vertical federated learning VFL architecture;
[0054] Figure 2 The apparatus of some embodiments is shown;
[0055] Figure 3 The first process of some embodiments is shown;
[0056] Figure 4 The second process of some embodiments is shown;
[0057] Figure 5 The third process of some embodiments is shown;
[0058] Figure 6 A first method of some embodiments is shown;
[0059] Figure 7 A second method is shown in some embodiments;
[0060] Figure 8 A third method is shown in some embodiments; and
[0061] Figure 9 A fourth method is shown in some embodiments. Detailed Implementation
[0062] This article further mentions federated learning, and more specifically, vertical federated learning (VFL).
[0063] Federated learning is a method for performing machine learning across multiple entities in a decentralized manner. In other words, federated learning can be viewed as a method that enables each of multiple entities to collaborate with other entities to train a globally shared model without uploading local data to a server for centralized training. Therefore, in federated learning, multiple entities are configured to collaboratively train a model, while the data used to train the model locally is retained within each entity. Consequently, in federated learning, there is no centralized data repository for training the model. In the following text, a federated learning architecture can be viewed as a network configuration in which at least one federated learning method can be implemented.
[0064] There are two types of federated learning: horizontal federated learning and vertical federated learning (VFL). The following discussion considers vertical federated learning in cellular or mobile telecommunications systems (often referred to as telecommunications systems).
[0065] In VFL, one problem is to find training samples with different features that have the same sample among training participants (e.g., user equipment (UE) of a telecommunications system) in a way that protects the privacy of the participants and their data during training (i.e., to achieve sample alignment).
[0066] Sample alignment is the process of selecting UEs with different features associated with the same samples. This ensures that data from different UEs can be used together to train machine learning models in VFL.
[0067] In VFL, collaborating entities (also known as parties or participants) may possess different features of the same sample, and different parties can jointly supplement the feature space of the same sample to train an ML model together to provide the desired output. The desired output can be prediction, classification, and / or another desired output.
[0068] For example, a UE (User Equipment) is a wearable device that collects data related to health-related characteristics (such as heart rate, body temperature, etc.) of a specific person (i.e., a specific sample); while a medical application on the user's device may have medical records of the same specific person (i.e., the same specific sample), also bringing data related to other characteristics of the same person (i.e., the same sample) (such as medical experiments, medication, etc.). For example, the wearable device and the medical application can jointly train an AI / ML model to predict the likelihood of that specific person experiencing a heart attack at a certain time.
[0069] In some applications of VFL, entities (e.g., parties or participants) can interact instantaneously within a shared pool of samples, enabling them to collaborate in training AI / ML models to optimize desired outputs. A challenge may be selecting all collaborating entities within the shared pool that possess different features of the same specific sample, and performing that sample alignment across all samples within the pool, without compromising the privacy of the collaborating entities (e.g., entities participating in VFL, also known as VFL members).
[0070] For example, in V2X (Vehicle-to-Everything) applications, numerous vehicles and pedestrians are able to traverse certain routes within a city during their daily commutes. Vehicles and pedestrians may possess devices capable of collecting data, such as user devices and / or one or more clients deployed on the respective user devices. Conversely, installed cameras can capture images of traffic conditions. In this scenario, the installed cameras can be considered user devices and / or one or more clients deployed on the respective user devices. The cameras and devices installed on vehicles and / or pedestrians can capture data about useful and / or unique, informative features that can be used for VFL training (e.g., for training AI / ML models using VFLs). For instance, the VFL can be used to train AI / ML models that can provide improved navigation services or traffic condition predictions for roads and highways in smart cities.
[0071] In this example, the shared sample pool is the collection of all roads, streets, etc., among all pedestrians and vehicles. The challenge of sample alignment is to identify which vehicles and / or pedestrians (i.e., UEs and / or one or more clients deployed on the UEs) traversed the same specific roads without exposing their locations and data (i.e., data associated with one or more features). In some cases, location can be considered a feature. In smart city applications, data on features from various sources such as vehicles, pedestrians, and traffic cameras can be aggregated to create a comprehensive dataset. This dataset can then be used to train an AI / ML model that predicts traffic conditions and provides navigation recommendations. Aggregating data associated with unique features from individual participants (i.e., UE clients) belonging to the same specific road (the same sample) in the VFL allows for joint training of AI / ML models, potentially leading to better traffic condition prediction and navigation services.
[0072] Another example application of VFL is augmented reality services (such as AR interior design) widely used by smartphones. AR interior design services will be described as illustrative AR services, but VFL can also be used for many other AR services.
[0073] For AR interior design services, furniture retailers can offer AR interior design options that include their products. These AR design services use generative artificial intelligence (GenAI) models to provide different interior design options for their products using data obtained from different customer cameras in their homes. However, a limitation of these AR interior design services is that customer data privacy may not be maintained. In other words, customers may not want to expose their data to train the GenAI model. VFL can be used to enable customer cameras to participate in training a generative model that provides interior design options without compromising customer data privacy.
[0074] Therefore, in this example, the shared sample pool among all participants consists of all products (i.e., furniture) supplied by the furniture retailer, and users (i.e., UE clients) who possess data related to the characteristics of their homes (i.e., interior photos) with specific samples (specific furniture types, such as specific tables, desks, sofas, etc.). The challenge of sample alignment lies in how to select interior photos of different customers who own the same specific furniture products without exposing their interior photos and infringing on their privacy.
[0075] Other AR services may use AR glasses and / or head-mounted devices to deliver generative AI services.
[0076] Another example application of VFL might involve passengers using public transportation services within a city. In this example, the shared sample pool among participants (UEs) is all public transportation services (e.g., buses, subways, etc.). Passengers carry devices (e.g., UEs and / or one or more clients deployed on the UEs) that enable them to participate in VFL tasks. In this task, UE clients can co-train models using unique features collected by devices associated with the same samples (e.g., the same buses, the same subway lines, etc.). The trained model can provide, for example, delay prediction, public transportation route recommendation, event prediction, etc. In this scenario, the challenge of sample alignment lies in selecting participants (e.g., UEs and / or one or more clients deployed on the UEs) that have features associated with the same public transportation samples without compromising the privacy of the UEs.
[0077] See Figure 1 , Figure 1 This diagram illustrates a UE configured to communicate with the AIMLE server and VAL server during VFL. VFL is a federated learning method in which a central server (e.g., ...) is configured to communicate with the AIMLE server and VAL server during VFL. Figure 1 The VAL server (shown) receives the outputs of AI / ML models trained on individual entities participating in VFL (e.g., VFL client devices) and facilitates the secure aggregation of the outputs of AI / ML models trained on the entities participating in VFL. The AI / ML model on each entity is updated after the training loss of the aggregated output is computed by one or more clients.
[0078] exist Figure 1 In the example, the central server is referred to as AI / ML Enabled AIMLE Server 5. This central server may sometimes also be referred to as a VFL server. In this document, the terms AIMLE Server and VFL Server are used interchangeably. An AIMLE Server provides a service that facilitates the execution or enabling of one or more AI / ML operations on the server side.
