Edge terminal participating in cloud-edge federated learning and operation method, and cloud-edge federated learning system

Abstract

This invention belongs to the field of edge computing technology. It discloses an edge device participating in cloud-edge federated learning and its manipulation method, as well as a cloud-edge federated learning system. In the edge device manipulation method, if participating in the current round of federated learning, a global model and a global estimated gradient are obtained, trained to obtain a local model g, and the local model g and its local estimated gradient are uploaded to the central server, where the local estimated gradient is reset. If not participating in the current round of federated learning, the gradient of each new sample on the local model g is calculated. Based on the updated local estimated gradient, the sample with the smallest gradient projection in the dataset is placed in the recycling area. When the sample data in the recycling area reaches a preset upper limit, the sample with the lowest importance is deleted from the local dataset, and the samples with increased data importance are reinserted into the dataset. Based on this method, in a cloud-edge fusion architecture, the storage pressure on edge devices can be alleviated while simultaneously preventing data forgetting.

Description

Technical Field

[0001] This invention belongs to the field of edge computing technology, and more specifically, relates to an edge device and its control method for participating in cloud-edge federated learning, and a cloud-edge federated learning system. Background Technology

[0002] With the rapid development of IoT and 5G technologies, the demand for real-time processing of massive amounts of data is increasing daily, and traditional cloud computing struggles to meet the needs of optimizing data processing efficiency and reducing latency. Cloud-edge converged systems push data processing to edge devices, enabling dynamic scheduling and collaborative optimization of resources to meet the requirements of low latency, bandwidth optimization, and privacy and security protection.

[0003] In cloud-edge converged storage and computing architectures, for streaming data federated learning scenarios, such as connected vehicles, financial data, and video surveillance, the local data of clients is constantly increasing, while the storage space of edge devices is limited, making it difficult to store the massive amount of streaming data. In order to alleviate the storage pressure on edge devices, it is usually chosen to delete relatively old historical data. However, in application scenarios such as intelligent driving, robot control, and human-computer interaction, which require the model to accumulate knowledge over a long period of time, deleting historical data will significantly lead to the forgetting of previously learned task knowledge, causing data forgetting problems.

[0004] Therefore, in a cloud-edge converged architecture, how to alleviate the storage pressure on edge devices while preventing data loss is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides an edge device and control method for participating in cloud-edge federated learning, and a cloud-edge federated learning system. Its purpose is to alleviate the storage pressure of edge devices and at the same time avoid data forgetting in the cloud-edge fusion architecture.

[0006] To achieve the above objectives, the following technical solution is proposed.

[0007] According to a first aspect of the present invention, an edge terminal manipulation processing method for participating in cloud-edge federated learning is provided, comprising: selecting to perform corresponding operations based on whether the edge terminal participates in the current round of federated learning; If the edge device participates in this round of federated learning, a model update operation is performed, including: obtaining the global model distributed by the central server. and global gradient estimation Using old samples for the global model Training is performed to obtain an updated local model g, and the gradients of each old sample are updated to the gradients of each old sample on the updated local model g. The updated local model g and the updated local estimated gradients are uploaded to the central server for aggregation, and the local estimated gradients are reset to the initial gradients. Here, the old samples are the samples in the local dataset. If the edge does not participate in this round of federated learning, then data preparation operations are performed, including: periodically screening samples to be retrieved based on a screening mechanism, wherein the screening mechanism includes: periodically adding new samples to the dataset and calculating the performance of each new sample added. Gradient on the latest local model g Based on gradient Update the local estimated gradient , where n is the index of the new sample. This represents the nth new sample updated after the local gradient estimation is reset. When n=1, When n > 1, Compare the gradients of each sample in the dataset with the latest globally estimated gradient obtained locally. The gradient projection is used to select the sample with the smallest gradient projection as the sample to be recycled and remove it from the dataset and place it in the recycling area. When the sample data in the recycling area reaches a preset upper limit, the data importance of each sample to be recycled is compared and the sample with the lowest importance is deleted from the local dataset. If the data importance of a sample to be recycled increases, it is removed from the recycling area and put back into the dataset. The data importance of each sample is the gradient projection of the sample, the sample gradient, and the globally estimated gradient. The result is obtained by weighted summation of the similarities.

