Federal incremental learning framework and method for edge intelligence

Abstract

The invention discloses a federal incremental learning framework and method for edge intelligence. The framework comprises a central cloud server, a plurality of edge servers and a plurality of edge intelligent devices, the central cloud server is connected with the plurality of edge servers, one edge server is connected with the plurality of edge intelligent devices, and the central cloud server comprises a first initial neural network model and a second initial neural network model. The method solves the problems that the performance and resource requirements of edge equipment are too high, efficient operation on the edge equipment with low calculation and storage resources cannot be achieved, catastrophic forgetting exists in a class incremental learning method, when a model receives new class data, the performance of the model on an old class is sharply reduced, effective protection on data privacy is lacked, and the data privacy cannot be effectively protected. Especially in federal learning, the model is easy to leak sensitive information of local data due to non-uniform data distribution on different edge devices.

Description

Technical Field

[0001] The present invention relates to the field of edge computing technology, and in particular to a federated incremental learning framework and method for edge intelligence. Background Art

[0002] With the development of deep learning, artificial intelligence has received widespread attention in the field of edge computing. However, the characteristics of edge devices bring two major challenges: the contradiction between high precision and strong data privacy requirements, and the contradiction between high computing resource consumption and low performance of edge devices. Although the existing federated learning (FL) technology provides solutions in terms of privacy protection and communication costs, most solutions assume that the data distribution is fixed and known. In practical applications, this assumption is too idealistic and cannot cope with the ever-changing data streams on edge devices. Summary of the invention

[0003] The embodiments of the present invention provide a federated incremental learning framework and method for edge intelligence, so as to at least solve the problems that the performance and resource requirements of edge devices are too high and they cannot run efficiently on edge devices with low computing and storage resources. The incremental learning method has the problem of catastrophic forgetting. When the model receives new category data, its performance on old categories will drop sharply. There is a lack of effective protection for data privacy. Especially in federated learning, due to the uneven distribution of data on different edge devices, the model is prone to leak sensitive information of local data.

[0004] According to one aspect of an embodiment of the present invention, a federated incremental learning framework for edge intelligence is provided. The framework may include: a central cloud server, multiple edge servers and multiple edge intelligent devices, wherein the central cloud server is connected to multiple edge servers, and one edge server is connected to multiple edge intelligent devices, wherein the central cloud server includes a first initial neural network model and a second initial neural network model; the central cloud server is used to train the first initial neural network model through an initial data set to obtain a trained initial feature extractor, weights of the initial feature extractor, an initial classifier and weights of the initial classifier, obtain an initialized feature autocorrelation matrix corresponding to the initial data set, add an expansion layer after the initial feature extractor, replace the initial classifier with a parsing classifier, obtain a second initial neural network model, train the second initial neural network model through the initial data set to obtain a trained target feature extractor, weights of the target feature extractor, parsing classifier and weights of the parsing classifier, and send the target feature extractor, weights of the target feature extractor, parsing classifier and weights of the parsing classifier to multiple edge servers; the edge intelligent device is used to send a plurality of images of multiple target time periods collected to the corresponding edge server; the edge server is used to receive the first edge intelligent device sent by the first edge intelligent device. A plurality of images of a target time period are extracted by a target feature extractor, and the extracted features are learned by analytical incremental learning to update the analytical classifier. The process of analytical incremental learning of the extracted features is to update the initialization feature autocorrelation matrix, and the weight of the analytical classifier is updated based on the updated initialization feature autocorrelation matrix using a recursive formula and the number of the plurality of images of the first target time period is recorded until the plurality of images of the last target time period of the last edge intelligent device corresponding to the edge server are received, and the weight of the analytical classifier of the last target time period of the last edge intelligent device and the total number of images of each edge intelligent device are obtained, and uploaded to the central cloud server; the central cloud server is also used to generate the weight of the global analytical classifier by the weight of the analytical classifier of the last target time period of each edge server and the total number of images by the weighted average method, and send the weight of the global analytical classifier to multiple edge servers, so that when each edge server processes the plurality of images of a target time period of multiple target time periods obtained in the next round, the weight of the analytical classifier of the last target time period of the previous round is replaced by the weight of the global analytical classifier.