[0079] The VFL client device can be provided by user equipment 2. In this document, the terms "VFL client device" and "user equipment" are used interchangeably. Figure 1 In the diagram, a UE 2 is shown. In an embodiment, there will actually be multiple UEs. The VFL client device can be provided by an AIMLE client 4 (sometimes referred to as a VFL client) deployed on UE 2. The AIMLE client provides a service that facilitates the execution or enabling of one or more AI / ML operations on the client (i.e., UE) side.
[0080] AIMLE clients can be edge AI / ML applications. Edge AI / ML applications can consume 3GPP data analytics services.
[0081] Vertical Application Layer (VAL) server 6 provides server application functionality for a specific VAL service. VAL client 3 provides client functionality for a specific VAL service. VAL client 3 is deployed on the UE. In some embodiments, the AIMLE client and VAL client may be provided by a single client that provides the functionality of both the AIMLE client and the VAL client. In other embodiments, the AIMLE client and VAL client are separate clients.
[0082] VAL allows AI / ML services to be exposed to third-party VAL services provided by a VAL server, where the VAL service is provided by a third party outside the 3GPP network serving communication with the UE. The VAL server provides VFL tasks to consumers of the VAL service (i.e., VAL clients). The VAL server requests assistance from the AIMLE server for its VFL tasks. In some embodiments, the VAL server may be provided within the 3GPP network.
[0083] VAL server 6 communicates with VAL client 2 via 3GPP network 8, where VAL server 6 is provided by a third party.
[0084] The AIMLE server 5 can be provided by a third party. The AIMLE server communicates with the AIMLE client via 3GPP network 8, where the AIMLE server 5 is provided by a third party. Alternatively, the AIMLE server can also be provided within the 3GPP network.
[0085] A feature is a single measurable attribute or characteristic of the data collected and processed by the AIMLE client. These features can be viewed as data points or variables within the dataset.
[0086] A dataset is a collection of records that may include data associated with multiple different features. For example, a feature could be temperature, and the data could be temperature values.
[0087] A data domain refers to a specific area or category of data related to a VFL task. Each data domain contains a different set of features or data types that are collected and processed by the AIMLE client. For example, in a VFL application, a UE's dataset may belong to more than one different data domain. In other words, datasets provided by different UEs may contain data associated with different features relevant to the VFL task.
[0088] A sample refers to a specific instance or entity within the data used in VFL. Each sample can be associated with various features collected by different AIMLE clients and / or the UE on which the corresponding AIMLE client is deployed. The goal of VFL is to align these samples across different data domains without compromising privacy.
[0089] A sample pool (sometimes also called a sample space) is the collection of all possible samples that can be used in a VFL task. For example, in traffic prediction applications, a sample pool might include all roads, routes, and streets in a city.
[0090] Sample alignment refers to the process of selecting UEs with different features associated with the same samples to participate in the VFL task. This ensures that data about different features from different UEs can be used together to train machine learning models.
[0091] In some embodiments, the sample space or sample pool is unknown, and the sample ID used for the sample is also unknown.
[0092] In order to train VFL among the possible candidates, a minimum number of data samples and / or one or more portions of one or more datasets aligned among the potential participants are required. This is the challenge of sample alignment.
[0093] Potential VFL participants (UEs) may identify their data samples differently from each other. In this scenario, given that data sample identification may be private to the respective UE, it may be impossible to label the datasets of the corresponding UEs. For example, one UE might use timestamps to identify data points, while another UE might use location-based labels to identify the same data points. In this example, each client has a different method to identify its data samples. Therefore, the challenge for the VFL server is to identify UEs with aligned data samples. This can cause problems when determining the feasibility of VFL training.
[0094] Some implementations can address this sample alignment challenge.
[0095] Different UEs can assign (multiple) different feature IDs to their data samples based on their respective capabilities. Feature IDs can be identifiers for specific features, which are collected by the UE, for example.
[0096] For example, different UEs equipped with different sensors (such as cameras, accelerometers, and / or GPS (Global Positioning System)) and / or different network capabilities (such as connection type and / or mobility profiles) may differ in how effectively the respective UE can extract and construct feature sets from its local data. In this case, checking the feasibility of VFL training can be challenging. Some embodiments can address this challenge. Some embodiments can provide a minimum threshold for sample alignment confidence between different UE datasets.
[0097] In cases where sample alignment is ambiguous and the VAL server requests the AIMLE server to perform VFL across different UE clients, some embodiments can be used. For example, sample alignment ambiguity may be caused by data samples from different UEs being identified in different ways. In some embodiments, a UE's eligibility as a VFL candidate can be evaluated by determining whether the corresponding values for different UEs in the shared feature space match.
[0098] In some embodiments, UE candidates for VFL can provide the UE’s feature set capabilities, for example, during the registration process.
[0099] In some embodiments, the candidate list of UEs is provided by a VAL server. The feature set capability of the UE may be unknown.
[0100] In some embodiments, the AIMLE server can evaluate sample alignment among potential VFL participants.
[0101] In some embodiments, the VAL server (consumer) can determine the reliability of model training based on its initial request.
[0102] In some embodiments, sample feature negotiation may be provided.
[0103] In some embodiments, alignment of features from different datasets from different UEs can be provided.
[0104] In some embodiments, alignment of samples taking into account the values of corresponding features can be provided.
[0105] In some embodiments, a minimum level of confidence threshold required for the VFL process can be provided.
[0106] See Figure 3 The figure is used to illustrate a first example process of some embodiments.
[0107] In this example, VFL members (such as AIMLE clients deployed on the corresponding UE) have registered their VFL profiles with the AIMLE server. VFL members can update the status or information in their VFL profiles with the AIMLE server. A VFL member is a VFL client that can participate in a specific VFL task.
[0108] In this example, the datasets of different UEs within a UE may belong to more than one different data domain. A UE is a UE deployed on an AIMLE client (VFL member). For example, a UE can be associated with one data domain, while different UEs can be associated with different data domains.
[0109] In this example, the VAL server has successfully subscribed to and / or registered with the AIMLE server for model training notifications.
[0110] As shown by reference numeral 0 in the attached figure, the AIMLE client registers with the AIMLE server. The AIMLE client may provide a registration message to the AIMLE server. The registration message may include a list of feature IDs. These feature IDs identify features or feature data. A feature or feature data represents an attribute or characteristic of one or more data points in sample data collected or collectable by or on which the AIME client and / or AIMLE client are deployed. The registration message may include the status of the corresponding feature identified by the feature ID. The feature may be available (or activated) or unavailable (or deactivated). The registration message may include information regarding the expiration of the corresponding status of the corresponding feature. This may indicate the time when the corresponding status of the corresponding feature may change.
[0111] As shown in Figure 1, the AIMLE server receives ML model training requests from the VAL server.
[0112] If an ML model training request is received and the AIMLE server determines that VFL training will be used, then part of the process in Figure 2 can proceed. VFL training is performed across different data domains.
[0113] Optionally, as shown in Figure 2, the AIMLE server can use the ML model retrieval process to retrieve ML models for ML training tasks. If the retrieved model has been trained and meets the machine learning requirements requested by the VAL server, the next part of the process is marked with 7. Otherwise, the next part of the process is marked with 3.
[0114] Optionally, as shown in Figure 3, the AIMLE server can continuously monitor and select AIMLE clients for VFL training. When a candidate AIMLE client list marked with 1 is selected as part of the process, the AIMLE server selects the provided AIMLE client for VFL training.
