Methods, devices, and systems for federated learning

Abstract

Methods, devices, and systems for use in updating a global ML model in a federated learning system are provided. A computer-implemented method includes receiving first and second model updates. The method also includes estimating a first level of label noise in first local data used to generate the first model update and a second level of label noise in second local data used to generate the second model update, and determining first and second weightings based on the first and second levels of label noise, respectively. The method also modifies the first model update using the first weighting and modifies the second model update using the second weighting. The method also includes initiating transmission of the modified first and second model updates to a global device for use in updating the global ML model.

Description

Technical Field

[0001] The embodiments described herein relate to computer-implemented methods and apparatus for machine learning (ML), particularly for use in updating a global ML model in a federated learning system. Background Technology

[0002] ML is increasingly becoming an integral part of various technological fields, including natural language processing, computer vision, speech recognition, and the Internet of Things (IoT) related to automation and digitization tasks. The successful implementation of ML systems often relies on collecting and processing large amounts of data in a suitable environment. For some ML applications, the data collection process can raise privacy concerns, particularly regarding the sharing of potentially private and / or confidential data with external entities (such as the entity providing the ML model) for model training purposes.

[0003] Some ML techniques can be implemented in ways that help address privacy concerns. One such technique is federated learning (FL), a machine learning approach where training data never needs to leave the user's device. FL is a collaborative form of machine learning where the training process is distributed among many local users. Typically, in an FL system, a single global user (e.g., a server) communicates with multiple local users (e.g., users on local devices); local users typically do not communicate directly with each other. The global user (e.g., the server) may have a role in coordinating among the local users, but most (if not all) of the model training is not performed by the global user (e.g., the server), but rather collaboratively by the local users. Instead of sharing data, local users themselves can use their locally available data to compute ML model updates (e.g., node weights and biases updates within an ML model, such as a neural network (NN)). ML model updates from local users can then be shared with another entity (typically the global user) without requiring any sharing of the local user data itself.

[0004] The typical operation of an FL system is as follows: A global user provides an ML model to at least some of a group of local users. After the ML model is initialized by the local users, a number of local users (possibly all local users) can be selected to improve the model. Each selected local user can train the model using data available to them locally, thus computing model updates. The computed model updates are then sent back to the global user, where they are combined in some way (e.g., averaged, weighted by the number of training samples that may have been used by the local users). The global user then typically applies the combined updates to the ML model using some form of gradient descent to generate an updated ML model. The updated ML model can then be shared with the local users again for use and / or further training.

[0005] ML applications, including deep learning, have achieved success in many fields through the use of massive amounts of data. However, the lack of high-quality labels (i.e., accurate labels) in many real-world scenarios can be a cause for concern. Low-quality labels (i.e., labels that are imprecise or use non-standard terminology) can severely degrade the performance of deep neural networks; however, in some cases, it may be necessary to train using data with low-quality labels. Therefore, in modern deep learning applications, accurate learning from data with low-quality labels (also known as noisy labels) is becoming increasingly important.

[0006] FedCorr, a general multi-level framework for handling heterogeneous label noise in FL, was disclosed by Joohyung, J et al. on November 2, 2023, at https: / / arxiv.org / abs / 2204.04677. FedCorr aims to handle data heterogeneity and uses an adaptive local approximation regularization term based on the estimated local noise level to improve training stability.

[0007] The goal is to provide a method for efficiently estimating label noise on a per-local-user basis and using the label noise estimate to assign relative weights to contributions from the local model for use in updating the global model. Summary of the Invention

[0008] The purpose of this disclosure is to provide methods, apparatus, and computer-readable media that at least partially address one or more of the challenges described above. In particular, the purpose of this disclosure is to allow for label noise estimation without requiring access to local user data other than that of the local user.

[0009] This disclosure provides methods, apparatus, and systems for updating a global ML model in a federated learning system. Embodiments can support accurate label noise estimation and thus support the efficient generation of accurate global ML models by using data impressions that can correspond to local user data.

[0010] An embodiment provides a computer-implemented method for updating a global ML model in a federated learning system. The method includes receiving a first model update from a first local computing device of the federated learning system and a second model update from a second local computing device of the federated learning system. The method further includes estimating label noise at a first level in first local data used to generate the first model update and label noise at a second level in second local data used to generate the second model update, and determining a first weighting based on the estimated first level of label noise and a second weighting based on the estimated second level of label noise. The method also includes modifying the first model update using the first weighting and modifying the second model update using the second weighting; and initiating a transfer of the modified first model update and the modified second model update to a global computing device for use in updating the global ML model.