[0005] Optionally, the expression for obtaining the initialization characteristic autocorrelation matrix corresponding to the initial data set is:

[0006] in, is the number of data participating in the p-1 stage, is the high-dimensional feature of the i-th data in the p-1 stage, is the transposition of the high-dimensional features of the i-th data in the p-1 stage, represents the regularization coefficient, is the initialization feature autocorrelation matrix, I is the unit matrix, the p-1 stage is the stage corresponding to the first initial neural network model, and (0) is used to characterize the stage corresponding to the initial data set as 0.

[0007] Optionally, the expression of the weight of the parsing classifier is:

[0008] in, is the high-dimensional feature of the initial data set, is the transpose of the high-dimensional features of the initial data set, -1 is the matrix inversion operation, is the data label of the initial data set, is the weight of the parsing classifier corresponding to the initial dataset.

[0009] Optionally, the objective loss function of the second initial neural network model is expressed as:

[0010] Among them, FNC is the second initial convolutional neural network, F is the F norm, is the regularization term, is the initial loss function of the second initial neural network model.

[0011] Optionally, the expression for updating the initialization characteristic autocorrelation matrix is:

[0012] in, For The updated initialized feature autocorrelation matrix corresponding to the stage, For The updated initialized feature autocorrelation matrix corresponding to the stage, is the transposition of the high-dimensional features of the i-th data in stage p, is the high-dimensional feature of the i-th data in stage p, where the P-th stage corresponds to the stage in which the edge server receives several images of a target time period sent by the edge smart device.

[0013] Optionally, the expression for updating the weight of the parsing classifier is:

[0014] in, is the weight of the parsing classifier after updating at the Pth order, is the weight of the parsing classifier after updating in the P-1 phase, is the transposition of the high-dimensional features of several images of a target time period obtained in stage p, is the high-dimensional features of several images of a target time period obtained in stage p, is the label of several images of a target time period obtained in stage p, is the weight of the parsing classifier after updating in the P-1 stage.

[0015] According to one aspect of an embodiment of the present invention, a federated incremental learning method for edge intelligence is provided. The method may include: constructing a federated incremental learning framework for edge intelligence, the framework including a central cloud server, multiple edge servers and multiple edge intelligent devices, the central cloud server is connected to multiple edge servers, and one edge server is connected to multiple edge intelligent devices, wherein the central cloud server includes a first initial neural network model and a second initial neural network model; the central cloud server is used to train the first initial neural network model through an initial data set to obtain a trained initial feature extractor, a weight of the initial feature extractor, an initial classifier and a weight of the initial classifier, obtain an initialized feature autocorrelation matrix corresponding to the initial data set, add an expansion layer after the initial feature extractor, replace the initial classifier with a parsing classifier, obtain a second initial neural network model, train the second initial neural network model through the initial data set to obtain a trained target feature extractor, a weight of the target feature extractor, a parsing classifier and a weight of the parsing classifier, and send the target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier to multiple edge servers; the edge intelligent device is used to send a plurality of images of multiple target time periods collected to the corresponding edge server; the edge server is used to receive the first edge intelligent The method comprises the following steps: receiving a plurality of images of the first target time period sent by the intelligent device, extracting features of the images by a target feature extractor, and updating the analytical classifier after learning the extracted features by analytical incremental learning, wherein the process of learning the extracted features by analytical incremental learning is to update the initialization feature autocorrelation matrix, and updating the weight of the analytical classifier by using a recursive formula based on the updated initialization feature autocorrelation matrix and recording the number of the plurality of images of a target time period, until the plurality of images of the last target time period of the last edge intelligent device corresponding to the edge server are received, and obtaining the weight of the analytical classifier of the last target time period of the last edge intelligent device and the total number of images of each edge intelligent device, and uploading them to the central cloud server; the central cloud server is also used to generate the weight of the global analytical classifier by using the weighted average method of the analytical classifier of the last target time period of each edge server and the total number of images, and sending the weight of the global analytical classifier to multiple edge servers, so that when each edge server processes the plurality of images of a target time period of multiple target time periods obtained in the next round, the weight of the analytical classifier of the last target time period of the previous round is replaced by the weight of the global analytical classifier.