[0008] According to a second aspect of the present invention, an edge device participating in cloud-edge federated learning is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0009] According to a third aspect of the present invention, a cloud-edge federated learning system is provided, comprising a central server and edge devices; the central server is used to aggregate the local models uploaded by each edge device that participated in the previous round of federated learning to obtain a global model in each round of federated learning. The global estimated gradient is obtained by aggregating the local estimated gradients uploaded by each edge node participating in the previous round of federated learning. and the global model and global gradient estimation Distribute to the edge devices participating in the current round of federated learning; The edge device is the edge device described above, used to select and execute corresponding operations based on whether the edge device participates in this round of federated learning. If the edge device participates in this round of federated learning, a model update operation is performed; if the edge device does not participate in this round of federated learning, a data processing operation is performed.

[0010] In summary, compared with the prior art, the technical solutions conceived in this invention have the following main advantages: The method provided by this invention selects and executes corresponding operations at the edge based on whether it participates in the current round of federated learning. If the edge participates, a model update operation is performed; if it does not participate, a data preparation operation is performed. During data preparation, the gradient projection of the sample gradient onto the global estimated gradient is used as an evaluation metric to filter out samples to be recycled. This allows for the selection of high-quality data samples that are more conducive to model training and retain them in the local training dataset. Low-quality samples with small projection values ​​(i.e., a large difference between the sample gradient direction and the model update direction), or samples that have already been sufficiently trained, are placed in the recycling region. For samples in the recycling region, the gradient projection of the sample and the sample gradient versus the global estimated gradient are compared. The weighted sum of similarities is used as an evaluation metric for sample importance. This allows for a reasonable assessment of the impact of each data sample in the recycling region on the current model update. Data samples whose gradient direction aligns with the current model update direction—that is, those more beneficial to model training—are retained, while samples with low contribution to model training are deleted to alleviate storage pressure at the edge. Furthermore, by reintegrating data with increased importance into the local training dataset, these forgotten data samples can be reintegrated into the model's training. This allows the model to continue training on previously trained but now forgotten high-quality samples, thus mitigating the problem of data forgetting in the model. In summary, the method provided by this invention can perform online data selection and solve the data forgetting problem in streaming data federated learning scenarios under cloud-edge fusion systems, under the premise of limited storage space on edge clients. This effectively improves the model performance of cloud-edge federated learning and reduces the storage burden at the edge. Attached Figure Description

[0011] Figure 1 This is a flowchart of the steps of the edge manipulation processing method participating in cloud-edge federated learning in one embodiment of the present invention; Figure 2 This is an architecture diagram of a cloud-edge federated learning system according to an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0013] In a first aspect, the present invention claims a method for edge manipulation processing that participates in cloud-edge federated learning.

[0014] like Figure 1 The diagram shown is a flowchart of the edge manipulation processing method participating in cloud-edge federated learning according to an embodiment of the present invention. The following is a summary of the steps. Figure 1 This method will be described in detail.

[0015] The corresponding operation is selected based on whether the edge device participates in this round of federated learning. If the edge device participates in this round of federated learning, the model update operation is performed; if the edge device does not participate in this round of federated learning, the data cleaning operation is performed.

[0016] Specifically, since federated learning requires the participation of multiple edge devices, and considering that edge devices may be unable to participate in each training session due to their own reasons, two operation modes are designed, corresponding to the cases of participating in training and not participating in training, respectively. When choosing to participate in training, a model update operation is performed; when not participating in training, a data cleanup operation is performed.