[0011] In some embodiments, the method may further include a global computing device initiating a transfer of an initial global model to a first local computing device and a second local computing device. In some embodiments, the method may further include a first local computing device training an initial global model using first local data to generate a first model update, and a second local computing device training an initial global model using second local data to generate a second model update.

[0012] Another embodiment provides an arbiter for use in updating a global ML model in a federated learning system. The arbiter includes processing circuitry, one or more interfaces, and memory containing instructions executable by the processing circuitry. The arbiter is operable to receive a first model update from a first local computing device of the federated learning system and a second model update from a second local computing device of the federated learning system. The arbiter is also operable to estimate a first level of label noise in first local data used to generate the first model update and a second level of label noise in second local data used to generate the second model update, and to determine a first weighting based on the estimated first level of label noise and a second weighting based on the estimated second level of label noise. The arbiter is also operable to modify the first model update using the first weighting and the second model update using the second weighting, and to initiate a transfer of the modified first model update and the modified second model update to the global computing device for use in updating the global ML model.

[0013] The federated learning system according to some embodiments may include an arbitrator and may also include a global computing device, a first local computing device, and a second local computing device. The global computing device may be configured to initiate a transfer of an initial global model to the first and second local computing devices. The first local computing device may be configured to train the initial global model using first local data to generate a first model update. The second local computing device may be configured to train the initial global model using second local data to generate a second model update.

[0014] Embodiments of this disclosure are discussed below. The scope of this disclosure is defined by the claims. Attached Figure Description

[0015] The present disclosure is described by way of example only with reference to the following figures, in which: Figure 1 This is a flowchart of the method according to an embodiment; Figure 2A This is a schematic diagram of an arbitrator according to an embodiment; Figure 2B This is a schematic diagram of another arbitrator according to an embodiment; Figure 2C This is a schematic diagram of a global computing device according to an embodiment; Figure 2D This is a schematic diagram of a local computing device according to an embodiment; Figure 3 This is a schematic diagram of an example FL system according to an embodiment; Figure 4 This is a sequence diagram illustrating an example method according to an embodiment; and Figure 5 This is a graph comparing the performance of the comparative embodiment and the conventional FL system. Detailed Implementation

[0016] For purposes of explanation, details are set forth in the following description in order to provide a full understanding of the disclosed embodiments. However, it will be apparent to those skilled in the art that embodiments may be implemented without these specific details or with equivalent arrangements.

[0017] In existing flint and flare (FL) systems, it is typically assumed that all local users have a similar data distribution across all users, and that the data is free of label noise. However, in many cases, this assumption may not hold and could degrade the performance of the FL system. Furthermore, the labels used for a given set of data may depend on the perspective of the users labeling the data. For example, one user might label a feature pattern with 'A', while another user might label the same feature pattern with 'B', leading to label noise.

[0018] In some known systems, label noise can be estimated by analyzing data; however, this requires data sharing and may introduce privacy issues as described above. Therefore, this option is generally not compatible with FL systems designed to provide data privacy.

[0019] The embodiments provide a system for effectively taking into account label noise in label noise. The embodiments can result in the global ML model being trained more efficiently to a given level of accuracy and / or providing a higher level of accuracy over a given number of training epochs. Robust ML modeling with reduced sensitivity to label noise can be provided.

[0020] In some embodiments, a validated dataset can be used to compute label noise in a federated learning system. The validated dataset can be publicly available to all local users (i.e., users of the local computing device) and global users (i.e., users of the global computing device). In some embodiments, instead of sending ML model updates directly to the global computing device, the ML model updates can be sent to an arbitrator. The arbitrator can then create an anonymized data impression and thereby estimate the label noise level by comparing the constructed data impression with the actual features in the validated data. Global users can then use the estimated label noise level to assign relative weights to local device ML model updates for use when updating the global ML model. Therefore, the label noise level can be accurately estimated without requiring access to the original user data beyond the local users.