[0016] Beneficial effects of the present invention: 1. Continuous update capability: The framework supports continuous updating of models and is suitable for edge devices with constantly changing data distribution.

[0017] 2. Low resource consumption: Through local parsing incremental learning, the computing and storage requirements of edge devices are reduced, which is suitable for resource-constrained devices such as drones and wearable devices.

[0018] 3. Privacy protection mechanism: The framework achieves data privacy protection by relatively forgetting incremental data on edge devices, saving storage space while avoiding the risk of data leakage. It is particularly suitable for scenarios with high data security requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 It is a schematic diagram of a federated incremental learning framework for edge intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only embodiments of a part of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0021] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0022] Example 1 According to an embodiment of the present invention, a federated incremental learning framework for edge intelligence is provided. It should be noted that the steps shown in the schematic diagrams of the accompanying drawings can be executed in a computer system comprising at least one set of computer executable instructions, and although a logical order is shown in the schematic diagrams, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0023] Figure 1 is a schematic diagram of a federated incremental learning framework for edge intelligence according to an embodiment of the present invention, such as Figure 1 As shown, the framework may include: a central cloud server, multiple edge servers and multiple edge intelligent devices, the central cloud server is connected to multiple edge servers, and one edge server is connected to multiple edge intelligent devices, wherein the central cloud server includes a first initial neural network model and a second initial neural network model, and the framework may include the following steps: Step S101, the central cloud server is used to train the first initial neural network model through the initial data set to obtain the trained initial feature extractor, the weight of the initial feature extractor, the initial classifier and the weight of the initial classifier, obtain the initialized feature autocorrelation matrix corresponding to the initial data set, add an extension layer after the initial feature extractor, replace the initial classifier with the parsing classifier, obtain the second initial neural network model, train the second initial neural network model through the initial data set to obtain the trained target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier, and send the target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier to multiple edge servers.

[0024] In the technical solution provided in step S101 of the present invention, if Figure 1 As shown, the central cloud server uses an initial data set (large basic data set) to train the first initial neural network model, wherein the first initial convolutional neural network model includes an initial feature extractor, an initial classifier (basic parsing classifier), and the basic parsing classifier is trained by a parsing incremental learning algorithm. After the training, the trained initial feature extractor, the weight of the initial feature extractor, the initial classifier and the weight of the initial classifier are obtained, and the initialized feature autocorrelation matrix corresponding to the initial data set is obtained. An extension layer is added after the initial feature extractor of the first convolutional neural network model, and the initial classifier is replaced by the parsing classifier to obtain a second initial neural network model. The second initial neural network model is trained by the initial data set to obtain a trained target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier. The target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier are sent to each edge server, for example, Figure 1 There is a central server and multiple edge servers.

[0025] Step S102: The edge intelligent device is used to send the collected images of multiple target time periods to the corresponding edge server.

[0026] In the technical solution provided in the above step S102 of the present invention, the edge intelligent device sends the collected images of multiple target time periods to the corresponding edge server, and the 10 edge intelligent devices corresponding to the edge server 1; for example, the target time period can be every half hour, the first batch of data includes 10 target time periods, one target time period includes 10 half hours, the first half hour of the first target time period includes 20 images, the second half hour includes 15 images... the tenth half hour includes 16 images; the first edge intelligent device sends the first batch of data to the edge server 1, and the other edge intelligent devices corresponding to the edge server 1 obtain data in the same way as the first edge intelligent device, and send the obtained data to the edge server 1.