[0017] Perform model update operations, including: obtaining the global model distributed by the central server. and global gradient estimation Using old samples for the global model The process involves training to obtain an updated local model g, updating the gradients of each old sample to their corresponding gradients on the updated local model g, uploading the updated local model g and the updated local estimated gradients to the central server for aggregation, and resetting the local estimated gradients to 0. Here, "old samples" refers to samples from the local dataset. These samples can be, for example, lost data such as video frames.

[0018] Specifically, the global model issued by the central server The global gradient estimate is obtained by the central server aggregating the local models g uploaded by each edge device participating in the previous training round. The central server provides the locally estimated gradients uploaded by each edge device participating in the previous training round. The result of aggregation.

[0019] Specifically, since the edge continuously collects new samples, when the edge needs to participate in training, the newly collected samples (new samples) are not used to train the model. Only the samples that have already been cleaned and put into the local dataset are used to train the model. The new samples are temporarily stored and the data cleanup operation is performed when the edge is no longer involved in training.

[0020] Specifically, after obtaining the global model Then, the local old samples can be used to train the updated local model g. The specific training method is not limited here. During the training of the updated local model g using old samples, the gradients of each old sample on the updated local model g can be obtained simultaneously. The gradient of a sample on the model refers to the partial derivative of the model parameters with respect to the loss function of a single sample. It is used to guide parameter updates to minimize the loss. Therefore, the gradients of each old sample can be updated simultaneously during training.

[0021] Subsequently, the updated local model g and the updated local estimated gradients will be used. The data is uploaded to the central server for aggregation, and the locally estimated gradients are reset to 0. The updated locally estimated gradients... This is the locally estimated gradient updated when the edge did not participate in the most recent training. That is, the locally estimated gradient updated by the edge when it did not participate in the most recent training. It is necessary to wait until the edge device participates in training before uploading to the central server, and then immediately reset the local estimated gradient to the initial gradient, which will serve as the initial value for the next update of the local estimated gradient. In a specific embodiment, resetting the local estimated gradient to the initial gradient means resetting the local estimated gradient to 0.

[0022] Performing data processing operations includes: periodically screening samples to be retrieved based on a screening mechanism, wherein the screening mechanism includes: periodically adding new samples to the dataset and calculating the value of each new sample added. Gradient on the latest local model g Based on gradient Update the local estimated gradient , where n is the index of the new sample. This represents the nth new sample updated after the local gradient estimation is reset. When n=1, When n > 1, Compare the gradients of each sample in the dataset with the latest globally estimated gradient obtained locally. The gradient projection is used to select the sample with the smallest gradient projection as the sample to be recycled and remove it from the dataset and place it in the recycling area. When the sample data in the recycling area reaches a preset upper limit, the data importance of each sample to be recycled is compared and the sample with the lowest importance is deleted from the local dataset. If the data importance of a sample to be recycled increases, it is removed from the recycling area and put back into the dataset. The data importance of each sample is the gradient projection of the sample, the sample gradient, and the globally estimated gradient. The result is obtained by weighted summation of the similarities.

[0023] Specifically, considering the limited storage space at the edge, and the continuous collection of new samples at the edge, making it difficult to store all samples, the purpose of data cleanup is to delete less important samples.

[0024] First, a periodic screening mechanism is used to select samples for reclaiming. This mechanism involves adding new samples to the local dataset and then filtering out those to be reclaimed. Specifically, this periodic screening can be performed at regular intervals or after a certain number of new samples have been acquired.

[0025] 1) Determine the gradient of each sample in the dataset on the latest local model g.

[0026] For old samples, the gradient of the old samples has been updated synchronously when the latest local model g was obtained in the most recent training. Therefore, when performing data cleanup, it is not necessary to recalculate the gradient of the old samples. It is only necessary to calculate the gradient of the new samples on the latest local model g. The new samples are those that did not participate in the local model training.