[0021] Depend on Figure 1 A method according to an embodiment is shown. Figure 1 This is a flowchart illustrating a computer-implemented method for use in updating a global ML model in a FL system. The embodiment is applicable to a wide variety of ML model types, such as NN, random forest, etc. This computer-implemented method can be executed by any suitable device, such as... Figure 2A and Figure 2B Arbitrators such as arbitrators 20A and 20B, as shown, perform the operation. When the network is a communication network (or part of it), the method can be performed by a device (e.g., an arbitrator) that is or forms part of a network node (e.g., a base station or core network node) (or can be incorporated into the base station or core network node). The communication network can be, for example, a 5G network or a 6G network. Alternatively, the arbitrator can be or forms part of a system for controlling autonomous and / or semi-autonomous vehicles. In some embodiments, the method can be performed at least partially by a global computing device 31 and / or multiple local computing devices 32, such as... Figure 2C and Figure 2D The devices shown in the diagram perform this function.

[0022] like Figure 1As shown in step S102, the method includes receiving a first model update from a first local computing device of the federated learning system and a second model update from a second local computing device of the federated learning system. Figure 2A As shown, the steps of receiving the first and second model updates can be executed according to a computer program stored in memory 23, which is executed by processor 21 in conjunction with one or more interfaces 22 of arbitrator 20A. Alternatively, the steps of receiving the first and second model updates can be executed by transceiver 24 of arbitrator 20B, as shown. Figure 2B As shown. In some embodiments, model updates may include updated node weights and biases of the ML model (e.g., a neural network), which have been obtained through training at a local computing device (using local computing device data). The term "node" in this context refers to the software structure within the ML model, rather than the physical device. In some embodiments, the method also includes a transfer of an initial global ML model initiated by global computing device 31 to local computing device 32; typically, the same initial global model can be sent to all local computing devices. Figure 2C As shown, the step of initiating the transfer of the initial global ML model can be executed according to a computer program stored in memory 35, which is executed by processor 33 in conjunction with one or more interfaces 34 of global computing device 31. The initial global model can then be used as a starting point for training by local computing devices; as an example, a first local computing device can use first local data to train the initial global model to generate a first model update, and a second local computing device can use second local data to train the initial global model to generate a second model update. For completeness, note that all embodiments discussed herein may include additional local computing devices in addition to the first and second local computing devices, i.e., equivalent steps performed by the first and second local computing devices may also be performed by a third local computing device, a fourth local computing device, and so on. The number of local devices involved in an FL system depends on the specific functionality of the system; some FL systems may include dozens or hundreds of local computing devices. In some embodiments, local data of local computing devices may be kept confidential at the local computing device; in other embodiments, some or all of the local data may be shared with other local computing devices or global computing devices.

[0023] ML models can be trained at local computing devices in any suitable manner. An example of a suitable training method is to utilize reinforcement learning (RL). RL allows ML models to learn by attempting to maximize the expected cumulative reward of a set of actions through trial and error. RL agents (i.e., systems that use RL to improve the performance of ML models on a given task over time) can be closely associated with the system (environment) they are being used to model / control and learn through experience in performing actions that change the state of the environment. Additionally or alternatively, RL can be used to train ML models (where simulations of the environment are used) before the trained ML model is used to control the actual environment; this option can be useful when it is desirable to avoid having an untrained or partially trained ML model control the actual operating environment. After the training epoch, the updated ML model parameters (ML model updates) from each local computing device that performed the model training can then be passed to a global computing device (which may include an arbitrator) or to separate arbitrators, depending on the system configuration. The steps of training the initial global ML model and initiating the transfer of ML model updates to the global computing device 31 or arbitrator 20 can be executed according to a computer program stored in memory 38, which is executed by processor 36 in conjunction with one or more interfaces 37 of local computing device 32, such as... Figure 2D As shown.

[0024] When the first and second ML model updates have been received (e.g., by an arbitrator that may be part of a global computing device), the method continues to estimate a first level of labeled noise in the first local data used to generate the first model update and a second level of labeled noise in the second local data used to generate the second model update, such as Figure 1 As shown in step S104. As described above, in some embodiments, an estimate can be provided even when the arbitrator cannot access local data on the local computing device. Figure 2A As shown, the step of estimating the tag noise level can be performed according to a computer program stored in memory 23, which is executed by processor 21 in conjunction with one or more interfaces 22 of arbitrator 20A. Alternatively, the step of estimating the tag noise level can be performed by estimator 25 of arbitrator 20B, such as... Figure 2B As shown.