[0115] As shown in Figure 4, the AIMLE server sends a client feature selection request to multiple AIMLE clients. This client feature selection request is a request for the corresponding AIMLE client to activate (or make available) and / or deactivate (or disable) one or more features associated with that AIMLE client. For example, the client feature selection request may include one or more features, and for each of the one or more features, an indication is provided as to whether that feature will be activated (or made available) or deactivated (or disabled). The one or more features may be identified by corresponding feature identifiers.
[0116] The client feature selection request may include information indicating the priority of which features will be activated (or made available) or deactivated (or made unavailable) for VFL training.
[0117] The client feature selection request may include information indicating the time at which the requested activation / deactivation of the corresponding feature will be enforced. In other words, if a feature is activated or made available, the time information will indicate the minimum time for which the feature will be activated or made available. If a feature is deactivated or made unavailable, the time information will indicate the minimum time for which the feature will be deactivated or made unavailable.
[0118] The client feature selection request may include information indicating the identifier of the AIMLE server.
[0119] In some embodiments, the client feature selection request may include one or more of the information described in the table below. In the following examples, some of the information indicated in the table is indicated as "M"—meaning mandatory. However, in one or more other embodiments, one or more of the information in the table below may be optional. In the following examples, some of the information indicated in the table is indicated as "O"—meaning optional. However, in one or more other embodiments, one or more of the information in the table below may be mandatory.
[0120] As shown in Figure 5, the AIMLE client sends one or more client feature selection responses to the AIMLE server.
[0121] The client feature selection response may include information indicating the status of one or more features that the AIMLE server requests to be activated / deactivated.
[0122] The client feature selection response may include information indicating whether the feature activation / deactivation requested by the AIMLE server was successful or unsuccessful.
[0123] If a feature activation / deactivation request from the AIMLE server fails, the client feature selection response may include information indicating the reason for the failure.
[0124] The client-side feature selection response may include information indicating the percentage of available features that are prioritized for VFL training.
[0125] When the time provided by the AIMLE server is unavailable to the UE, the client function selection response may include information related to the time adjustment that will be enforced by the UE for the availability (or unavailability) of the feature. The UE may provide a reason or justification for the time adjustment.
[0126] In some embodiments, the client feature selection response may include one or more of the information described in the table below. In the following examples, some of the information indicated in the table is indicated as "M" – i.e., mandatory. However, in one or more other embodiments, one or more of the information in the table below may be optional. In the following examples, some of the information indicated in the table is marked as "O" – i.e., optional. However, in one or more other embodiments, one or more of the information in the table below may be mandatory.
[0127] Based on the response message, the AIMLE server finds one or more common IDs by comparing the client's feature descriptions and / or feature IDs. The AIMLE server considers feature IDs shared by different candidate datasets and constructs pairwise relationships between candidate datasets for use in VFL training.
[0128] Feature descriptions can be considered as information describing the nature and / or characteristics of features collected by the UE. Feature information may include one or more of the following: the type of feature, the purpose of the feature, and / or any other information.
[0129] If one or more feature IDs are a subset or part of the feature space union, the AIMLE server receives one or more hashed feature values for each data sample from each UE. The feature space union is the feature space that may be used for VFL training.
[0130] As shown in Figure 6, the AIMLE server possesses information about VFL members. VFL member information can be included in a request as part of the process described in Figures 1 and / or 3. The AIMLE server interacts with the VFL members (AIMLE clients deployed on the UE) for each data domain for ML model training capability evaluation.
[0131] As shown in Figure 7, the AIMLE server determines the VFL members for this VFL training process based on information received in a portion of the process in Figure 5 and one or more parameters received in a portion of the process in Figure 1. The AIMLE server can determine the VFL members for each data domain.
[0132] The criteria used by the AIMLE server to determine VFL members may include one or more of the following: Available data and minimum number of data samples for the same samples used in VFL members (AIMLE clients); Features of samples / datasets and data labels within VFL members (AIMLE clients); VFL members (AIMLE clients) support the available time for VFL training operations; The ability and minimum number of VFL members (AIMLE clients) to perform VFL training operations; and / or AIML model information for VFL members (AIMLE clients) and for AIMLE servers.
[0133] As shown in Figure 8, the AIMLE server coordinates the selected VFL members (AIMLE clients) for VFL training. During VFL training, VFL members (AIMLE clients) send intermediate results to the AIMLE server, and the AIMLE server responds to the VFL members with updated information (e.g., gradients) for the ML model being trained. Information from the AIMLE server can be used to update the model parameters maintained at each VFL member (AIMLE client) for different data domains.
[0134] As shown in Figure 8a, the AIMLE server can report the training status to the VAL server. This can include intermediate training results. The VAL server can adjust its requests for training the ML model. If the VAL server is providing training data labels to complete the process, it can send updated training parameters to the AIMLE server for distribution to VFL members. The updated training parameters apply to the model of the VFL member (AIMLE client) associated with each data domain.
[0135] As shown in Figure 9, the AIMLE server sends an ML model training notification to the VAL server. If the training schedule is not complete (e.g., there are one or more rounds of training remaining), the AIMLE server can configure the next training schedule set, and parts of processes 3 through 8 can be repeated for the next round of training.
[0136] Such as about Figure 3 As discussed, these embodiments can address the problem of sample alignment where different participants identify the same data points in different ways.
[0137] See Figure 4 The diagram is used to illustrate a second example process of some embodiments. This process can be used for sample alignment.
[0138] This example procedure involves a UE with an AIMLE client and a VAL client, a VAL server (consumer), and a VFL server (AIMLE server).
[0139] As shown by reference 0 in the attached diagram, the AIMLE client deployed on the UE registers its VFL profile with the AIMLE server. This indicates that the AIMLE client and / or the UE on which the AIMLE client is deployed are willing to participate in the VFL process. The AIMLE client sends a registration request to the AIMLE server.
[0140] The registration request may include information about the AIMLE client and / or the UE on which the AIMLE client is deployed.
[0141] The registration request may include information about the AIMLE client and / or the identifier of the UE on which the AIMLE client is deployed.
[0142] The registration request may include information about one or more supported VAL services and / or AIML client profiles. For each client profile provided in the list, a corresponding list of identifiers for one or more VAL services may be provided. One or more permissions corresponding to one or more of the one or more VAL services may be provided.
[0143] The registration request may include information about one or more VAL service IDs that the AIMLE client supports.
[0144] The registration request may include information about one or more service permission levels associated with one or more services. These one or more services may be one or more VAL services. Permission levels may be associated with one or more of, for example, advanced resource usage, standard resource usage, and / or limited resource usage.
[0145] The registration request may include information about the AIMLE client's ability to support AI / ML operations for one or more of the identified VAL services.
[0146] The registration request may include information about one or more features supported by the AIMLE client and / or the UE on which the AIMLE client is deployed.
[0147] Feature data or features can be viewed as individual measurable attributes or characteristics of a data sample collected by the UE. Each feature represents a specific aspect of the data that can be used for analysis and model training. Feature IDs can be used to identify different feature data or features that can be provided by the UE.
[0148] The registration request may include information about the status of one or more features, each identified by a corresponding identifier ID. This indicates whether the feature is currently available or unavailable. For example, the feature may be activated or deactivated. Feature availability may change. Feature availability may depend on the UE's current state. For example, if the UE has low battery, it may deactivate one or more features. Of course, different criteria can be used by the UE to determine whether a particular feature is available.