[0027] Step S103, the edge server is used to receive several images of the first target time period sent by the first edge smart device, extract the features of the image through the target feature extractor, and update the parsing classifier after learning the extracted features through parsing incremental learning. The process of parsing the extracted features through parsing incremental learning is to update the initialization feature autocorrelation matrix, and based on the updated initialization feature autocorrelation matrix, the weight of the parsing classifier is updated using a recursive formula and the number of several images of the first target time period is recorded until the last target time period of the last edge smart device corresponding to the edge server is received, and the weight of the parsing classifier of the last target time period of the last edge smart device and the total number of images of each edge smart device are obtained, and uploaded to the central cloud server.

[0028] In the technical solution provided in the above step S103 of the present invention, for example, the edge server 1 receives 20 images of the first target time period sent by the first edge smart device corresponding to the edge server, extracts features of the 20 images through a target feature extractor, and updates the parsing classifier after learning the extracted features through parsing incremental learning. The process of parsing the extracted features through parsing incremental learning is to update the initialization feature autocorrelation matrix, and the weight of the parsing classifier is updated based on the updated initialization feature autocorrelation matrix using a recursive formula and the number of 20 images of the first target time period (20) is recorded until the last target time period of the last edge smart device (the 10th) corresponding to the edge server is received. Several images (16 images) of the last target time period of the last edge smart device (the 10th) are obtained, and the weight of the parsing classifier of the last target time period of the last edge smart device and the total number of images of each edge smart device (that is, the total number of images of 10 target time periods) are obtained, and uploaded to the central cloud server.

[0029] Step S104, the central cloud server is also used to generate the weight of the global parsing classifier by weighted averaging the weight of the parsing classifier of the last target time period of each edge server and the total number of images, and send the weight of the global parsing classifier to multiple edge servers, so that when each edge server processes several images of a target time period of multiple target time periods obtained in the next round, the weight of the parsing classifier of the last target time period of the previous round is replaced by the weight of the global parsing classifier.

[0030] In the technical solution provided in the above step S104 of the present invention, the central cloud server generates the weight of the global parsing classifier by weighted averaging the weight of the parsing classifier received from each edge server in the last target time period and the total number of images (the sum of the total number of images obtained by each edge server) .

[0031] The weight of the parsing classifier of each edge server in the last target time period and the total number of images are weighted averaged to generate the expression of the weight of the global parsing classifier:

[0032] in, is the weight of the global parsing classifier, , represents the amount of data of edge server i in stage p-1. Here, stage p-1 is the stage before stage p. Since the cycle is always in progress, stage p-1 can be the stage corresponding to the initial data set, or it can be the stage before stage p in the cycle process. That is, the image currently receiving a target time period can be stage p. Indicates the edge server in the p-1 phase The parsing classifier weight of , in this embodiment of the present invention, the value range of P starts from 1.

[0033] The above method of this embodiment is further introduced below.

[0034] As an optional implementation manner, in step S101, the expression for obtaining the initial characteristic autocorrelation matrix corresponding to the initial data set is:

[0035] in, is the number of data participating in the p-1 stage, is the high-dimensional feature of the i-th data in the p-1 stage, is the transposition of the high-dimensional features of the i-th data in the p-1 stage, represents the regularization coefficient, is the initialization feature autocorrelation matrix, I is the unit matrix, the p-1 stage is the stage corresponding to the first initial neural network model, and (0) is used to characterize the stage corresponding to the initial data set as 0.

[0036] In this embodiment, p is 1. When p is 1, is 0, which is the stage corresponding to the initial data set, and p is 1, which is the stage of the image of the first target time period of the first edge smart device corresponding to edge server 1 obtained by edge server 1.

[0037] As an optional embodiment, in step S101, the expression of the weight of the parsing classifier is:

[0038] in, is the high-dimensional feature of the initial data set, is the transpose of the high-dimensional features of the initial data set, -1 is the matrix inversion operation, is the data label of the initial data set, is the weight of the parsing classifier corresponding to the initial dataset.