[0025] In some embodiments, the arbitrator may use a validated dataset when estimating the first level of label noise and the second level of label noise. The validated dataset may be used to provide a benchmark dataset against which model updates generated at a local computing device using unvalidated local data can be compared. In some embodiments, the validated dataset may be publicly available. The validated dataset may be hosted by the arbitrator, thereby providing the arbitrator with easy access to the validated dataset; alternatively, the arbitrator may retrieve the validated dataset from a server 28 separate from but connected to the arbitrator (e.g., ...). Figure 2B (As shown).

[0026] In an embodiment where the arbitrator uses a validated dataset when estimating label noise levels, the arbitrator can process data from the validated dataset using a given model update to generate a given output, and then process the given output in reverse using the given model update to generate a given data impression corresponding to the given model update. The arbitrator can then compare the given data impression with data from the validated dataset and estimate the label noise of the given model update based on the divergence between the given data impression and the data from the validated dataset.

[0027] According to an embodiment, an example of how label noise related to model updates can be estimated is as follows. In the example below, the public dataset is... In this example, the public dataset is hosted at the arbitrator and is available to both the local and global computing devices. In this example, the arbitrator first initializes the global model. The data is sent to the local computing device. Subsequently, at iteration 'i', the local computing device updates the model by training the model using local data. and update the model Send to the arbitrator.

[0028] The arbitrator then estimates the label noise level in the local data used to generate the model update. In the example according to the embodiment, anonymized data impressions can be used to estimate the label noise level. Given a model M, which is correlated between input X (which is local user data) and output model update y, where It is the available feature set (where N is the number of features in the local user data) and Anonymous data impressions are anonymized feature sets X* that have the same properties as X. To determine X*, softmax values ​​from a Dirichlet distribution are sampled. To match the obtained softmax values ​​to the real scene, a class similarity matrix is ​​used to control the sampling distribution. The class similarity matrix contains important information about the degree of similarity between classes. If classes are similar, the softmax values ​​will typically be uniformly concentrated on these labels, and vice versa. The class similarity matrix is ​​obtained by considering the weights of the last layer of the model (which has been trained using local user data). In general, any classification model has a final layer that is a fully connected layer with softmax non-linearity. If the classes are similar, we find similar weights between the connections from the previous layer to the nodes of the class. Obtain the category similarity matrix (C), where w i It is the weight vector that connects the previous layer node to the category node i, and This is the similarity matrix of K categories in the data. When the category similarity matrix (C) is already obtained, the softmax value can be sampled as... Where K is a concentration parameter that controls the distribution of softmax values ​​across classes in the data.

[0029] Once the softmax values ​​have been determined, these values ​​can then be used to generate an anonymous impression of the data. This involves using... To calculate the softmax value (Y) corresponding to category k. k Sampling from the Dirichlet distribution constructed as described above, the data impression X* can be obtained by using model M and softmax value Y. k Solving optimization problems To obtain this, the input X is set to a random input, and then iterated until the cross-entropy loss (L) between two iterations is reached. CE The change is less than the significance value. This process is repeated for each of the K categories to obtain a data impression for each category, thus providing anonymized data features for each category.

[0030] Once the anonymized data impression has been calculated, the difference between the input data (from the validated dataset) and the anonymized data impression is then obtained.

[0031] To summarize the above examples, model updates from N local computing devices Empirical data used to process validated datasets. x To generate output Then, the output is a reconstruction of the empirical data used to generate the input (the reconstruction is an anonymized impression of the data). Then the anonymized data impressions can be... Compared with the empirical data input x The comparison is made between the divergence levels of the two. D use (Hereinafter referred to as Equation 1) is used for calculation.

[0032] Given the divergence level of the model update D This is related to the level of label noise in the data used to generate model updates, and therefore to the reliability of the model updates. Thus, iterations can be obtained. i Local computing device j The label noise level.

[0033] Once the first level of labeled noise in the first local data used to generate the first model update and the second level of labeled noise in the second local data used to generate the second model update have been estimated (e.g., along with any additional labeled noise levels used for the other local datasets, as described above), the weighting of the model update can then be determined, such as... Figure 1 As shown in step S106. The step of determining the first weighting based on the estimated first level of tag noise and the second weighting based on the estimated second level of tag noise can be executed according to a computer program stored in memory 23, which is executed by processor 21 in conjunction with one or more interfaces 22 of arbitrator 20A, such as Figure 2A As shown. Alternatively, the step of determining the first weighting based on the estimated first level of tag noise and the second weighting based on the estimated second level of tag noise can be performed by the determiner 26 of the arbitrator 20B, as follows. Figure 2B As shown. Typically, the deterministic weighting of a given model update is inversely proportional to the estimated level of label noise corresponding to the model update, such that model updates generated using local data with more noisy labels are given lower weighting and therefore have less impact when updating the global ML model.