[0149] A registration request may include information about the availability or status of one or more features. For example, a registration request may include information indicating when the corresponding feature will no longer be available.
[0150] The registration request may include information about the location of the AIMLE client / UE on which the AIMLE client is deployed.
[0151] The registration request may include information about the dataset task description. The dataset task description may include information about the task objectives (e.g., indicating the possibility of performing one or more different learning operations, such as clustering, classification, prediction, etc.) and / or the data domain (e.g., whether the data type is image, text, etc.) of the local dataset (e.g., UE dataset).
[0152] In some embodiments, the registration request may include one or more of the information described in the table below. In the following examples, some of the information indicated in the table is indicated as "M" – i.e., mandatory. However, in one or more other embodiments, one or more of the information in the table below may be optional. In the following examples, some of the information indicated in the table is indicated as "O" – i.e., optional. However, in one or more other embodiments, one or more of the information in the table below may be mandatory.
[0153] As shown in Figure 1, the VAL server (consumer) sends a registration request to the AIMLE server, providing its profile.
[0154] As shown in Figure 2, the VAL server sends an ML model training request to the AIMLE server for VFL training in different data domains.
[0155] As shown in Figure 3, the AIMLE server determines the VFL task to be performed. The AIMLE server can authorize ML model training requests from the VAL server. The AIMLE server can use the ML model retrieval process to retrieve ML models for ML training tasks.
[0156] As shown in Figure 4, the AIMLE server sends a client feature selection request to the AIMLE client. A corresponding client feature selection request can be sent to multiple AIMLE clients. A client feature selection request is a request from the corresponding AIMLE client to activate (or make available) and / or deactivate (or disable) one or more features associated with that AIMLE client. For example, a client feature selection request may include one or more features, and for each of the one or more features, an indication is provided as to whether that feature will be activated (or made available) or deactivated (or disabled). The one or more features can be identified by corresponding feature identifiers.
[0157] Considering the current state of the features of the specific UE and / or the AIMLE client deployed on that UE, the client feature selection request can be associated with the desired features that will be activated or deactivated in the specific UE and / or the AIMLE client deployed on that UE.
[0158] However, in some embodiments, the AIMLE server broadcasts information about the feature set and the status of its requests to multiple different UEs. Each UE and / or AIMLE client deployed on a UE will perform feature-based activation / deactivation within the feature set based on whether the UE and / or the AIMLE client deployed on the UE can provide the corresponding feature and the current status of that feature. The client feature selection request may include information indicating the priority of the corresponding features to be activated (or made available) or deactivated (or made unavailable) for VFL training.
[0159] A client feature selection request may include information indicating the time at which the requested activation / deactivation of the corresponding feature will be enforced. In other words, if a feature is to be activated or made available, the time information will indicate the minimum time for which the feature will be activated or made available. If a feature is to be deactivated or made unavailable, the time information will indicate the minimum time for which the feature will be deactivated or made unavailable.
[0160] The client feature selection request may include information indicating the identifier of the AIMLE server.
[0161] Therefore, the AIMLE server can consider the feature capability availability status of the AIMLE client and / or the UE on which the AIMLE client is deployed, and decide to negotiate with one or more AIMLE clients. This could be to activate one or more features that can be used for VFL within a time period.
[0162] The AIMLE server can request the deactivation (or unavailability) of one or more feature capabilities for the dataset used by the UE. For example, features unrelated to the VFL model training objective task. To avoid additional complexity or reduce the heterogeneity between datasets from different clients, the AIMLE server can request that one or more features of the UE be deactivated or made unavailable.
[0163] In some embodiments, a client feature selection request may include one or more of the information described in the table below. In the examples below, some of the information indicated in the table is indicated as "M", meaning mandatory. However, in one or more other embodiments, one or more of the information in the table below may be optional. In the examples below, some of the information indicated in the table is indicated as "O", meaning optional. However, in one or more other embodiments, one or more of the information in the table below may be mandatory.
[0164] As shown in Figure 5, the AIMLE client sends one or more client feature selection responses to the AIMLE server.
[0165] The client feature selection response may include information indicating the status of one or more features that the AIMLE server requests to be activated / deactivated.
[0166] The client feature selection response may include information indicating whether the feature activation / deactivation requested by the AIMLE server was successful or unsuccessful.
[0167] If a feature activation / deactivation request from the AIMLE server fails, the client feature selection response may include information indicating the reason for the failure.
[0168] The client feature selection response may include information indicating the percentage of features that meet the criteria. This information may indicate the percentage of available features that are prioritized for VFL training. If the request from the AIMLE server provides a priority level indicator for the requested features, the AIMLE client may provide information in the response indicating the weighted (e.g., based on the priority of the features) percentage of the requested features(s) to be activated or deactivated. The AIMLE server may consider this information to filter potential UE(s) participants(s).
[0169] When the time provided by the AIMLE server is not supported by the UE, the client feature selection response may include information about the availability (or unavailability) of the feature by the UE and the time adjustment that will be enforced. The UE may provide a reason or justification for the time adjustment.
[0170] In some embodiments, the client feature selection response may include one or more of the information described in the table below. In the examples below, some of the information indicated in the table is indicated as "M" – i.e., mandatory. However, in one or more other embodiments, one or more of the information in the table below may be optional. In the examples below, some of the information indicated in the table is indicated as "O" – i.e., optional. However, in one or more other embodiments, one or more of the information in the table below may be mandatory.
[0171]
[0172] As shown in Figure 6, the AIMLE server searches for one or more common features provided by different UEs. AIMLE can compare features from different UEs to find one or more common features. AIMLE can compare feature IDs and / or feature descriptions to determine one or more common features.
[0173] The AIMLE server can construct pairwise relationships between two different UEs based on their respective UE datasets, considering shared feature IDs and / or feature descriptions. Relationships between different client datasets are analyzed, allowing the server to identify which clients might share similar samples based on common features. The AIMLE server captures pairwise relationships between each client's dataset and other clients' datasets.
[0174] The shared feature ID and / or feature description can be used by the AIMLE server to determine one or more AIMLE clients and / or UEs on which AIMLE clients are deployed as candidates for VFL tasks.
[0175] As shown in Figure 7, the AIMLE server sends a request to one or more AIMLE clients of the corresponding UE for VFL candidate. Separate requests may be sent to different AIMLE clients.
[0176] This request can be used by the corresponding AIMLE client to send the hash values of one or more features of one or more data samples of the UE on which the AIMLE client is deployed, wherein the feature identifiers or feature descriptions of one or more features are in the feature space and set for the feature identifiers or descriptions to be used in the VFL task.
[0177] The request sent by the AIMLE server may include information about the identifier of the AIMLE client that will send the hash value.
[0178] Requests sent by the AIMLE server may include information about the AIMLE server's identity.
[0179] The request sent by the AIMLE server may include information about the feature ID (or feature description) that will be sent for its hash value.
[0180] In some embodiments, a request sent by the AIMLE server may include one or more of the information described in the table below. In the following examples, the information indicated in the table is indicated as “M” – i.e., mandatory. However, in one or more other embodiments, one or more of the information in the table below may be optional.