[0039] In this embodiment, according to the above formula, the weight corresponding to the initial data set in the second convolutional neural network is obtained.

[0040] As an optional implementation manner, in step S103, the expression of the target loss function of the second initial neural network model is:

[0041] Among them, FNC is the second initial convolutional neural network, F is the F norm, is the regularization term, is the initial loss function of the second initial neural network model.

[0042] In this embodiment, the target loss function of the second initial convolutional neural network is determined by the regularization term and the initial loss function.

[0043] As an optional implementation mode, in step S103, the expression for updating the initialization characteristic autocorrelation matrix is:

[0044] in, For The updated initialization feature autocorrelation matrix corresponding to the stage is: For The updated initialization feature autocorrelation matrix corresponding to the stage is: is the transposition of the high-dimensional features of the i-th data in stage p, is the high-dimensional feature of the i-th data in the p-th stage, where the P-th stage is the stage corresponding to the edge server receiving several images of a target time period sent by the edge intelligent device.

[0045] In this embodiment, by The updated initial feature autocorrelation matrix of the stage is updated The stage corresponds to the initialization feature autocorrelation matrix. When edge server 1 obtains the stage of the image of the second target time period of the first edge intelligent device corresponding to edge server 1, stage P is the stage of the image of the second target time period of the first edge intelligent device corresponding to edge server 1 obtained by edge server 1, and stage p-1 is the stage of the image of the first target time period of the first edge intelligent device corresponding to edge server 1 obtained by edge server 1, and the cycle is repeated.

[0046] As an optional embodiment, in step S103, the expression for updating the weight of the parsing classifier is:

[0047] in, is the weight of the parsing classifier after updating at the Pth order, is the weight of the parsing classifier after updating in the P-1 phase, is the transposition of the high-dimensional features of several images of a target time period obtained in stage p, is the high-dimensional features of several images of a target time period obtained in stage p, is the label of several images of a target time period obtained in stage p, is the weight of the parsing classifier after updating in the P-1 stage.

[0048] In this embodiment, by The weights of the updated parsing classifier are updated The weight of the parsing classifier corresponding to the stage, such as Figure 1 As shown, the parsing classifiers corresponding to the first edge service device are 11 to parsing classifier 1i, and each parsing classifier corresponds to a parsing classifier weight.

[0049] In an embodiment of the present invention, a central cloud server is used to train a first initial neural network model through an initial data set to obtain a trained initial feature extractor, a weight of the initial feature extractor, an initial classifier and a weight of the initial classifier, obtain an initialized feature autocorrelation matrix corresponding to the initial data set, add an expansion layer after the initial feature extractor, replace the initial classifier with a parsing classifier, obtain a second initial neural network model, train the second initial neural network model through the initial data set to obtain a trained target feature extractor, a weight of the target feature extractor, a parsing classifier and a weight of the parsing classifier, and send the target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier to multiple edge servers; an edge intelligent device is used to send a plurality of images of multiple target time periods collected The image is sent to the corresponding edge server; the edge server is used to receive several images of the first target time period sent by the first edge intelligent device, extract the features of the image through the target feature extractor, and update the parsing classifier after learning the extracted features through parsing incremental learning, and the process of parsing incremental learning is to update the initialization feature autocorrelation matrix, and use the recursive formula based on the updated initialization feature autocorrelation matrix to update the weight of the parsing classifier and record the number of several images of the first target time period until the last target time period of the last edge intelligent device corresponding to the edge server is received, and the weight of the parsing classifier of the last target time period of the last edge intelligent device and the total number of images of each edge intelligent device are obtained, and uploaded to the central cloud server;The central cloud server is also used to generate the weight of the global parsing classifier by weighted averaging the weight of the parsing classifier of the last target time period of each edge server and the total number of images, and send the weight of the global parsing classifier to multiple edge servers, so that each edge server replaces the weight of the parsing classifier of the last target time period of the previous round with the weight of the global parsing classifier when processing several images of a target time period of multiple target time periods obtained in the next round, which solves the problem that the performance and resource requirements of edge devices are too high and cannot be efficiently run on edge devices with low computing and storage resources. The incremental learning method has a catastrophic forgetting problem, and the model will be lost when receiving a new number of categories. When the data is updated, its performance on the old categories will drop sharply, lacking effective protection of data privacy, especially in federated learning, due to the uneven distribution of data on different edge devices, the model is prone to leak sensitive information of local data. The framework in the embodiment of the present invention supports the continuous update of the model, which is suitable for edge devices with constantly changing data distribution. Through local parsing incremental learning, the computing and storage requirements of edge devices are reduced, which is suitable for resource-constrained devices. The framework achieves data privacy protection by relatively forgetting incremental data on edge devices, saving storage space while avoiding the risk of data leakage, and is particularly suitable for scenarios with high data security requirements. ;