[0034] like Figure 1 As shown in step S108, once the first and second weights have been determined, the method continues to modify the first model update using the first weight and modify the second model update using the second weight. In some embodiments, the modification of the model update may include appending the determined weights to the model update, or by applying the determined weights to the model update (e.g., as a multiplier). Figure 2AAs shown, the steps for modifying the model update can be executed according to a computer program stored in memory 23, which is executed by processor 21 in conjunction with one or more interfaces 22 of arbitrator 20A. Alternatively, the steps for modifying the model update can be executed by modifier 27 of arbitrator 20B, such as... Figure 2B As shown.

[0035] After updating based on the weighted modification model, such as Figure 1 As shown in step S110, the method continues to initiate the transmission of a modified first model update and a modified second model update to the global computing device for use in updating the global ML model. In some embodiments, the arbitrator may perform the transmission itself; alternatively, the arbitrator may instruct the transmission to be performed. Where the arbitrator is part of the global computing device, the transmission may be within said device. Where the arbitrator is separate from but connected to the global computing device, any suitable wired or wireless transmission method may be used. Figure 2A As shown, the steps for initiating the transmission of a modified model update can be executed according to a computer program stored in memory 23, which is executed by processor 21 in conjunction with one or more interfaces 22 of arbitrator 20A. Alternatively, as Figure 2B As shown, the step of initiating the transmission of the modified model update can be performed by the transceiver 24 of the arbitrator 20B.

[0036] In some embodiments, when a modified model update (e.g., a modified first model update, a modified second model update, a modified third model update, etc.) has been provided to the global computing device, the global computing device can use the modified first model update, the modified second model update, etc., to generate an updated global ML model. As an example, the updated global model can be based on... (Hereinafter referred to as Equation 2) is generated using the nomenclature illustrated in the example above. In some embodiments, the global computing device can then initiate a transfer of the updated global ML model to the local computing device (either by itself or by instructing another device). The steps of generating the updated global ML model and initiating the transfer of the updated global ML model can be executed according to a computer program stored in memory 35, which is executed by processor 33 in conjunction with one or more interfaces 34 of the global computing device 31, such as... Figure 2C As shown. As described above, the local computing device can then use the updated global ML model and / or initiate further training periods for the updated global ML model. The steps of using and / or further training the updated global ML model can be executed according to a computer program stored in memory 38, which is executed by processor 36 in conjunction with one or more interfaces 37 of the local computing device 32, such as Figure 2D As shown.

[0037] Figure 3 This is a schematic diagram illustrating an example FL system according to an embodiment. Figure 4 It corresponds to Figure 3 The sequence diagram. In Figure 3 and Figure 4 In the example shown, global user 31 (i.e., the user of the global computing device) sends the initial global ML model to local users 32A, 32B, and 32C (collectively referred to as 32), i.e., the users of the local computing device (see [link to example]). Figure 4 (S401, S402, and S403 in the text). Figure 3 and Figure 4 The example shows local users 1, 2, and N; as explained above, the number of local computing devices participating in the FL system varies depending on the specific implementation. Local user 32 then trains a global ML model using (private) local data (see S404, S405, and S406) and sends the model update to arbitrator 20 (see S407, S408, and S409). Local data is not sent to the arbitrator along with the model update. Figure 3 In the example FL system shown, the arbitrator is separate from the global computing unit. The arbitrator then estimates the label noise level in the (private) local data used to generate the model update (using Equation 1 as described above) and modifies the model update accordingly (see S410 and S411). The modified model update is then sent to the global computing unit (see S412) and used to generate an updated global ML model using Equation 2 as described above (see S413). This updated global ML model can then be sent by the global user to the local user (e.g., ...). Figure 3 (as shown), to restart the loop.

[0038] Figure 5 The graph in the figure shows a comparison of the effectiveness of an FL system implementing the method according to the embodiment (solid line) and a known FL system not utilizing the method according to the embodiment (dashed line). In the known FL system, model updates from the local computing device are combined by a simple average at the global computing device to generate an updated global ML model; there is no weighting or modification of the model updates. The x-axis of the graph shows the number of training iterations, while the y-axis shows the accuracy of the global ML model generated using model updates from the local computing device.

