Method and apparatus for training an image recognition model

Abstract

Embodiments of the present specification provide a method and apparatus for training an image recognition model. At each first member device, frequency domain transformation based data desensitization processing is performed on training sample image data to obtain first desensitized image data; the first desensitized image data is provided to a hyperparameter selection model to select first hyperparameters; the first desensitized image data is subjected to data enhancement based image mixing processing using the first hyperparameters to obtain second desensitized image data and labeled label data subjected to label mixing processing. Then, the image recognition model is trained using the second desensitized image data and the corresponding labeled label data subjected to label mixing processing. The second member device updates the image recognition model using the model training results received from each first member device.

Description

TECHNICAL FIELD

[0001] Embodiments of the present specification generally relate to the technical field of artificial intelligence, and in particular, to an image data processing method, a training method of an image recognition model, an image recognition method, and apparatuses thereof. BACKGROUND

[0002] Business processing schemes based on image recognition models have been widely applied in a large number of applications, such as face payment business based on a face recognition model. Image data related to data privacy information (for example, user privacy information, etc.) is usually distributed in different data owners or different regions and countries. In order to protect data privacy information, data sharing of these private data is not allowed between data owners or between different regions. However, in order to provide better services for users by using sufficient data, it is necessary to fully mine the information between data for image recognition model training for specific tasks. For this purpose, a federated learning method is proposed, which can use private image data of multiple data owners to train an image recognition model without data leaving the domain. However, in the federated learning method, each data owner needs to share gradient information or weight information with the model owner for aggregation after completing the image recognition model training locally, which may result in information leakage of the gradient information or the weight information. How to avoid reconstructing the original image data from the shared gradient information or weight information becomes a problem to be solved. SUMMARY

[0003] In view of the above, the embodiments of the present specification provide a training method and apparatus of an image recognition model. By using the training method and apparatus of the image recognition model, the image data is subjected to data desensitization processing based on frequency domain transformation to obtain desensitized image data, the obtained desensitized image data is subjected to image mixing processing based on data enhancement, and then the desensitized image data subjected to the image mixing processing is used to train the image recognition model locally, and the model update quantity of the image recognition model trained locally is shared with the model owner, which can avoid reconstructing the original image data from the shared model update quantity, thereby realizing data privacy protection of the original image data.

[0004] According to an aspect of the embodiments of the present specification, a method for training an image recognition model is provided, the method is performed by a first member device having local training data, and the method comprises: cyclically performing a model training process as follows until a model training end condition is met: obtaining current training sample image data and its labeled label data; performing frequency domain transformation-based data desensitization processing on the current training sample image data to obtain first desensitized image data of the current training sample image data; providing the first desensitized image data to a hyperparameter selection model to select a first hyperparameter for indicating the number of images participating in image mixing processing from a candidate hyperparameter set; using the first hyperparameter to perform Mixup data enhancement-based image mixing processing on the first desensitized image data to obtain second desensitized image data and corresponding label mixing processed labeled label data; training the current image recognition model using the second desensitized image data and the corresponding label mixing processed labeled label data; and providing a model training result of the current image recognition model to a second member device for maintaining an image recognition model, so that the second member device uses model training results from a plurality of first member devices to perform image recognition model updating, and receives the updated image recognition model from the second member device for the next round of image recognition model training.

[0005] Optionally, in one example of the above aspect, before providing the first desensitized image data to the hyperparameter selection model to select a first hyperparameter for indicating the number of images participating in image mixing processing from a candidate hyperparameter set, the method can further comprise: in response to satisfying a first predetermined condition, updating the hyperparameter selection model through a model updating process as follows: providing the first desensitized image data to the hyperparameter selection model to select a second hyperparameter for indicating the number of images participating in image mixing processing from the candidate hyperparameter set; using the second hyperparameter to perform data enhancement-based image mixing processing on the first desensitized image data to obtain third desensitized image data and corresponding label mixing processed labeled label data; providing the third desensitized image data to the current image recognition model to obtain second predicted label data of the third desensitized image data, and determining a first loss function according to the second predicted label data and the corresponding label mixing processed labeled label data; updating the current image recognition model according to the first loss function; providing the third desensitized image data to the updated current image recognition model to obtain third predicted label data of the third desensitized image data, and determining a second loss function according to the third predicted label data and the corresponding label mixing processed labeled label data; determining the third loss function based on the first loss function and the second loss function; and updating the model parameters of the hyperparameter selection model based on the third loss function.

[0006] Optionally, in one of the above aspects, the first predetermined condition can include that a round number interval between a current training round number of the image recognition model and a training round number at a last update processing time of the hyperparameter selection model reaches a first specified round number.

[0007] Optionally, in one of the above aspects, providing the model training result of the current image recognition model to the second member device can include providing the model training result of the current image recognition model to the second member device in response to a second predetermined condition being met.

[0008] Optionally, in one of the above aspects, the second predetermined condition can include that a round number interval between a current training round number of the image recognition model and a training round number at a last time when the model training result is sent reaches a second specified round number.