[0050] Example 2 According to an embodiment of the present invention, a federated incremental learning method for edge intelligence is also provided. It should be noted that the federated incremental learning method for edge intelligence can be used to execute a federated incremental learning framework for edge intelligence in Example 1.

[0051] A federated incremental learning method for edge intelligence includes: building a federated incremental learning framework for edge intelligence, the framework including a central cloud server, multiple edge servers and multiple edge intelligent devices, the central cloud server is connected to multiple edge servers, and one edge server is connected to multiple edge intelligent devices, wherein the central cloud server includes a first initial neural network model and a second initial neural network model.

[0052] The central cloud server is used to train the first initial neural network model through the initial data set to obtain the trained initial feature extractor, the weight of the initial feature extractor, the initial classifier and the weight of the initial classifier, obtain the initialized feature autocorrelation matrix corresponding to the initial data set, add an extension layer after the initial feature extractor, replace the initial classifier with the parsing classifier, and obtain the second initial neural network model, train the second initial neural network model through the initial data set to obtain the trained target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier, and send the target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier to multiple edge servers.

[0053] The edge intelligent device is used to send a number of images collected in multiple target time periods to the corresponding edge server.

[0054] The edge server is used to receive several images of the first target time period sent by the first edge intelligent device, extract the features of the images through the target feature extractor, and update the parsing classifier after learning the extracted features through parsing incremental learning, wherein the process of parsing the extracted features through parsing incremental learning is to update the initialization feature autocorrelation matrix, and based on the updated initialization feature autocorrelation matrix, the weight of the parsing classifier is updated using a recursive formula and the number of several images of a target time period is recorded, until the last target time period of the last edge intelligent device corresponding to the edge server is received, the weight of the parsing classifier of the last target time period of the last edge intelligent device and the total number of images of each edge intelligent device are obtained, and uploaded to the central cloud server.

[0055] The central cloud server is also used to generate the weight of the global parsing classifier by weighted averaging the weight of the parsing classifier of the last target time period of each edge server and the total number of images, and send the weight of the global parsing classifier to multiple edge servers, so that when each edge server processes several images of a target time period of multiple target time periods obtained in the next round, the weight of the parsing classifier of the last target time period of the previous round is replaced by the weight of the global parsing classifier.