[0039] exist Figure 5 In the example, the private user data at the local computing device includes three tags. To introduce label noise, labels corresponding to the same features are modified to artificially introduce label noise. The label noise is as follows: User 1: Noise level is zero, meaning this user's label pattern is the same as the pattern in the validated dataset. User 2: 50% noise level, meaning 50% of the labels were manually mislabeled. User 3: 30% noise level, meaning 30% of the labels were manually mislabeled.

[0040] To generate the results, the global user shares the initial global model with all three local users. The model chosen is a 3-layer convolutional neural network (CNN) model. Figure 5 The results were obtained using 10 training iterations. The final accuracy obtained at the end of 10 iterations using different methods is shown in Table 1 below: Table 1 As from Figure 5 As can be seen from Table 1, the example method according to the embodiments results in higher accuracy compared to typical FL methods. The embodiments can robustly handle label noise (especially compared to typical FL methods) and can provide high accuracy and fast convergence.

[0041] The embodiments can be used in any system employing FL. As an example, the embodiments can be used in a system where a first local computing device and a second local computing device are controllers of an autonomous vehicle, and where the global ML model is an autonomous vehicle control model. In such an embodiment, the global ML model can be used for the control of the autonomous vehicle. The embodiments may be particularly well-suited for applications where autonomous vehicles in a given area are likely to be provided by different vendors. Vendors may use their own labels and label data using their domain expertise, resulting in inconsistent training data used by different local computing devices. To address privacy and security concerns, FL can be used to create global models. Without considering inherent labeling differences, the performance of global ML models generated by FL systems is likely to be unsatisfactory. Using the method according to the embodiments can help circumvent potential data labeling problems and provide a globally performing ML model that performs satisfactorily.

[0042] As another example, the embodiment can also be used in a system where the first and second local computing devices are communication network controllers, and where the global ML model is a network anomaly detection model. In such an embodiment, the global ML model can be used for network anomaly detection. Anomaly detection from AI can effectively enhance and automate early detection, prediction, and decision-making regarding operational and business processes. When the timing of detection deviations is improved, incident resolution can be achieved faster, reducing costs associated with outages. The embodiment can be used to understand incorrect or toxic labels and analyze multiple dimensions of data sources—viewing cell, subscriber, and device-level key performance indicators (KPIs), fault monitoring in network devices, and cross-domain correlated alerts for noise reduction and root cause analysis.

[0043] It will be understood that the examples disclosed herein can be virtualized so that the methods and processes described herein can run in a cloud environment.

[0044] The methods disclosed herein can be implemented in hardware or as a software module running on one or more processors. The methods can also be executed according to the instructions of a computer program, and this disclosure also provides a computer-readable medium having a program stored thereon for performing any of the methods described herein. The computer program embodying this disclosure can be stored on a computer-readable medium, or it can take the form of, for example, a signal, such as a downloadable data signal provided from an Internet website, or it can take any other form.

[0045] Generally, various exemplary embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. For example, some embodiments may be implemented in hardware, while others may be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device, although this disclosure is not limited thereto. While various aspects of exemplary embodiments of this disclosure may be illustrated and described as block diagrams, flowcharts, or using some other graphical representation, it is well understood that, by way of non-limiting example, these blocks, devices, systems, techniques, or methods described herein may be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0046] Therefore, it should be understood that at least some aspects of the exemplary embodiments of this disclosure can be implemented in various components such as integrated circuit chips and modules. Therefore, it should be understood that the exemplary embodiments of this disclosure can be implemented in devices embodied as integrated circuits, wherein the integrated circuits may include circuitry (and possibly firmware) embodying at least one or more of a data processor, digital signal processor, baseband circuitry, and radio frequency circuitry, which are configurable to operate according to the exemplary embodiments of this disclosure.

[0047] It should be understood that at least some aspects of the exemplary embodiments of this disclosure can be embodied in computer-executable instructions, such as one or more program modules, which are executed by one or more computers or other devices. Typically, program modules include routines, programs, objects, components, data structures, etc., which perform specific tasks or implement specific abstract data types when executed by a processor in a computer or other device. The computer-executable instructions can be stored on a computer-readable medium, such as a hard disk, optical disk, removable storage medium, solid-state memory, RAM, etc. As those skilled in the art will appreciate, the functionality of a program module can be combined or distributed as needed in various embodiments. Furthermore, this functionality can be wholly or partially embodied in firmware or hardware equivalents, such as integrated circuits, field-programmable gate arrays (FPGAs), etc.