[0009] Optionally, in one of the above aspects, the first hyperparameter is k, and a maximum weight coefficient of the mixed image is W max The data augmentation based image mixing processing of the first desensitized image data using the first hyperparameter can include: performing k-1 shuffle processes on a set of image data of the first desensitized image data to obtain k sets of image data; constructing an image super matrix of size m*k based on the k sets of image data, a first column in the image super matrix corresponding to an original set of image data of the first desensitized image data, and m being a number of image data in the original set of image data; randomly generating a weight coefficient for each image data in the image super matrix; performing row normalization on the weight coefficients of the image data in the image super matrix, so that the sum of the weight coefficients of each row of image data is 1, and the weight coefficient of each image data is not greater than W max ; and performing weighted summation on each row of image data of the image super matrix to obtain a mixed image super matrix of size m*1, the image data in the mixed image super matrix being desensitized image data after data augmentation.

[0010] Optionally, in one example of the above aspect, the performing frequency domain transformation based data de-sensitization processing on the current training sample image data can include: performing local frequency domain transformation processing on the current training sample image data to obtain at least one feature map, each feature map including a plurality of elements and corresponding to one data block in the current training sample image data, each element corresponding to one frequency in the frequency domain; constructing a frequency component channel feature map corresponding to each frequency using the element corresponding to the frequency in the at least one feature map; and screening at least one target frequency component channel feature map from the constructed frequency component channel feature maps to obtain the de-sensitized image data of the current training sample image data, the screened target frequency component channel feature map being a key channel feature for image recognition.

[0011] Optionally, in one example of the above aspect, after the at least one target frequency component channel feature map is screened from the constructed frequency component channel feature maps, the method can further include: performing first shuffling processing on the target frequency component channel feature map to obtain a first shuffled feature map; and performing normalization processing on the first shuffled feature map to obtain the de-sensitized image data of the current training sample image data.

[0012] Optionally, in one example of the above aspect, after the first shuffled feature map is subjected to the normalization processing, the method can further include: performing channel mixing processing on the normalized first shuffled feature map; performing second shuffling processing on the first shuffled feature map subjected to the channel mixing processing to obtain a second shuffled feature map; and performing normalization processing on the second shuffled feature map to obtain the de-sensitized image data of the current training sample image data.

[0013] According to another aspect of embodiments of the present specification, there is provided a method for training an image recognition model, the method being performed by at least two first member devices having local training data and a second member device for maintaining an image recognition model, the method comprising: cyclically performing a model training process as follows until a model training end condition is met: at each first member device, training a current image recognition model locally using local training sample image data according to the method as described above; and at the second member device, updating the current image recognition model using model training results of the current image recognition model received from each first member device, and sending the updated image recognition model to each first member device for local model training.

[0014] According to another aspect of embodiments of the present specification, there is provided an apparatus for training an image recognition model, the apparatus being applied to a first member device with local training data for execution, the apparatus comprising: an image recognition model receiving unit, acquiring a current image recognition model from a second member device for maintaining an image recognition model; a training sample data acquisition unit, acquiring current training sample image data and its labeled label data; a data desensitization processing unit, performing data desensitization processing based on frequency domain transformation on the current training sample image data to obtain first desensitized image data of the current training sample image data; a hyperparameter selection unit, providing the first desensitized image data to a hyperparameter selection model to select a first hyperparameter for indicating the number of images participating in image mixing processing from a candidate hyperparameter set; an image mixing processing unit, using the first hyperparameter to perform image mixing processing based on Mixup data augmentation on the first desensitized image data to obtain second desensitized image data and corresponding labeled label data after label mixing processing; a model training unit, training the current image recognition model using the second desensitized image data and the corresponding labeled label data after label mixing processing; and a model training result sending unit, sending a model training result of the current image recognition model to the second member device for the second member device to use to update the current image recognition model, wherein the image recognition model receiving unit, the training sample data acquisition unit, the data desensitization processing unit, the hyperparameter selection unit, the image mixing processing unit, the model training unit, and the model training result sending unit perform the following operations in a loop until a model training end condition is met.

[0015] Optionally, in one example of the above aspect, the apparatus can further include: an hyper-parameter selection model updating unit, in response to satisfying a first predetermined condition, updating the hyper-parameter selection model through a model updating process including: providing the first desensitized image data to the hyper-parameter selection model to select a second hyper-parameter for indicating a number of images participating in image mixing processing from the candidate hyper-parameter set; performing data augmentation based image mixing processing on the first desensitized image data using the second hyper-parameter to obtain third desensitized image data and corresponding label mixed processing marked label data; providing the third desensitized image data to the current image recognition model to obtain second predicted label data of the third desensitized image data, and determining a first loss function according to the second predicted label data and the corresponding label mixed processing marked label data; updating the current image recognition model according to the first loss function; providing the third desensitized image data to the updated image recognition model to obtain third predicted label data of the third desensitized image data, and determining a second loss function according to the third predicted label data and the corresponding label mixed processing marked label data; determining the third loss function based on the first loss function and the second loss function; and updating model parameters of the hyper-parameter selection model based on the third loss function.

[0016] According to another aspect of embodiments of the present specification, there is provided a system for training an image recognition model, including: at least two first member devices, each first member device having local training sample image data and including an apparatus for training an image recognition model as described above; and a second member device, the second member device maintaining an image recognition model, and the second member device including: a model training result receiving unit receiving model training results of an image recognition model from each first member device; a model updating unit updating the current image recognition model using the model training results of the current image recognition model received from each first member device; and a model sending unit sending the updated image recognition model to each first member device for local model training.