[0027] For the nth new sample Calculate its gradient on the latest obtained local model g. Based on gradient Update the local estimated gradient When n=1, When n > 1, Therefore, during the data preparation phase, the local estimated gradient is continuously updated based on new samples until the edge needs to participate in federated learning, at which point the update ends, yielding the updated local estimated gradient. During the federated learning phase, the data is uploaded to the central server for aggregation, and the local estimated gradient is reset to 0, which will serve as the initial value for updating the local estimated gradient when the next data processing operation is performed.

[0028] 2) After obtaining the gradients of each sample in the local dataset on the latest local model g, compare the gradients of each sample with the latest globally estimated gradient obtained locally. Gradient projection on For example, the samples in the dataset are sorted based on the magnitude of their gradient projections. The sample with the smallest gradient projection among all samples in the local dataset is marked as a sample to be recycled and removed from the local dataset and moved to the recycling area. The most recently acquired global estimated gradient is used as the basis for this process. This refers to the globally estimated gradient issued by the central server during the most recent training execution. The gradient projection of a sample gradient onto the globally estimated gradient is used as an evaluation metric to filter out samples for recycling. Since a larger projection value indicates that the gradient direction of the sample is more consistent with the model update direction, and the gradient value is larger, it means that the sample is more beneficial to model training updates. This allows us to select high-quality data samples that are more conducive to model training and keep them in the local training dataset. Low-quality samples with small projection values ​​(i.e., samples whose gradient direction differs significantly from the model update direction), or samples that have already been sufficiently trained, are placed in the recycling area.

[0029] In one embodiment, the sample with the smallest gradient projection is removed from the dataset and placed into the recycling area as the sample to be recycled. The number of samples to be recycled each time can be flexibly adjusted. Only one sample can be selected each time, or multiple samples can be selected.

[0030] Secondly, determine whether the sample data in the area to be recycled has reached the preset upper limit. If it has, perform a deletion operation to alleviate the pressure on local storage.

[0031] 1) Determine the importance of each sample in the area to be recovered.

[0032] The importance of each sample is determined by the gradient projection onto that sample, the sample gradient, and the globally estimated gradient. The similarity is obtained by weighted summation. In this invention, the gradient projection of the sample and the sample gradient and the globally estimated gradient are used. The weighted sum of similarities is used as an evaluation metric for sample importance. However, during gradual training, the projected values ​​of data that were once large and highly beneficial for model updates will gradually decrease. If the projected value is still used as the sole evaluation criterion for data samples, and data samples with small projected values ​​are deleted from the recycle bin, it may lead to data forgetting, meaning the model will forget data that was previously sufficiently trained. The sample gradients are then compared with the global gradient estimator. Using a weighted sum of similarities as the sample importance score can take into account the update direction of old samples, making samples that are consistent with the model update direction more important. Therefore, it is possible to reasonably evaluate the impact of each data sample in the recycling region on the current model update, retain data samples whose gradient direction is consistent with the current model update direction, which is more beneficial to model training, and delete samples that contribute little to model training.

[0033] In one embodiment, the data importance of sample x The calculation formula is: ; In the formula, This represents the gradient of sample x on the latest local model g. The latest globally estimated gradient obtained locally , Gradient In gradient The projection on Indicates the calculation of gradient and gradient cosine similarity, The weighting coefficients are set.

[0034] For samples in the pending recovery area whose latest data importance has already been calculated, there is no need to repeat the calculation.

[0035] For samples in the pending recovery area whose importance has not been calculated using the latest data, the importance needs to be recalculated. These samples are divided into two categories: One type is newly added samples, whose data importance has never been calculated. For this type of sample, we need to calculate its gradient projection and the latest global gradient estimate. Calculate the importance of its data; Another type is based on the projected gradient and the old global estimated gradient in the early stage. For samples whose data importance has been calculated, the projected gradient and the globally estimated gradient of these samples in this round are considered. All these processes have been updated, so their data importance may also change, thus requiring recalculation. For example, the edge component did not participate in federated learning in round 3, participated in federated learning in round 4, and did not participate in federated learning in round 5; in round 3, the importance of a sample to be retrieved had already been calculated, but in round 4, the local model g and the globally estimated gradient of that edge component have changed. All of these have been updated, so their importance will change in the 5th round, and therefore their importance needs to be recalculated.