[0181] As shown in Figure 8, the AIMLE client sends a response to the AIMLE server. This response may include one or more hash values when the feature identifier or feature description is located within the feature space union of the feature identifier or description to be used in the VFL task. In other words, the response may include one or more hash values of the feature specified in the AIMLE request.
[0182] The AIMLE client will determine the hash value of one or more features used to identify the user in the AIMLE request for one or more data samples of the UE deployed on it. The hash value of the features can be calculated by running a cryptographic hash function on the data.
[0183] The response may include information about the identifier of the AIMLE client that provided one or more hash values.
[0184] In some embodiments, the response sent by the AIMLE client may include one or more of the information described in the table below. In the following examples, the information indicated in the table is marked "M" – meaning mandatory. However, in one or more other embodiments, one or more of the information in the table below may be optional.
[0185]
[0186] As shown in Figure 9, the AIMLE server uses a common feature hash value for each candidate pairwise relationship with other UE datasets to filter whether one or more AIMLE clients are candidates for the VFL task. This filtering list can consider the level of alignment between the samples of a corresponding AIMLE client and the samples of one or more other AIMLE clients. The AIMLE server can provide a confidence matrix representing the alignment of each dataset pair.
[0187] Consider the following example. A first client has features A and B, and a second client has features B and C to describe its data sample. Both clients share a common feature B, so the hash values of feature B for their data samples can be compared. If the hash feature values (the hash values used for feature B) match exactly, they correspond to the same data sample with a certain probability (confidence value).
[0188] In the presence of multiple potential clients, a confidence matrix can be constructed, where each element of the matrix represents a pairwise confidence value for two potential clients with respect to the sample alignment of their datasets.
[0189] Confidence values can be determined, for example, using methods such as Private Set Intersection (PSI) or Probabilistic Record Link (PRL). The resulting confidence matrix can then be constructed. The confidence matrix can be constructed using methods such as Siamese networks, cosine similarity, or Pearson correlation between client pairs.
[0190] Confidence scores can be determined using the similarity between hash feature values from different UEs. This similarity can be determined through one or more methods or approaches, such as:
[0191] Private set intersection (PSI) is the overlap between sets that measure hash feature values.
[0192] Probabilistic record links (PRLs) estimate the probability that records from different datasets refer to the same entity.
[0193] Siamese networks are neural networks that learn to compare pairs of data points and output similarity scores.
[0194] Cosine similarity measures the cosine of the angle between vectors of two hash feature values; and / or
[0195] Pearson correlation measures the linear correlation between two sets of hash feature values.
[0196] Confidence scores can be represented as numerical ratings, ranging from 0 to 1, where:
[0197] 0: Indicates misalignment between datasets.
[0198] 1: Indicates perfect alignment between datasets.
[0199] As shown in Figure 10, the AIMLE server sends information to the VAL server regarding the confidence value of each UE participant in the VFL task. The AIMLE server may also send information about candidate pairwise confidence matrices. The portion of the process marked by 10 can be optional. If the VAL server instructs the VAL server to require a minimum confidence threshold for UE participation in VFL training, then the portion of the process marked by 10 can be executed.
[0200] Optionally, the message sent by the AIMLE server to the VAL server includes information about one or more criteria for analyzing related client pairs based on specific VFL characteristics. This information may be referred to as one or more discovery filters. This information may indicate the number of common features for each client pair; therefore, in this case, it indicates the confidence values used to determine client A and client B when client A and client B share n feature IDs. This information may indicate the minimum confidence value recommended by the AIMLE server for performing VFL tasks with specific requirements. This information may also indicate data domain filters for the clients to provide more information to the VAL server.
[0201] In some embodiments, the message may include one or more of the information described in the table below. In the following examples, each of the information indicated in the table is indicated as “M” – i.e., mandatory. However, in one or more other embodiments, one or more of the information in the table below may be optional. In the following examples, each of the information indicated in the table is indicated as “O” – i.e., optional. However, in one or more other embodiments, one or more of the information in the table below may be mandatory.
[0202]
[0203] As shown in Figure 11, the VAL server (consumer) analyzes potential participants for the VFL task and determines a minimum confidence threshold for participants to participate in the VFL task. This minimum confidence threshold can be determined based on the characteristics of the model training request and the pairwise confidence factors of potential participants. The part of the process marked by 11 can be optional. If the VAL server instructs that it requires a minimum confidence threshold for UE participation in VFL training, then the part of the process marked by 11 can be executed.
[0204] As shown in Figure 12, the VAL server (consumer) sends a minimum confidence threshold to the AIMLE server. Optionally, the VAL server may provide the AIMLE server with one or more sets of criteria to analyze the feasibility of performing VFL on behalf of the requester based on confidence factors. This part of the process, marked by 12, can be optional. This part of the process can be executed when the VAL server is instructed to use the minimum confidence threshold for VFL training.
[0205] In some embodiments, the message sent by the VAL server may include one or more of the information described in the table below. In the examples below, each of the information indicated in the table is indicated as "M" – i.e., mandatory. However, in one or more other embodiments, one or more of the information in the table below may be optional. In the examples below, each of the information indicated in the table is indicated as "O" – i.e., optional. However, in one or more other embodiments, one or more of the information in the table below may be mandatory.
[0206]
[0207] As shown in Figure 13, the AIMLE server supervised evaluation process involves testing a given AIMLE client and / or a UE on which AIMLE is deployed and / or a VAL client on which AIMLE is deployed to verify its model training capability.
[0208] As shown in Figure 14, the AIMLE server selects one or more qualified candidates for VFL training that meet the required level of data sample alignment and have the capability required to participate in the VFL process.
[0209] See Figure 5 The figure illustrates a third example of the process in some embodiments.
[0210] As shown in Figure 1, the VAL server (consumer) sends a registration request to the AIMLE server, providing its profile.
[0211] As shown in Figure 2, the VAL server sends an ML model training request to the AIMLE server for VFL training in different data domains.
[0212] As shown in Figure 3, the AIMLE server determines the VFL task to be performed. The AIMLE server can authorize ML model training requests from the VAL server. The AIMLE server can use the ML model retrieval process to retrieve ML models for ML training tasks.
[0213] As shown in Figure 4, in order to discover candidate data sample alignments for the VFL task, the AIMLE server sends a request to the AIMLE client deployed on the UE (potential candidate). The request can be sent to different potential candidates. The request can be a request to the AIMLE client sending feature descriptions and / or (multiple) feature IDs. The request can also be a request to the AIMLE client sending (multiple) hashed feature IDs. Considering the unique capabilities of each UE, this may apply to one or more private feature sets of (multiple) UEs. The requested information is to allow the AIMLE server to find (multiple) public feature IDs in the UE's dataset.
[0214] The list of (multiple) client UEs is provided by the VAL server. However, the capabilities of these UEs are unknown to the AIML server. The AIML server sends requests to find (multiple) public feature IDs. However, depending on the VFL task, the client may be unwilling to share its (multiple) feature IDs. This may be related to its privacy concerns. The client may not disclose its ability to provide certain features. To resolve this privacy issue, the client may send a hashed version of its (multiple) feature IDs.