[0017] According to another aspect of embodiments of the present specification, there is provided an apparatus for training an image recognition model, including: at least one processor, a memory coupled with the at least one processor, and a computer program stored in the memory, the at least one processor executing the computer program to implement a method for training an image recognition model as described above.

[0018] According to another aspect of embodiments of the present specification, there is provided a training device for an image recognition model, comprising at least one processor, a memory coupled with the at least one processor, and a computer program stored in the memory, the at least one processor executing the computer program to implement the method for training an image recognition model as above.

[0019] According to another aspect of embodiments of the present specification, there is provided a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the method for training an image recognition model as above.

[0020] According to another aspect of embodiments of the present specification, there is provided a computer program product comprising a computer program executed by a processor to implement the method for training an image recognition model as above. BRIEF DESCRIPTION OF DRAWINGS

[0021] A further understanding of the nature and advantages of the present specification can be realized by reference to the following drawings. In the drawings, like reference numerals can designate similar components or features.

[0022] Figure 1 An example schematic diagram of a federated learning system for training an image recognition model according to embodiments of the present specification is shown.

[0023] Figure 2 An example flowchart of an image recognition model training method performed on a first member device side according to embodiments of the present specification is shown.

[0024] Figure 3 An example flowchart of a data de-sensitization process based on frequency domain transformation according to embodiments of the present specification is shown.

[0025] Figure 4 An example schematic diagram of converting image feature data from a spatial domain to a frequency domain according to embodiments of the present specification is shown.

[0026] Figure 5 An example schematic diagram of a local frequency domain transformation process according to embodiments of the present specification is shown.

[0027] Figure 6 An example schematic diagram of a frequency component channel feature map according to embodiments of the present specification is shown.

[0028] Figure 7 Another example flowchart of a data de-sensitization process based on frequency domain transformation according to embodiments of the present specification is shown.

[0029] Figure 8An example flowchart of an updating process of a hyperparameter selection model according to an embodiment of the present specification is shown.

[0030] Figure 9 An example flowchart of a hyperparameter selection process based on a hyperparameter selection model according to an embodiment of the present specification is shown.

[0031] Figure 10 An example flowchart of an image mixing processing process based on data augmentation according to an embodiment of the present specification is shown.

[0032] Figure 11 An example structure diagram of an image recognition model according to an embodiment of the present specification is shown.

[0033] Figure 12 An example block diagram of an image recognition model training apparatus according to an embodiment of the present specification is shown.

[0034] Figure 13 An example block diagram of a data desensitization processing unit according to an embodiment of the present specification is shown.

[0035] Figure 14 An example block diagram of a data desensitization processing unit according to another embodiment of the present specification is shown.

[0036] Figure 15 An example block diagram of an image mixing processing unit according to an embodiment of the present specification is shown.

[0037] Figure 16 An example block diagram of a hyperparameter selection unit according to an embodiment of the present specification is shown.

[0038] Figure 17 An example block diagram of a hyperparameter selection model updating unit according to an embodiment of the present specification is shown.

[0039] Figure 18 An example block diagram of an image recognition model training apparatus of a second member device according to an embodiment of the present specification is shown.

[0040] Figure 19 An example schematic diagram of an image recognition model training apparatus applied to a first member device side implemented based on a computer system according to an embodiment of the present specification is shown. DETAILED DESCRIPTION

[0041] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that these implementations are discussed solely for the purpose of illustrating aspects of the subject matter described herein and are not a limitation of the scope, applicability, or examples set forth in the claims. Changes in the function and arrangement of elements discussed can be made without departing from the scope of the subject matter described herein. Various examples can omit, substitute, or add various procedures or components as appropriate. For instance, the methods described can be performed in an order different than that described, and / or various steps can be added, omitted, or combined. Also, features described with respect to some examples can be combined in other examples.

[0042] As used herein, the terms "includes," "including," "has," "having," "contains," "containing," "comprises," "comprising," "including," "contains," "containing" or variants thereof do not exclude additional elements or steps. The terms "based on" or "based upon" are not exclusive. The terms "a" or "an" do not denote a limitation of quantity, but rather denote the presence of at least one of the referenced item. The term "another" is used herein to denote at least a second or more of the referenced item. The terms "first," "second," or the like can be used herein to denote different or identical objects. The definitions provided below can apply equally to the use of the terms in the claims. Unless explicitly stated otherwise, the definitions are consistent throughout the specification and the claims.

[0043] Business processing schemes based on image recognition models have been widely applied in a large number of applications, for example, a face recognition model-based face payment business, etc. Image data involving data privacy information (for example, user privacy information, etc.) is usually distributed in different data owners or different regions and countries. In order to protect data privacy information, data sharing of these private data is not allowed between data owners or between different regions, but in order to utilize sufficient data to provide better services for users, it is necessary to fully mine the information between data for image recognition model training for specific tasks.

[0044] To this end, a federated learning method is proposed, which can use private image data of multiple data owners to train an image recognition model without data leaving the domain. However, in the federated learning method, each data owner needs to share gradient information or weight information with the model owner for aggregation after completing the image recognition model training locally, thereby causing information leakage of the gradient information or the weight information.