[0036] 2) Based on data importance, delete unimportant samples to be recycled.

[0037] After obtaining the latest data importance of each sample in the reclaim area, unimportant samples are identified based on this data importance and completely deleted from the local storage. For example, each sample to be reclaimed in the reclaim area is sorted according to its importance, and one or more samples with the lowest importance are completely deleted from the local storage to alleviate local storage pressure.

[0038] 3) Based on the changes in data importance, remove the samples to be recycled that have increased importance from the recycling area and add them back to the local dataset.

[0039] For old samples in the uncollected region, if their data importance has increased compared to their previous importance, they are removed from the uncollected region and added back to the local dataset as training samples for the next local model training. In this invention, importance is determined by the gradient projection onto the sample and the sample gradient versus the globally estimated gradient. The result is obtained by weighted summation of similarity scores. Therefore, when the importance of a sample increases, it indicates that the latest update direction of the model differs from the gradient direction of the sample, which is when the data forgetting problem occurs. Putting the data with increased importance back into the local training dataset allows these forgotten data samples to participate in the model's training again. This allows the model to continue training on high-quality samples that were previously trained but are now forgotten, thereby alleviating the data forgetting problem.

[0040] Secondly, the present invention also relates to an edge device that participates in cloud-edge federated learning.

[0041] The device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0042] The device can be a desktop computer, laptop, handheld computer, or other computing device. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory can be used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory.

[0043] Thirdly, this invention also relates to a cloud-edge federated learning system.

[0044] like Figure 2 The diagram shown is an architecture diagram of a cloud-edge federated learning system according to an embodiment of the present invention. The cloud-edge federated learning system includes a central server and edge devices.

[0045] The central server is used to aggregate the local models uploaded by each edge device that participated in the previous round of federated learning to obtain a global model in each round of federated learning. The global estimated gradient is obtained by aggregating the local estimated gradients uploaded by each edge node participating in the previous round of federated learning. and the global model and global gradient estimation It will be distributed to the edge devices participating in the current round of federated learning.

[0046] Specifically, at the beginning of each training round, the central server aggregates all local gradients and local gradient estimators collected in the previous training round to obtain the global gradient G and the global gradient estimator. Distribute to edge clients participating in training.

[0047] Specifically, the aggregation formula used by the central server for aggregation is as follows: ; ; In the formula, This refers to the set of edge devices participating in federated learning in the current round. For set The local model uploaded by the i-th edge device. For set The locally estimated gradient uploaded from the i-th edge.

[0048] The edge device mentioned above is used to select and execute corresponding operations based on whether the edge device participates in this round of federated learning. If the edge device participates in this round of federated learning, a model update operation is performed; if the edge device does not participate in this round of federated learning, a data processing operation is performed.

[0049] In this invention, under a cloud-edge collaborative architecture, the contribution of each data sample to the global model update is evaluated by utilizing the projection values ​​of local data samples onto the global gradient direction, thereby providing real-time data storage decision support for edge devices with limited storage resources. This method not only accelerates the convergence process of federated learning models in a cloud-edge fusion environment but also improves the accuracy of the final model. Furthermore, this invention introduces gradient direction information during the data selection process, prioritizing the retention of data samples that provide broad gradient direction coverage. This ensures that the data stored in local memory effectively represents discarded samples, thereby correcting the model update direction and mitigating the data forgetting problem in streaming data scenarios within cloud-edge collaboration. In summary, this invention can achieve a collaborative solution to online data selection and data forgetting problems in a streaming data federated learning environment under a cloud-edge fusion system, combined with the limited storage space of edge clients, significantly improving the overall performance of federated learning models while reducing storage pressure on the edge.

[0050] The technical features of the embodiments described above can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that the terms "in one embodiment," "for example," and "again" in this invention are intended to illustrate the invention and are not intended to limit the invention.