[0215] For example, consider a scenario where some UEs have humidity and temperature sensors on their devices, and these UEs are able to provide characteristics representing the humidity and temperature at the time of data recording. If these capabilities are proprietary, (multiple) UEs can send (multiple) hashed feature IDs instead of (multiple) feature IDs themselves. The server can then compare these client datasets based on (multiple) hashed feature IDs.
[0216] The request may include information about the identifier of the AIMLE server.
[0217] This request may include information about the ML task. This information may indicate the VFL task and / or the ML model. This is to allow the AIMLE client and / or the UE on which the AIMLE client is deployed to prepare and build datasets related to the ML task.
[0218] The request may include information about the feature set management type requirement. This information can indicate what feature management process is being requested. For example, the feature management process could be feature preparation or feature importance score analysis.
[0219] This request may include information regarding the required feature type information. This information may include details about the data type of the feature (e.g., numeric, category, text, image).
[0220] The request may include information regarding feature dimension requirements. This information may include desired details about the shape or number of dimensions in the feature representation.
[0221] The request may include information regarding feature privacy constraints. This information may include requirements regarding privacy rules or constraints that affect the feature-sharing process.
[0222] In some embodiments, the request may include one or more of the information described in the table below. In the following examples, some of the information indicated in the table is marked "M" – i.e., mandatory. However, in one or more other embodiments, one or more of the information in the table below may be optional. In the following examples, some of the information indicated in the table is marked "O" – i.e., optional. However, in one or more other embodiments, one or more of the information in the table below may be mandatory.
[0223] As shown in Figure 5, the AIMLE client sends a response to the AIMLE server. The response sent by the AIMLE client may include information about the feature description or feature ID. The feature ID may be sent as a hash value.
[0224] The response may include information about the feature set management type output. This information may include the status of the feature management operation.
[0225] The response may include information about the feature type output. This information may include details about the data type of the feature (e.g., numeric, categorical, text, image).
[0226] The response may include information about the feature dimension output. This information may include information about the shape or number of dimensions in the feature representation.
[0227] The response may include information about the feature privacy constraint output. This information may include one or more privacy rules or constraints that affect the feature sharing process.
[0228] The response may include information about whether the request was successful.
[0229] The response can include information about whether the request was unsuccessful. If the request was unsuccessful, the response can indicate the reason for the failure.
[0230] In some embodiments, messages sent by the VAL server may include one or more of the information described in the table below. In the examples below, some of the information indicated in the table is indicated as "M" – i.e., mandatory. However, in one or more other embodiments, one or more of the information in the table below may be optional. In the examples below, some of the information indicated in the table is indicated as "O" – i.e., optional. However, in one or more other embodiments, one or more of the information in the table below may be mandatory.
[0231] Part of the process, as indicated by reference numerals 6 to 14, is usually as follows: Figure 4 The process shown in the second example is described as part of the process in the accompanying drawings 6 to 14.
[0232] See Figures 6 to 9 It illustrates some methods of some embodiments.
[0233] Each method can be executed by a device. The device may include suitable components, such as a circuit system for providing the corresponding method. Alternatively or additionally, the device may include at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the device to at least provide the corresponding method. Alternatively or additionally, the device may be, for example, regarding... Figure 2 The methods discussed herein can be provided by computer program code or computer executable instructions.
[0234] Figure 6 and Figure 8 The method can be performed by a device. The device can be a user equipment, or it can be provided in a user equipment.
[0235] Figure 7 and Figure 9 The method can be performed by a device. This device can be a server, or it can be provided in a server.
[0236] See Figure 6 .
[0237] The method includes, as shown by reference numeral A1, providing a vertical federated learning (VFL) client profile for an AI machine learning enabling client for a user device, the VFL client profile including information about one or more features of a data sample of the user device, wherein the user device is a candidate user device for participating in a vertical federated learning machine learning task.
[0238] The method includes, as shown by reference numeral A2 in the figure, receiving from a first server a request for changing the availability of one or more features of one or more features of a data sample of a user equipment, the request including identifiers of one or more features of the data sample of the user equipment.
[0239] The method includes, as shown by reference numeral A3 in the figure, changing the availability of one or more features according to the request.
[0240] The method includes, as shown by reference numeral A4 in the attached figure, sending a response to a first server in response to a request, the response indicating that the availability of one or more of the features has been changed.
[0241] See Figure 7 .
[0242] The method includes, as shown by reference numeral B1, receiving a vertical federated learning (VFL) client profile for an AI machine learning enabled client of a corresponding user device, the VFL client profile including information about one or more features of a data sample of the corresponding user device.
[0243] The method includes, as shown by reference numeral B2, sending a request to the corresponding user equipment to change the availability of one or more features of one or more features of the corresponding user equipment's data sample, the request including an identifier of one or more features of the corresponding user equipment's data sample.
[0244] The method includes, as shown by reference numeral B3 in the figure, receiving a response to a request from the corresponding user equipment, the response indicating that the state of one or more features has been changed.
[0245] See Figure 8 .
[0246] The method includes, as indicated by reference numeral C1, receiving a first request for information regarding one or more features of a data sample of a user equipment.
[0247] The method includes, as shown by reference numeral C2 in the figure, sending a response to a first request to a first server, the response including information about one or more features of a data sample of a user device, wherein the user device is a candidate user device for participating in a vertical federated learning task.
[0248] The method includes, as indicated by reference numeral C3, receiving a second request from a first server for one or more integrity protection values of one or more features of one or more features of a data sample for a user equipment.
[0249] The method includes, as shown by reference numeral C4 in the figure, providing a first server, in response to a first request, one or more integrity protection values of one or more features of one or more features of a data sample for a user equipment.
[0250] See Figure 9 .
[0251] The method includes, as shown by reference numeral D1 in the figure, sending a first request to the corresponding user equipment for information about one or more features of a data sample of the corresponding user equipment.
[0252] The method includes, as shown by reference numeral D2 in the figure, receiving a response to a first request from a corresponding user equipment, the response including information about one or more characteristics of a data sample of the corresponding user equipment.
[0253] The method includes, as shown by reference numeral D3 in the figure, sending a second request to the corresponding user equipment, the second request being for one or more integrity protection values of one or more features of one or more features of a data sample for the corresponding user equipment.
[0254] The method includes, as shown by reference numeral D4 in the figure, receiving, in response to a second request, one or more integrity protection values of one or more features of one or more features of a data sample for a user equipment.
[0255] Some examples of this disclosure can be implemented using communication networks such as 5G (also known as NR), 6G, or any future communication network.
[0256] The term "User Equipment" can refer to any terminal device capable of wireless communication. By example, a User Equipment can be referred to as a communication device, a subscriber station (SS), or a mobile station (MS). A User Equipment can include mobile phones, cellular phones, smartphones, Voice over IP (VoIP) phones, wireless local loop phones, tablets, wearable terminal devices, personal digital assistants (PDAs), portable computers, desktop computers, image capture terminal devices (such as digital cameras), gaming terminal devices, music storage and playback devices, in-vehicle wireless terminal devices, USB dongles, Internet of Things (IoT) devices, watches or other wearables, head-mounted displays (HMDs), vehicles, drones, medical devices and applications (e.g., remote surgery), industrial devices and applications (e.g., robots and / or other wireless devices operating in the context of industrial and / or automated processing chains), consumer electronics devices, devices operating on commercial and / or industrial wireless networks, and so on.