[0045] To avoid unnecessary losses caused by such information leakage, an embodiment of this specification proposes a federated learning scheme based on mixed data desensitization. In this federated learning scheme, when the first member device trains an image recognition model using local training sample image data, firstly, the training sample image data undergoes frequency transformation-based data desensitization processing, and then an image mixing processing method based on MixUp data augmentation is used to mix the desensitized image data. Subsequently, the desensitized image data after image mixing processing is used to train the image recognition model, thereby improving data privacy protection during federated learning. Furthermore, during image recognition model training, a hyperparameter selection model is used to adaptively select appropriate image mixing parameters (i.e., the number of images participating in image mixing) based on the first desensitized image data, thereby ensuring that multiple desensitized image data can be fused during image recognition model training while maintaining model training performance without significant impact.

[0046] The image recognition model training method and apparatus according to embodiments of this specification are described below with reference to the accompanying drawings.

[0047] Figure 1 An example schematic diagram of a federated learning system 100 for training an image recognition model according to an embodiment of this specification is shown.

[0048] like Figure 1 As shown, the federated learning system 100 includes at least two first member devices 110-1 to 110-n and a second member device 120. The at least two first member devices 110-1 to 110-n and the second member device 120 can communicate with each other via a network, such as, but not limited to, the Internet or a local area network (LAN). In some embodiments, the network can be any one or more of a wired network or a wireless network. Examples of the network 330 may include, but are not limited to, cable networks, fiber optic networks, telecommunications networks, corporate intranets, the Internet, local area networks (LANs), wide area networks (WANs), wireless local area networks (WLANs), metropolitan area networks (MANs), public switched telephone networks (PSTNs), Bluetooth networks, ZigZee networks, near field communication (NFC), device internal buses, device internal lines, etc., or any combination thereof.

[0049] In the embodiments described in this specification, the first member devices 110-1 to 110-n can be devices or device owners used for locally collecting image data samples, such as smart terminal devices, server devices, etc. In this specification, the terms "first member device," "data owner," and "client device" are used interchangeably. The collected image data samples can be referred to as local image data samples.

[0050] In this specification, the local image data samples possessed by each of the first member devices 110-1 to 110-n together constitute the training sample data of the image recognition model, and the local image data samples possessed by each of the first member devices 110-1 to 110-n are the secret of that first member device and cannot be known or fully known by other first member devices or second member devices.

[0051] The second member device 120 can be a device or device owner used to maintain the image recognition model, such as a smart terminal device, a server device, etc. In this specification, the terms "second member device," "model owner," and "server device" are used interchangeably.

[0052] In a practical application example, the second member device 120 may be, for example, a server of a service provider or operator, such as a server device of a third-party payment platform that provides facial payment services. Each first member device may be, for example, a client device that interacts with the user for business services, such as a facial scanning payment device deployed at a checkout counter.

[0053] In this specification, each of the first member devices 110 and the second member device 120 can be any suitable electronic device with computing capabilities. Such electronic devices include, but are not limited to: personal computers, server computers, workstations, desktop computers, laptop computers, notebook computers, mobile electronic devices, smartphones, tablet computers, cellular phones, personal digital assistants (PDAs), handheld devices, messaging devices, wearable electronic devices, consumer electronic devices, etc.

[0054] The first member devices 110-1 to 110-n and the second member device 120 may each have an image recognition model training device. The image recognition model training devices at the first member devices 110-1 to 110-n and the second member device 120 can perform data interaction via network communication, thereby collaboratively processing to perform the model training process for the image recognition model.

[0055] like Figure 1 As shown, during image recognition model training, the second member device 120 distributes the image recognition model W to each of the first member devices 110-1 to 110-n. Upon receiving the image recognition model W, each of the first member devices 110-1 to 110-n uses local training sample image data to train the image recognition model W locally, and displays the local model training results of the image recognition model W (e.g., gradient information). Or model parameter information W i t) to the second member device 120. The second member device 120 updates the image recognition model W using the model training results received from the respective first member devices. For example, the second member device 120 aggregates the model training results received from the respective first member devices, and updates the image recognition model W based on the aggregated results. Then, the second member device 120 re-issues the updated image recognition model W' to the respective first member devices 110-1 to 110-n. This cycle is performed until a model training end condition is satisfied, thereby ending the image recognition model training.

[0056] In some embodiments, the first member device 110 provides the model training results to the second member device 120 whenever the first member device 110 completes the local image recognition model training. In some embodiments, the first member device 110 provides the model training results to the second member device 120 only when the first member device 110 completes the local image recognition model training and a predetermined condition (i.e., a second predetermined condition) is satisfied. Examples of the predetermined condition can include, but are not limited to, for example, a round number interval between a current training round number of the image recognition model and a training round number when the model training results were last sent reaches a specified round number (i.e., a second specified round number). For example, the first member device 110 can provide the model training results of the current training round to the second member device 120 for image recognition model updating only every t rounds. In other training rounds, the first member device 110 only locally completes the model updating of the image recognition model, and uses the updated image recognition model to complete the next round of model training.

[0057] Figure 2 An example flowchart of the image recognition model training method 200 performed on the first member device side according to an embodiment of the present specification is shown.

[0058] It is to be noted that the image recognition model training process performed on the first member device side is a loop process until a training end condition is satisfied. Examples of the training end condition include, but are not limited to, for example, reaching a training round number, the image recognition result satisfying a predetermined requirement, such as the image recognition rate reaching a predetermined value, the image recognition difference being within a predetermined range, etc. Figure 2 The image recognition model training process shown is a single round training process.