[0056] In this embodiment, the central cloud server is used to train the first initial neural network model through the initial data set to obtain the trained initial feature extractor, the weight of the initial feature extractor, the initial classifier and the weight of the initial classifier, obtain the initialized feature autocorrelation matrix corresponding to the initial data set, add an expansion layer after the initial feature extractor, replace the initial classifier with the parsing classifier, obtain the second initial neural network model, train the second initial neural network model through the initial data set to obtain the trained target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier, send the target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier to multiple edge servers; the edge intelligent device is used to send a plurality of images collected for multiple target time periods to the edge servers; The edge server is used to receive several images of the first target time period sent by the first edge intelligent device, extract the features of the image through the target feature extractor, and update the parsing classifier after learning the extracted features through parsing incremental learning. The process of parsing incremental learning is to update the initialization feature autocorrelation matrix, and the weight of the parsing classifier is updated based on the updated initialization feature autocorrelation matrix using a recursive formula and the number of several images of the first target time period is recorded until the last target time period of the last edge intelligent device corresponding to the edge server is received, and the weight of the parsing classifier of the last target time period of the last edge intelligent device and the total number of images of each edge intelligent device are obtained, and uploaded to the central cloud server;The central cloud server is also used to generate the weight of the global parsing classifier by weighted averaging the weight of the parsing classifier of the last target time period of each edge server and the total number of images, and send the weight of the global parsing classifier to multiple edge servers, so that each edge server replaces the weight of the parsing classifier of the last target time period of the previous round with the weight of the global parsing classifier when processing several images of a target time period of multiple target time periods obtained in the next round, which solves the problem that the performance and resource requirements of edge devices are too high and cannot be efficiently run on edge devices with low computing and storage resources. The incremental learning method has a catastrophic forgetting problem, and the model will be lost when receiving a new number of categories. When the data is updated, its performance on the old categories will drop sharply, lacking effective protection of data privacy, especially in federated learning, due to the uneven distribution of data on different edge devices, the model is prone to leak sensitive information of local data. The framework in the embodiment of the present invention supports the continuous update of the model, which is suitable for edge devices with constantly changing data distribution. Through local parsing incremental learning, the computing and storage requirements of edge devices are reduced, which is suitable for resource-constrained devices. The framework achieves data privacy protection by relatively forgetting incremental data on edge devices, saving storage space while avoiding the risk of data leakage, and is particularly suitable for scenarios with high data security requirements. ;

[0057] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0058] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0059] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0060] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0061] In addition, each functional unit in each embodiment of the present invention may be integrated into a first processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0062] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A federated incremental learning framework for edge intelligence, characterized in that: include: A central cloud server, multiple edge servers and multiple edge smart devices, the central cloud server is connected to the multiple edge servers, and one edge server is connected to the multiple edge smart devices, wherein the central cloud server includes a first initial neural network model and a second initial neural network model; The central cloud server is used to train the first initial neural network model through the initial data set to obtain the trained initial feature extractor, the weight of the initial feature extractor, the initial classifier and the weight of the initial classifier, obtain the initialized feature autocorrelation matrix corresponding to the initial data set, add an extension layer after the initial feature extractor, replace the initial classifier with the parsing classifier, obtain the second initial neural network model, train the second initial neural network model through the initial data set to obtain the trained target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier, and send the target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier to multiple edge servers; The edge intelligent device is used to send a plurality of images collected in a plurality of target time periods to a corresponding edge server; The edge server is used to receive several images of the first target time period sent by the first edge intelligent device, extract features of the image through the target feature extractor, and update the parsing classifier after learning the extracted features through parsing incremental learning, and the process of parsing the extracted features through parsing incremental learning is to update the initialization feature autocorrelation matrix, and use the recursive formula based on the updated initialization feature autocorrelation matrix to update the weight of the parsing classifier and record the number of several images of the first target time period, until the last target time period of the last edge intelligent device corresponding to the edge server is received, the weight of the parsing classifier of the last target time period of the last edge intelligent device and the total number of images of each edge intelligent device are obtained, and uploaded to the central cloud server; The central cloud server is also used to generate the weight of the global parsing classifier by weighted averaging the weight of the parsing classifier of the last target time period of each edge server and the total number of images, and send the weight of the global parsing classifier to multiple edge servers, so that when each edge server processes several images of a target time period of multiple target time periods obtained in the next round, the weight of the parsing classifier of the last target time period of the previous round is replaced by the weight of the global parsing classifier.

2. The method according to claim 1, characterized in that The expression for obtaining the initial characteristic autocorrelation matrix corresponding to the initial data set is: in, is the number of data participating in the p-1 stage, is the high-dimensional feature of the i-th data in the p-1 stage, is the transposition of the high-dimensional features of the i-th data in the p-1 stage, represents the regularization coefficient, is the initialization feature autocorrelation matrix, I is the unit matrix, the p-1 stage is the stage corresponding to the first initial neural network model, and (0) is used to characterize the stage corresponding to the initial data set as 0.