[0257] The UE can have radio access to communication networks (such as 3GPP networks).
[0258] While the terms "a," "an," or "some" may be used multiple times in this disclosure, this does not necessarily mean that each reference refers to the same example(s), or that a particular feature applies only to a single example. Individual features of different examples may also be combined to provide other examples. Furthermore, when a particular feature, structure, or characteristic is described with respect to an example, the application of such a feature, structure, or characteristic to other examples is within the knowledge of those skilled in the art, whether or not it is explicitly described.
[0259] It should be understood that although the terms "first," "second," etc., may be used in this document to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another.
[0260] It should be understood that the foregoing references to various servers can be implemented by means of devices performing at least some of the functions associated with these servers. Furthermore, the means configured to implement a server can also be configured to implement a virtual server instance of that server. This means can be a virtual machine provided by the infrastructure of a cloud computing system (e.g., a hypervisor, processor, and / or memory). This means can be part of a distributed computing system having multiple computing devices (e.g., servers) that communicate with each other via a data network. Any part or all of the operation of the server can be performed by the same means or different means, which can be one or more computing devices in the distributed computing system.
[0261] It should be noted that while some examples related to 5G or 6G networks have been described, similar examples can be applied to other networks and communication systems. Therefore, although some examples have been described above, by way of non-limiting and illustrative examples, and by referring to certain example architectures used in wireless networks, technologies, and standards, additional examples can be applied to any other suitable form of communication system besides those shown and described herein.
[0262] It should also be noted in this document that various variations and modifications may be made to the various examples described herein without departing from the scope of this disclosure.
[0263] As used herein, the phrases “at least one of A or B,” “at least one of A and B,” and “A and / or B” all refer to (A), (B), or (A and B). For the purposes of this disclosure, the phrase “A, B, and / or C” refers to (A), (B), (C), (A and B), (A and C), (B and C), or (A, B, and C). As used herein, “at least one” and “one or more” respectively refer to “any one of at least one” and “any one of one or more.”
[0264] As used herein, the term “or” means “or” in a non-exclusive sense, unless otherwise indicated (e.g., “otherwise” or “or as an alternative”).
[0265] As used herein, unless otherwise expressly stated, “responding to A” execution step does not indicate that the step is executed immediately after “A” occurs, and one or more intermediate steps may be included. Similarly, “based on A” execution step or function does not indicate that the step or function is executed solely based on “A”, as one or more additional conditions may be included.
[0266] Figure 2 A block diagram of apparatus 10 is illustrated by way of example. Apparatus 10 includes, for example, at least one processor 12 and at least one memory 14 storing instructions 15 that, when executed by the at least one processor, cause apparatus 10 to perform methods (or corresponding portions thereof) as disclosed herein. In one example, the at least one memory and instructions (e.g., computer program code, software) are configured, together with the at least one processor, to cause apparatus 10 to perform one or more methods (or corresponding portions thereof) as disclosed herein.
[0267] The processor 12 may include, or be constituted by, one or more circuit systems configured to perform phases of the method as described herein.
[0268] As used herein, the term “circuit system” may refer to one or more of the following: (a) a hardware circuit implementation only (such as an implementation only in analog, digital and / or quantum circuit systems) and (b) a combination of (multiple) hardware circuits and software, such as (if applicable): (i) a combination of (multiple) analog, digital and / or quantum hardware circuits with software / firmware and (ii) any or all portions of (multiple) hardware processors having software (including (multiple) digital and / or quantum processors) and (multiple) memories, which work together to enable a device (such as a mobile device, user device, computing device or server) to perform various functions and (c) any or all portions of (multiple) hardware circuits (such as (multiple) microprocessors, (multiple) processors or (multiple) quantum processors) that require software (e.g., firmware) to operate, but may be absent where operation does not require software.
[0269] This definition of "circuit system" applies to all uses of the term in this disclosure, including in any claim. As another example, as used in this disclosure, the term "circuit system" also covers only the implementation of hardware circuitry or a processor (or processors) or portions thereof and their accompanying software and / or firmware. For example, if applicable to a particular claim element, the term "circuit system" also covers baseband integrated circuits or processor integrated circuits for mobile devices or similar integrated circuits in servers, cellular network devices, or other computing or network devices.
[0270] The memory 14 can be implemented using any suitable data storage technology. The memory may include a database for storing data. The memory 14 may be at least partially located outside the device 10, but may be accessible to the device 10.
[0271] Instruction 15 may be included in a computer-readable medium or a non-transitory computer-readable medium. As used herein, the term “non-transitory” is a limitation of the medium itself (i.e., a tangible medium, not a signal), rather than a limitation of the persistence of data storage (e.g., random access memory (RAM) and read-only memory (ROM).
[0272] For example, device 10 can be implemented as a UE. As another example, the device can be included in such a UE, for example, as a chipset configured to control the UE. Device 10 can be made or configured to perform at least one or more of the methods described herein (or portions thereof).
[0273] As another example, device 10 can be implemented as a server, such as a VAL server or an AIMLE server. In another example, the device can be included in such a server, for example, as a chipset configured to control a server. Device 10 can be made or configured to perform at least one or more of the methods described herein (or portions thereof).
[0274] Device 10 may include a radio interface 16. Radio interface 16 may provide communication capabilities to device 10. Radio interface 16 may include a receiver configured to receive information according to at least one cellular or non-cellular standard. Radio interface 16 may include a transmitter configured to transmit information according to at least one cellular or non-cellular standard. Receivers may include more than one receiver. Transmitters may include more than one transmitter. Radio interface 16 may include a transceiver configured to receive and transmit information according to at least one cellular or non-cellular standard. Transceivers may include more than one transceiver.
[0275] Device 10 may include an interface (e.g., a user interface) 18, including at least one of, for example, a keyboard, microphone, touch screen, display screen, speaker, etc. Interface 18 may be used (e.g., by a user) to control the device. Interface 18 may be located externally to device 10. For example, device 10 may be connected to another device (such as a computer) via a wireless or wired connection, and device 10 may be controlled via the computer (e.g., by a user).
[0276] In some embodiments, example 3GPP proposals are provided.
[0277] The following modifications were proposed.
[0278] 8.18.2 Procedures for supporting VFL
[0279] Prerequisites:
[0280] 1. VFL members have registered their VFL profiles with the AIMLE server. VFL members can update the status or information in their VFL profiles with the AIMLE server.
[0281] 2. Each dataset in the UE (on which AIMLE clients (as VFL members) are deployed) belongs to more than one different data domain.
[0282] 3. The VAL server has successfully subscribed to / registered with the AIMLE server for model training notifications.
[0283] (It should be noted that this amendment concerns...) Figure 3 The topic of discussion Figure 3In this proposal and reference Figure 8 .18.2-1 is provided: for procedures that support VFL)
[0284] 0. The AIMLE client registers with the AIMLE server in accordance with Clause 8.7 of 3GPP TS 23.482[2]. The registration message includes a list of (multiple) feature IDs, their status, and their expiration time.
[0285] 1. The AIMLE server receives the ML model training request from the VAL server as described in Clause 8.3.2. If the ML model training request is received and the AIMLE server determines that VFL training (VFL training between different data domains) is used, the process continues to step 2.