[0059] Specifically, as Figure 2As shown, at 210, current training sample image data and its labeled label data are obtained. In the case where the image recognition model is a face recognition model, the training sample image data may, for example, include face image data, and the label thereof may be identity information corresponding to the face in the image, for example, a person's name, etc. The label of the training sample image data may be added manually or in other manners, and the embodiments are not limited thereto. In some embodiments, the training sample image data and its label may be obtained in the manner of reading or calling a data interface from a database.

[0060] The training sample image data may be original face image data, or image data obtained after face detection or face alignment on the original face image data. The original face image data may refer to image data collected directly by an image collection device (for example, a camera, etc.) without any processing. Face detection refers to detecting the position of a face in an image. The image data to be processed may be an image obtained by cropping based on the position of the face in the image, for example, removing the redundant part of the image except the face. Face alignment refers to correcting the angle of the face in the image. The face in the original face image may be tilted at a certain angle. Through face alignment, the face in the image can be straightened to facilitate subsequent recognition processing of the image, etc.

[0061] In some embodiments, the processing device may obtain the training sample image data through the camera of the terminal device, or may obtain the training sample image data in the manner of reading from a database, a storage device, or calling a data interface, etc.

[0062] It should be noted that the program / code for obtaining the training sample image data may run in a trusted execution environment deployed in the processing device. The security features of the trusted execution environment can ensure that the image data obtained by the processing device cannot be stolen. At the same time, the method and / or process disclosed in the embodiments of the present specification may also be executed in the trusted execution environment to ensure that the entire process from the source of obtaining the training sample image data to the processing of the training sample image data is secure and trusted, thereby improving the security of privacy protection of the training sample image data.

[0063] At 220, the current training sample image data is subjected to frequency domain transformation-based data desensitization processing to obtain first desensitized image data of the current training sample image data.

[0064] Figure 3 An example flowchart of the frequency domain transformation-based data desensitization processing process 300 according to the embodiments of the present specification is shown.

[0065] As Figure 3As shown in 310, a local frequency domain transformation processing is performed on the current training sample image data to obtain at least one feature map, each feature map including a plurality of elements and corresponding to a data block in the image data, each element corresponding to a frequency in the frequency domain.

[0066] The feature map refers to a plurality of sub-maps extracted from the to-be-processed image data by using certain image processing means, each sub-map carrying part of the features of the to-be-processed image data. The obtained feature map can be the same size as the training sample image data, for example, pixel-to-pixel correspondence, or can be different from the training sample image data.

[0067] In some embodiments, examples of the local frequency domain transformation processing can include, but are not limited to, local discrete cosine transformation, local wavelet transformation, or local discrete Fourier transformation.

[0068] Figure 4 An example schematic diagram of converting image data from a spatial domain to a frequency domain according to an embodiment of the present specification is shown. In Figure 4 The spatial domain is represented by a coordinate system (x, y), and the frequency domain is represented by a coordinate system (c, v), N*M represents the size of the image, for example, Figure 4 In the above example, N*M is 2*2. The number of feature points after the conversion of the spatial domain and the frequency domain can be consistent, one block in the spatial domain represents one pixel position, and one block in the frequency domain represents one frequency position.

[0069] In some embodiments, the following discrete cosine transformation formula (1) can be used to perform discrete cosine transformation on the to-be-transformed gray-scale image data.

[0070]

[0071] wherein F(u, v) is the value of the feature point (i.e., each frequency position) in the converted frequency domain, f(x, y) is the pixel value in the to-be-transformed image data (gray-scale image data), (u, v) is the coordinate of the feature point in the converted frequency domain, (x, y) is the coordinate of the to-be-transformed image data in the spatial domain, N is the number of rows of pixel points or feature points of the to-be-transformed image data, and M is the number of columns of pixel points or feature points of the to-be-transformed image data, for example, when the size of the to-be-processed image data is 8*8, N=M=8.

[0072] c(u) can be represented by the following formula (2).

[0073]

[0074] wherein c(u)=c(v).

[0075] The local frequency domain transformation processing is performed on the image data, and a plurality of transformation results, i.e., a plurality of feature maps, can be obtained. When performing the local frequency domain transformation processing, a smaller image block (local image block) than the image data to be transformed can be selected, for example, the size of the image data to be transformed is 256x256, and the size of the selected image block is 8x8. Then, the selected image block size is used to move and sample the image data to be transformed at a certain step (for example, 8), and the local data (i.e., the data block of 8x8 size) of the image data to be transformed obtained by each sampling is subjected to discrete cosine transformation based on formulas (1) and (2), thereby obtaining a plurality of transformation results, each of which can be 8x8 size. The smaller the moving step of the image block during the discrete cosine transformation, the more features contained in the obtained transformation result, which can be beneficial to improve the accuracy of subsequent image data processing.