[0048] References to "an embodiment," "embodiment," etc., in this disclosure indicate that the described embodiment may include specific features, structures, or characteristics, but it is not necessary for every embodiment to include specific features, structures, or characteristics. Furthermore, such phrases do not necessarily refer to the same embodiment. Additionally, when a specific feature, structure, or characteristic is described in connection with an embodiment, it is believed that implementing such a feature, structure, or characteristic in conjunction with other embodiments is within the knowledge of those skilled in the art, whether explicitly described or not.

[0049] It should be understood that although the terms “first,” “second,” etc., may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of this disclosure, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element. As used herein, the term “and / or” includes any and all combinations of one or more of the associated listed terms.

[0050] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. As used herein, the singular forms “a” (“a”, “an”) and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the terms “comprise” (“comprising”), “has” (“having”), and / or “include” (“including”), when used herein, specify the presence of the stated feature, element, and / or component, but do not exclude the presence or addition of one or more other features, elements, components, and / or combinations thereof. The term “connect” (“connects”, “connecting”, and / or “connected”) as used herein covers direct and / or indirect connections between two elements.

[0051] This disclosure includes any novel features or combinations of features expressly disclosed herein, or any generalization thereof. Various modifications and adaptations to the foregoing exemplary embodiments of this disclosure will become apparent to those skilled in the art when read in conjunction with the accompanying drawings, in light of the foregoing description. However, any and all modifications will still fall within the non-limiting scope of the exemplary embodiments of this disclosure. For the avoidance of doubt, the scope of this disclosure is defined by the claims.

Claims

1. A computer-implemented method for use in updating a global machine learning ML model in a federated learning system, the method comprising: Receive (S102) a first model update from a first local computing device of the federated learning system and a second model update from a second local computing device of the federated learning system; Estimation (S104) is used to generate a first level of label noise in the first local data updated by the first model and to generate a second level of label noise in the second local data updated by the second model; Determine (S106) a first weighted sum of tag noise based on the estimated first level and a second weighted sum of tag noise based on the estimated second level; The first model is updated using the first weighted modification (S108), and the second model is updated using the second weighted modification; as well as The transmission of the modified first model update and the modified second model update to the global computing device is initiated (S110) for use in updating the global ML model.

2. The method according to claim 1, wherein, When estimating the label noise at the first level and the label noise at the second level, a validated dataset is used.

3. The method according to claim 2, wherein, The arbitrator (20) hosts the verified dataset, or retrieves the verified dataset from the server (28).

4. The method according to any one of claims 2 and 3, wherein, When estimating a given level of label noise, the data from the validated dataset is processed using a given model update to generate a given output, and then the given output is processed in reverse using the given model update to generate a given data impression corresponding to the given model update.

5. The method according to claim 4, wherein, When estimating the label noise at the given level, the given data impression is compared with data from the validated dataset, and the label noise for the given model update is estimated based on the divergence between the given data impression and the data from the validated dataset.

6. The method according to any of the preceding claims, further comprising: The global computing device initiates the transmission of the initial global model to the first local computing device and the second local computing device; The first local computing device trains the initial global model using first local data to generate the first model update; as well as The initial global model is trained by the second local computing device using the second local data to generate the second model update.

7. The method according to claim 6, wherein, The first local data is kept confidential at the first local computing device, and the second local data is kept confidential at the second local computing device.

8. The method according to any of the preceding claims, further comprising, at the global computing device, generating an updated global model using the modified first model update and the modified second model update.

9. The method of claim 8, further comprising, at the global computing device, initiating a transfer of the updated global model to at least one of the first local computing device and the second local computing device.

10. The method according to any of the preceding claims, wherein, The first model update includes a first node weight and a first bias, and the second model update includes a second node weight and a second bias.

11. The method according to any of the preceding claims, wherein, The arbitrator is part of the global computing device, or the arbitrator is separate from the global computing device but connected to it.

12. The method according to any of the preceding claims, wherein, The first local computing device and the second local computing device are controllers for autonomous vehicles, and the global ML model is an autonomous vehicle control model.