[0286] 2. [Optional] The AIMLE server retrieves the ML model indicated in step 1 using the ML model retrieval process as described in Clause 8.1. If the retrieved model has been trained and meets the machine learning requirements requested by the consumer, the process continues to step 7.
[0287] 3. If the AIMLE client selection criteria were provided in step 1, the AIMLE server continuously monitors and selects AIMLE clients for VFL training, as described in step 4 of clause 8.13.2. If a list of AIMLE clients was provided in step 1, the AIMLE server selects the provided AIMLE client for VFL training.
[0288] 4. The AIMLE server sends client feature selection requests to some clients, requesting the activation or deactivation of some of their feature IDs / feature lists. This message includes the information described in the table below.
[0289] 5. The AIMLE client sends a response indicating the activation or deactivation status of its feature ID(s) or feature list. The response message includes the information described in the table below. Based on the response message, the AIMLE server looks up the common feature ID(s) by comparing the client's feature descriptions / feature ID(s). The AIMLE server considers the shared feature ID(s) of candidate datasets to construct pairwise relationships between them.
[0290] Note: If (multiple) feature IDs are a subset of the union of feature spaces, the AIMLE server obtains hash feature values from candidates for its data samples.
[0291] 6. The AIMLE server obtains VFL member information (VFL member information may be included in the request in step 1 (Table 8.3.3.1-1), or obtained through step 3). The AIMLE server interacts with the VFL members (AIMLE clients deployed on the UE) of each data domain for ML model training capability assessment, as described in clause 8.19.2.
[0292] 7. The AIMLE server determines the VFL members used for this VFL training process based on the information received in step 4 and the parameters received in step 1 (Table 8.3.3.1-1). The AIMLE server determines the VFL members for each data domain. The criteria used by the AIMLE server include:
[0293] - Available data and minimum number of data samples for the same samples among VFL members.
[0294] - The features of samples / datasets within VFL members are aligned with the data labels.
[0295] - Available time for VFL members to support VFL training operations.
[0296] - The capacity and minimum number of VFL members used for VFL training operations.
[0297] - AIML model information for VFL members and for the AIMLE server.
[0298] 8. The AIMLE server coordinates the selected VFL members for VFL training. During VFL training, VFL members send intermediate results to the AIMLE server, and the AIMLE server responds to the VFL members with updated information (e.g., gradients). Information from the AIMLE server can be used to update the model parameters maintained by each VFL member for different data domains.
[0299] 8a. The AIMLE server can report the training status to the VAL server, including intermediate training results. The VAL server can adjust its requests for training the ML model. If the VAL server is providing data labels to complete training, it sends updated training parameters for the AIMLE server to distribute to VFL members. The updated training parameters apply to the model of the VFL member associated with each data domain.
[0300] Note: How the AIMLE server coordinates the selected VFL members to perform VFL training outside the scope of this version.
[0301] 79. The AIMLE server sends a machine ML training notification to the VAL server as described in Clause 8.3.2. If the training schedule has not yet been completed (e.g., there are still training rounds remaining), the AIMLE server configures the next training schedule set, and steps 3 through 6 are repeated for the next training round.
[0302] In some examples, at least some of the processes described herein may be performed by means including components for performing at least some of the processes. Components for performing the method steps disclosed herein may include software and / or hardware components of means 10. For example, at least one processor 12, memory 14, and computer program code form components for performing one or more methods or any examples disclosed herein. As used herein, the term “component” should be interpreted as either the singular (i.e., a single element) or the plural (i.e., a combination of single elements). Thus, the term “component for [performing A, B, C]” should be interpreted to encompass means in which only one component exists for performing A, B, and C, or in which separate components exist for performing A, B, and C, or in which components partially or completely overlap for performing A, B, and C. Furthermore, the terms “component for performing A, component for performing B, component for performing C” should be interpreted as covering an apparatus in which only one component exists for performing A, B and C, or in which separate components exist for performing A, B and C, or in which components exist that partially or completely overlap for performing A, B and C.
[0303] The scope of protection sought by the various examples in this disclosure is defined by the independent claims. Examples and features described in this disclosure that do not fall within the scope of the independent claims, and their features (if any), should be interpreted as examples that aid in understanding the various examples in this disclosure.
[0304] Although various embodiments have been described above with reference to the accompanying drawings, it is apparent that these examples are not limited to the drawings and can be modified in many different ways. Therefore, all words and expressions should be interpreted broadly and are intended to illustrate rather than limit these examples. As technology advances, it will become apparent to those skilled in the art how certain examples can be further implemented and / or modified in various ways. Furthermore, it should be clear to those skilled in the art that the examples (or portions thereof) described herein can, but are not necessarily, combined in various ways with any other examples (or portions thereof) described herein.
Claims
1. A user equipment, comprising: At least one processor; as well as At least one memory storing instructions that, when executed by the at least one processor, cause the user equipment to at least: Provides a Vertical Federated Learning (VFL) client profile for an AI machine learning enabling client for the user device, the VFL client profile including information about one or more features of a data sample of the user device, wherein the user device is a candidate user device for participating in a vertical federated learning machine learning task; The user equipment receives a request from a first server to change the availability of one or more features of the data sample, the request including an identifier of the one or more features of the data sample. The availability of one or more of the one or more features is changed according to the request; as well as A response to the request is sent to the first server, the response indicating that the availability of one or more of the features has been changed.
2. The user equipment of claim 1, wherein the information regarding one or more features of the data sample includes information regarding the availability of a corresponding feature among the one or more features; and / or The information regarding the availability of a corresponding feature among the one or more features indicates that the state of the corresponding feature is either active or deactivated; and / or The information regarding the availability of a corresponding feature among the one or more features indicates that the state of the corresponding feature is either available or unavailable.
3. The user equipment of claim 2, wherein the information regarding one or more features of the data sample includes information regarding the expiration of the availability of the corresponding feature among the one or more features.
4. The user equipment according to any of the preceding claims, wherein the request includes information indicating the priority of a corresponding feature in a vertical federated learning machine learning task, and a change in availability is requested for the priority of the corresponding feature.
5. The user equipment according to any one of claims 1-3, wherein the request includes information indicating the time at which the change in the availability of the corresponding feature will be applied.
6. The user equipment according to any one of claims 1-3, wherein the request includes information indicating that the state of a corresponding feature will be set to active or deactivated to change the state of the corresponding feature; and / or The request includes information indicating whether the state of the corresponding feature will be set to available or unavailable to change the state of the corresponding feature.
7. The user equipment according to any one of claims 1-3, wherein the response includes information indicating the degree to which the request for changing the availability of one or more of the one or more features is satisfied.
8. The user equipment of claim 7, wherein the response includes information indicating the timing of the change in the availability of the corresponding feature relative to the adjustment to which it will be applied.
9. The user equipment according to any one of claims 1-3, wherein the user equipment is further configured to receive from the first server a request for one or more hash values of one or more feature values of one or more data samples, and to send a response to the first server to the request for one or more hash values of one or more hash values comprising one or more feature values of one or more data samples.
10. The user equipment according to any one of claims 1-3, wherein the user equipment is selected to participate in the vertical federated learning task, and the user equipment is further configured such that: Provide the first server with information related to the model trained in the vertical federated learning task; and The updated information is received from the first server, and the updated information updates the model trained in the vertical federated learning task.