[0076] Figure 5 An example schematic diagram of a local frequency domain transformation process according to an embodiment of the present specification is shown. In Figure 5 In an example, the size of the image data to be transformed is 6x6, the size of the selected local image block is 2x2, and the moving sampling is performed on the image data to be transformed at a step of 2, and the frequency domain transformation, for example, the discrete cosine transformation, is performed on the sampled local image block, thereby obtaining 9 transformation results, i.e., 9 feature maps, 51, 52, 53, 54, 55, 56, 57, 58 and 59. The values of each frequency position in each transformation result are represented by fi1, fi2, fi3, fi4, respectively, where i represents the i-th transformation result, and fij represents the value of the j-th frequency position in the i-th transformation result. As can be seen from the figure, each transformation result has 4 corresponding frequency positions.

[0077] At 320, using the elements corresponding to each frequency in the at least one feature map, a frequency component channel feature map corresponding to each frequency is constructed. Specifically, the elements (values) of the same frequency position in each transformation result are combined to obtain a frequency component channel feature map, and then a plurality of frequency component channel feature maps corresponding to different frequency positions in the corresponding transformation result are obtained. It is not difficult to understand that the number of frequency component channel feature maps is consistent with the number of pixel points of the image block used for sampling in the transformation process.

[0078] Figure 6 An example schematic diagram of a frequency component channel feature map according to an embodiment of the present specification is shown. Figure 6 The frequency component channel feature map shown in Figure 5 is the frequency component channel feature map corresponding to the transformation result in

[0079] As shown in Figure 6 , based on Figure 5The transformation result in the discrete cosine transformation can be obtained, and four frequency component channel feature maps (the number of pixels of the sampled image block is 4) can be obtained, each of which includes 9 elements. The frequency component channel feature map 61 corresponds to the first frequency position fi1, the frequency component channel feature map 62 corresponds to the second frequency position fi2, the frequency component channel feature map 63 corresponds to the third frequency position fi3, and the frequency component channel feature map 64 corresponds to the fourth frequency position fi4.

[0080] After the frequency component channel feature map is constructed as above, at least one target frequency component channel feature map is selected from the constructed frequency component channel feature map at 330. The selected target frequency component channel feature map is a key channel feature for image recognition.

[0081] In some embodiments, at least one target frequency component channel feature map can be selected from the constructed frequency component channel feature map based on channel importance or based on a preset selection rule.

[0082] In some embodiments, the processing device can input the plurality of transformation results into the trained SEnet network, and the SEnet network can give the channel importance (such as a score positively correlated with importance) of each feature map. Here, the channel importance refers to the channel importance relative to image recognition. The SEnet network can be trained together with the image recognition model (i.e., as a whole), for example, by adding the SEnet network in the image recognition model, adjusting the parameters of the SEnet network in the training process of the image recognition model, so as to obtain the SEnet network for determining the channel importance of the feature map.

[0083] In some embodiments, the preset selection rule can be to select a preset proportion of partial feature maps containing more feature information. For example, in the plurality of feature maps obtained after the discrete cosine transformation and reorganization, a preset proportion of low-frequency partial feature maps can be selected, and partial high-frequency feature maps can be discarded. For example, 50%, 60%, or 70% of low-frequency feature maps can be retained, and the remaining high-frequency feature maps can be discarded. For example, the low-frequency feature maps 61, 62, and 63 shown in FIG. 6 can be retained, and the high-frequency feature map 64 can be discarded. Figure 6 In the transformation result 51 in FIG. 5, f11 corresponds to low-frequency data, and f14 corresponds to high-frequency data. According to the above formula (1), when (u, v) is (0, 0) Figure 5 In the transformation result 51 in FIG. 5, f11 corresponds to low-frequency data, and f14 corresponds to high-frequency data. According to the above formula (1), when (u, v) is (0, 0) F(0,0) in the upper left corner of the transformation result, the value corresponding to the lowest frequency, as the position moves to the right lower corner, F(u, v) contains alternating current components, and the frequency increases, so the right lower corner of the transformation result value corresponding to the highest frequency.

[0084] In Figure 3 the example, the filtered at least one target frequency component channel feature map is taken as the first de-sensitized image data of the training sample image data. In this way, the first de-sensitized image data is the image data after de-sensitizing the plurality of feature maps, and can include one or more frequency component channel feature maps after de-sensitizing. The de-sensitized image data is different from the image data to be processed, since the de-sensitized image data is a feature map, and thus the original facial information of the image data cannot be directly obtained from the de-sensitized image data.

[0085] In Figure 3 the example, the de-sensitizing method includes frequency component channel feature map reconstruction of the feature map and feature map filtering of the frequency component channel feature map. In other embodiments, in addition to the above operations, the de-sensitizing method can also include shuffling processing, normalization processing, channel mixing processing, etc., or any combination thereof.

[0086] Figure 7 Another example flowchart of the frequency domain transformation-based data de-sensitizing process 700 according to an embodiment of the present specification is shown. Figure 7 The embodiment shown is for Figure 3 a modification of the embodiment shown. Figure 7 Steps 710 to 730 in Figure 2 are exactly the same as 210 to 230 in . For the sake of simplicity of description, the same parts will not be described below, and only the differences will be described.

[0087] As Figure 7 shown, after the target frequency component channel feature map is filtered in 730, the filtered target frequency component channel feature map is not taken as the first de-sensitized image data, but the operations of 740 to 760 are continued, thereby obtaining more secure first de-sensitized image data.