[0051] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. An edge-end manipulation processing method for participating in cloud-edge federated learning, characterized in that, include: The corresponding operation will be executed depending on whether the edge device participates in this round of federated learning. If the edge device participates in this round of federated learning, a model update operation is performed, including: obtaining the global model distributed by the central server. and global gradient estimation Using old samples for the global model Training is performed to obtain an updated local model g, and the gradients of each old sample are updated to the gradients of each old sample on the updated local model g. The updated local model g and the updated local estimated gradients are uploaded to the central server for aggregation, and the local estimated gradients are reset to the initial gradients. Here, the old samples are the samples in the local dataset. If the edge does not participate in this round of federated learning, then data preparation operations are performed, including: periodically screening samples to be retrieved based on a screening mechanism, wherein the screening mechanism includes: periodically adding new samples to the dataset and calculating the performance of each new sample added. Gradient on the newly obtained local model g Based on gradient Update the local estimated gradient , where n is the index of the new sample. This represents the nth new sample updated after the local gradient estimation is reset. When n=1, When n > 1, Compare the gradients of each sample in the dataset with the latest globally estimated gradient obtained locally. The gradient projection is used to select the sample with the smallest gradient projection as the sample to be recycled and remove it from the dataset and place it in the recycling area. When the sample data in the recycling area reaches a preset upper limit, the data importance of each sample to be recycled is compared and the sample with the lowest importance is deleted from the local dataset. If the data importance of a sample to be recycled increases, it is removed from the recycling area and put back into the dataset. The data importance of each sample is the gradient projection of the sample, the sample gradient, and the globally estimated gradient. The result is obtained by weighted summation of the similarities.

2. The edge-end manipulation processing method as described in claim 1, characterized in that, Data importance of sample x The calculation formula is: ； In the formula, This represents the gradient of sample x on the latest local model g. The latest globally estimated gradient obtained locally , Gradient In gradient The projected length on, Indicates the calculation of gradient and gradient cosine similarity, The weighting coefficients are set.

3. The edge-end manipulation processing method as described in claim 1, characterized in that, Reset the local estimated gradient to the initial gradient, specifically by resetting the local estimated gradient to 0.

4. The edge-end manipulation processing method as described in claim 1, characterized in that, Compare the gradients of each sample in the dataset with the latest globally estimated gradient obtained locally. The gradient projection on the dataset is used to select the sample with the smallest gradient projection as the sample to be recycled. Specifically, the samples in the dataset are sorted according to the size of their gradient projections, and one or more samples with the smallest gradient projections are selected as the samples to be recycled.

5. The edge-end manipulation processing method as described in claim 1, characterized in that, The data importance of each sample to be recycled is compared and the sample with the lowest importance is deleted from the local database. Specifically, each sample to be recycled in the recycling area is sorted according to its importance, and one or more samples with the lowest importance are completely deleted from the local database.

6. An edge device participating in cloud-edge federated learning, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 5.

7. A cloud-edge federated learning system, characterized in that, This includes central servers and edge devices; The central server is used to aggregate the local models uploaded by each edge device that participated in the previous round of federated learning to obtain a global model in each round of federated learning. The global estimated gradient is obtained by aggregating the local estimated gradients uploaded by each edge node participating in the previous round of federated learning. and the global model and global gradient estimation Distribute to the edge devices participating in the current round of federated learning; The edge device is the edge device as described in claim 6, used to select and execute corresponding operations based on whether the edge device participates in this round of federated learning. If the edge device participates in this round of federated learning, a model update operation is executed; if the edge device does not participate in this round of federated learning, a data processing operation is executed.

8. The cloud-edge federated learning system as described in claim 7, characterized in that, The aggregation formula used by the central server for aggregation is: ； ； In the formula, This refers to the set of edge devices participating in federated learning in the current round. For set The local model uploaded by the i-th edge device. For set The locally estimated gradient uploaded from the i-th edge.

Images