3. The method according to claim 1, characterized in that The expression of the weight of the parsing classifier is: in, is the high-dimensional feature of the initial data set, is the transpose of the high-dimensional features of the initial data set, -1 is the matrix inversion operation, is the data label of the initial data set, is the weight of the parsing classifier corresponding to the initial dataset.

4. The method according to claim 3, characterized in that The expression of the objective loss function of the second initial neural network model is: Among them, FNC is the second initial convolutional neural network, F is the F norm, is the regularization term, is the initial loss function of the second initial neural network model.

5. The method according to claim 1, characterized in that The expression for updating the initialization characteristic autocorrelation matrix is: in, For The updated initialized feature autocorrelation matrix corresponding to the stage, For The updated initialized feature autocorrelation matrix corresponding to the stage, is the transposition of the high-dimensional features of the i-th data in stage p, is the high-dimensional feature of the i-th data in the p-th stage, where the P-th stage is the stage corresponding to the edge server receiving several images of a target time period sent by the edge intelligent device.

6. The method according to claim 5, characterized in that The expression for updating the weight of the parsing classifier is: in, is the weight of the parsing classifier after the P-th order update, is the weight of the parsing classifier after updating in the P-1 phase, is the transposition of the high-dimensional features of several images of a target time period obtained in stage p, is the high-dimensional features of several images of a target time period obtained in stage p, is the label of several images of a target time period obtained in stage p, is the weight of the parsing classifier after updating in the P-1 stage.

7. A federated incremental learning method for edge intelligence, characterized in that: include: Constructing a federated incremental learning framework for edge intelligence, the framework comprising a central cloud server, multiple edge servers and multiple edge intelligent devices, the central cloud server is connected to multiple edge servers, and one edge server is connected to multiple edge intelligent devices, wherein the central cloud server comprises a first initial neural network model and a second initial neural network model; The central cloud server is used to train the first initial neural network model through the initial data set to obtain the trained initial feature extractor, the weight of the initial feature extractor, the initial classifier and the weight of the initial classifier, obtain the initialized feature autocorrelation matrix corresponding to the initial data set, add an extension layer after the initial feature extractor, replace the initial classifier with the parsing classifier, obtain the second initial neural network model, train the second initial neural network model through the initial data set to obtain the trained target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier, and send the target feature extractor, the weight of the target feature extractor, the parsing classifier and the weight of the parsing classifier to multiple edge servers; The edge intelligent device is used to send a plurality of images collected in a plurality of target time periods to a corresponding edge server; The edge server is used to receive several images of the first target time period sent by the first edge smart device, extract features of the images through a target feature extractor, and update the parsing classifier after learning the extracted features through parsing incremental learning, wherein the process of parsing the extracted features through parsing incremental learning is to update the initialization feature autocorrelation matrix, and based on the updated initialization feature autocorrelation matrix, the weight of the parsing classifier is updated using a recursive formula and the number of several images of a target time period is recorded, until the last target time period of the last edge smart device corresponding to the edge server is received, the weight of the parsing classifier of the last target time period of the last edge smart device and the total number of images of each edge smart device are obtained, and uploaded to the central cloud server; The central cloud server is also used to generate the weight of the global parsing classifier by weighted averaging the weight of the parsing classifier of the last target time period of each edge server and the total number of images, and send the weight of the global parsing classifier to multiple edge servers, so that when each edge server processes several images of a target time period of multiple target time periods obtained in the next round, the weight of the parsing classifier of the last target time period of the previous round is replaced by the weight of the global parsing classifier.

8. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the apparatus of claim 1.

9. A computer-readable storage medium, characterized in that Computer executable instructions are stored, and when the instructions are executed, they are used to implement the device of claim 1.

10. A computer program product, characterized in that The invention comprises computer executable instructions, which are used to implement the apparatus of claim 1 when being executed.

Images