[0088] Specifically, after screening at least one target frequency component channel feature map from the constructed frequency component channel feature map, at 740, a first shuffle processing is performed on the screened target frequency component channel feature map to obtain a first shuffled feature map, and a normalization processing is performed on the first shuffled feature map. In some embodiments, the data after the normalization processing on the first shuffled feature map can be directly used as the first desensitized image data. In some embodiments, subsequent processing needs to be performed on the data after the normalization processing on the first shuffled feature map to obtain the first desensitized image data.

[0089] In some embodiments, the first shuffle processing can be a sequential randomization on the screened target frequency component channel feature map. Sequential randomization refers to disordering the arrangement order of multiple feature maps. For example, the screened multiple target frequency component channel feature maps are 61, 62 and 63, and the arrangement order after the sequential randomization can be 63, 61, 62. Figure 6

[0090] In some embodiments, the normalization parameter refers to a parameter used when performing the normalization processing on the multiple target frequency component channel feature maps. When performing the normalization processing, the normalization coefficient of each frequency component channel feature map can be determined based on the frequency component channel feature map, so that the normalization parameter used when performing the normalization processing on each frequency component channel feature map is only related to the frequency component channel feature map itself and is irrelevant to other frequency component feature maps, which can increase the difficulty of back-propagation on the voice data. For example, it is assumed that a frequency component channel feature map is back-propagated, but since the parameters used when performing the normalization processing on each frequency component channel feature map are different, the normalization parameter of the back-propagated frequency component channel feature map cannot be used to back-propagate other frequency component channel feature maps. The above normalization processing can also be referred to as self-normalization processing.

[0091] In some embodiments, the normalization parameter can be the mean or variance of all values of the frequency component channel feature map, or can be the maximum or minimum value of all values of the frequency component channel feature map. The normalization processing can be to remove the numerical value of each element in the frequency component channel feature map by using the normalization parameter, and to replace the original numerical value with the quotient obtained by division, so as to obtain the normalized frequency component channel feature map.

[0092] By using the above first shuffle processing and the corresponding normalization processing, it is impossible to obtain the original data of the screened target frequency component channel feature map, so as to protect the data privacy and security of the screened target frequency component channel feature map.

[0093] At 750, a channel mixing processing is performed on the normalized first shuffled feature map.

[0094] ​The mixing processing can refer to performing a pre-designed calculation on two or more feature maps in the plurality of frequency component channel feature maps. For example, the values of corresponding elements in the two or more frequency component channel feature maps can be calculated, and the calculated values can be taken as the values of the corresponding elements in the mixed frequency component channel feature map, so that the two or more frequency component channel feature maps can be mixed into one frequency component channel feature map. The pre-designed calculation can be mean value calculation, sum value calculation, difference value calculation, etc.

[0095] In some embodiments, the channel mixing can be mixing two adjacent frequency component channel feature maps. It should be noted that when the feature maps are combined, the combination rules for different frequency component channel feature maps should be the same. For example, starting from the first frequency component channel feature map, the current frequency component channel feature map is combined with the next frequency component channel feature map adjacent thereto, i.e., the first frequency component channel feature map is combined with the second frequency component channel feature map, and the second frequency component channel feature map is combined with the third frequency component channel feature map. In this way, for M frequency component channel feature maps, M-1 frequency component channel feature maps can be obtained, thereby playing a role in reducing the dimension.

[0096] In some embodiments, in the case of mixing two adjacent frequency component channel feature maps, the number of the selected target frequency component channel feature maps can be set to the feature dimension of the training sample image data plus one. In this way, the feature dimension of the obtained desensitized image data can be the same as that of the training sample image data, so that the model architecture of the image recognition model does not need to be modified.

[0097] After the channel mixing, the values of the elements in the mixed frequency component channel feature map will change compared with the values of the frequency component channel feature maps before the channel mixing, so that the relative relationship between the values of the elements in the original frequency component channel feature maps can be destroyed, thereby further increasing the difficulty of inversely deducing the original image data from the frequency component channel feature maps.

[0098] After the above channel mixing, in 760, the first shuffled feature map after the channel mixing is subjected to a second shuffling processing to obtain a second shuffled feature map, and the second shuffled feature map is subjected to a normalization processing. The second shuffling processing and the normalization processing thereof can refer to the description of 750, which will not be repeated here. In some embodiments, the data of the second shuffled feature map after the normalization processing can be directly taken as the first desensitized image data. In some embodiments, the data of the second shuffled feature map after the normalization processing needs to be subjected to subsequent processing to obtain the first desensitized image data.

[0099] By using the second shuffling processing and the corresponding normalization processing, the original data of the frequency component channel feature map after channel mixing cannot be obtained, so that the data privacy security of the frequency component channel feature map after channel mixing can be protected. It should be noted that in some embodiments, the first shuffling processing can adopt a pseudo-random shuffling processing, and the second shuffling processing can adopt a completely random shuffling processing.