13. The method of claim 12, further comprising using the global ML model in the control of the autonomous vehicle at one or more of the first local computing device and the second local computing device.

14. The method according to any one of claims 1 to 11, wherein, The first local computing device and the second local computing device are communication network controllers, and the global ML model is a network anomaly detection model.

15. The method of claim 14, further comprising using the global ML model in the detection of network anomalies at one or more of the first local computing device and the second local computing device.

16. An arbitrator (20A) for use in updating a global machine learning ML model in a federated learning system, the arbitrator comprising processing circuitry (21), one or more interfaces (23), and a memory (22) containing instructions executable by the processing circuitry (21), thereby enabling the arbitrator (20A) to: Receives a first model update from a first local computing device (32A) of the federated learning system and a second model update from a second local computing device (32B) of the federated learning system; Estimate the label noise at a first level in the first local data used to generate the first model update and the label noise at a second level in the second local data used to generate the second model update; Determine a first weighting based on the estimated first level of label noise and a second weighting based on the estimated second level of label noise; The first model update is modified using the first weighting, and the second model update is modified using the second weighting; as well as The transmission of the first modified model update and the second modified model update to the global computing device (31) is initiated for use in updating the global ML model.

17. The arbitrator (20A) according to claim 16, wherein, The arbitrator (20A) is configured to use a validated dataset to estimate the label noise at the first level and the label noise at the second level.

18. The arbitrator (20A) according to claim 17, wherein, The arbitrator (20A) hosts the verified dataset, or the arbitrator (20A) is configured to retrieve the verified dataset from the server (28).

19. The arbitrator (20A) according to any one of claims 17 and 18, wherein, The arbitrator (20A) is configured to process data from the validated dataset using the given model update when estimating a given level of label noise, to generate a given output, and to reverse process the given output using the given model update to generate a given data impression corresponding to the given model update.

20. The arbitrator (20A) according to claim 19, wherein, The arbitrator (20A) is configured to compare the given data impression with data from the validated dataset when estimating the given level of label noise, and to estimate the label noise of the given model update based on the divergence between the given data impression and the data from the validated dataset.

21. A federated learning system comprising an arbitrator (20A) according to any one of claims 16 to 20, and further comprising the global computing device (31), the first local computing device (32A), and the second local computing device (32B), wherein: The global computing device (31) is configured to initiate the transfer of the initial global model to the first local computing device (32A) and the second local computing device (32B); The first local computing device (32A) is configured to use first local data to train the initial global model to generate the first model update; as well as The second local computing device (32B) is configured to use second local data to train the initial global model to generate the second model update.

22. The joint learning system according to claim 21, wherein, The first local computing device (32A) is configured to keep the first local data confidential at the first local computing device (32A), and the second local computing device (32B) is configured to keep the second local data confidential at the second local computing device (32B).

23. The joint learning system according to any one of claims 21 and 22, wherein, The global computing device (31) is configured to generate an updated global model using the modified first model update and the modified second model update.

24. The joint learning system according to claim 23, wherein, The global computing device (31) is configured to initiate the transmission of the updated global model to at least one of the first local computing device (32A) and the second local computing device (32B).

25. The joint learning system according to any one of claims 21 to 24, wherein, The first model update includes a first node weight and a first bias, and the second model update includes a second node weight and a second bias.

26. The joint learning system according to any one of claims 21 to 25, wherein, The arbiter (20) is part of the global computing device (31), or the arbiter (20) is separate from the global computing device (31) but connected to the global computing device (31).

27. The joint learning system according to any one of claims 21 to 26, wherein, The first local computing device (32A) and the second local computing device (32B) are controllers for autonomous vehicles, and the global ML model is an autonomous vehicle control model.

28. The joint learning system according to claim 27, wherein, One or more of the first local computing device (32A) and the second local computing device (32B) are configured to use the global ML model in the control of autonomous vehicles.

29. The joint learning system according to any one of claims 21 to 26, wherein, The first local computing device (32A) and the second local computing device (32B) are communication network controllers, and the global ML model is a network anomaly detection model.

30. The joint learning system according to claim 29, wherein, One or more of the first local computing device (32A) and the second local computing device (32B) are configured to use the global ML model in the detection of network anomalies.

31. A computer-readable medium comprising instructions that, when executed on a computer, cause the computer to perform the method according to any one of claims 1 to 15.

Images