[0100] In Figure 7 In the data desensitization processing shown, because two shuffling and normalization processes are involved, the difficulty of brute force cracking can be greatly increased. For example, the first feature data is subjected to local cosine transformation according to 8*8 local image blocks, 64 frequency component channel feature maps can be constructed, and after shuffling processing of the frequency component channel feature maps, the placement of each small block frequency component (one frequency component corresponds to one frequency component channel feature map) in the transformation result is random, and the randomization of the brute force cracking space size is 64!, where "!" represents the factorial operation. Even if the channel filtering of the frequency component channel feature map is performed by channel importance to determine, for example, 36 main feature maps, the brute force cracking space is 36!, and the two randomization processes in the desensitization process make the brute force cracking space size 36! * 36!, which will be greater than the key cracking space of the 256-bit AES encryption algorithm, which makes it difficult to reverse the original voice data by brute force cracking. At the same time, the normalization parameter only depends on the corresponding frequency component channel feature map, and the normalization parameters of different frequency component channel feature maps are different, which makes it more difficult to infer the normalization parameter of each frequency component channel feature map. At the same time, the frequency component channel feature map is subjected to channel mixing processing, only the result after channel mixing is retained, which destroys the relative relationship between the values of the frequency component channel feature maps, thereby further improving the data cracking difficulty, thereby enhancing the privacy security protection of the voice data.

[0101] Back to Figure 2 After obtaining the first desensitized image data as above, it is determined whether a first predetermined condition is met at 230. Examples of the first predetermined condition include, but are not limited to, for example: the number of intervals between the current training rounds of the image recognition model and the training rounds at the last update processing of the hyperparameter selection model reaches a first specified number of rounds, for example, m rounds. That is, every m rounds, the model update processing of the hyperparameter selection model is performed once.

[0102] If the first predetermined condition is met, the hyperparameter model update is performed at 240. Then, proceed to 250. If the first predetermined condition is not met, proceed directly to 250.

[0103] Figure 8An example flowchart of an update process 800 for a hyperparameter selection model according to an embodiment of this specification is shown. In some embodiments, examples of hyperparameter selection models may include, but are not limited to, the ResNet18 network.

[0104] like Figure 8 As shown in step 810, the first desensitized image data is provided to the hyperparameter selection model to select a second hyperparameter from the candidate hyperparameter set to indicate the number of images participating in the image blending process. Here, the selected second hyperparameter is used for image blending during the hyperparameter model update process.

[0105] Figure 9 An example flowchart of a hyperparameter selection process 900 based on a hyperparameter selection model according to an embodiment of this specification is shown.

[0106] like Figure 9 As shown, at 910, the image size of the first desensitized image data is scaled to the original image size of the training sample image data. For example, the first desensitized image data can be provided to the adaptation layer of the ResNet18 network, where the adaptation layer performs width and height interpolation on the first desensitized image data, thereby scaling the image size of the first desensitized image data to the original image size of the training sample image data. Using this image scaling process, it is possible to apply the hyperparameter model to the embodiments of this specification simply by adjusting the input channel size of the first layer of the existing ResNet18 network to the increased number of channels, without any other modifications.

[0107] In step 920, the scaled first desensitized image data is provided to the feature extraction layer of the hyperparameter selection model to extract the feature map representation of the first desensitized image data. It should be noted that the feature map representation obtained here is the feature map representation of the entire batch of image data, i.e., batch-sized feature map representation. For example, the matrix dimension of the obtained feature map representation is b*c*h*w, where b is the batch size, c is the number of channels, h is the feature map height dimension, and w is the feature map width dimension.

[0108] In step 930, the feature map representation of the first desensitized image data is subjected to pooling processing, thereby pooling the image number dimension b of the feature map representation from batchsize dimension to 1 dimension, thus obtaining the feature map representation after pooling processing, whose matrix dimension is 1*c*h*w.

[0109] At 940, the pooling result is provided to the fully connected layer of the hyperparameter selection model to obtain a selection probability of each candidate hyperparameter in the candidate hyperparameter set. Here, the candidate hyperparameter is a hyperparameter used to indicate the number of images participating in the image mixing process. The candidate hyperparameter set is usually a discrete parameter set, for example, {2, 3, 4, 5, 6}, etc.

[0110] At 950, the second hyperparameter is selected based on the selection probability of each candidate hyperparameter.

[0111] Returning to Figure 8 At 820, the first desensitized image data is subjected to data augmentation-based image mixing processing using the second hyperparameter to obtain third desensitized image data and corresponding label mixed labeled data.

[0112] Mixup data augmentation usually has two hyperparameters. One is the maximum weight coefficient W max of the mixed image (the sum of all weight coefficients of the mixed image needs to be 1). Usually, the maximum weight coefficient W max is 0.65 by default. Preferably, the maximum weight coefficient W max may be set to 0.55 so that different images can contribute more data to the mixing, thereby bringing higher privacy protection capability. The other hyperparameter is the number k of images participating in the mixing operation. The larger the value of k, the more information mixed, the stronger the privacy protection capability, and the lower the recognition rate.

[0113] Figure 10 An example flowchart of a Mixup data augmentation-based image mixing process 1000 according to an embodiment of the present specification is shown.

[0114] As Figure 10 shown, at 1010, k-1 times of shuffling processing is performed on the obtained image data set in the desensitized image data to obtain k image data sets.

[0115] At 1020, based on the obtained k image data sets, an image super matrix of size m*k is constructed. The first column in the constructed image super matrix corresponds to the original image data set, and the remaining columns each correspond to an image data set obtained after each time of shuffling processing. m is the number of sub-image data in the original image data set.

[0116] At 1030, a weight coefficient is randomly generated for each image data in the image super matrix.

[0117] At 1040, the weight coefficients of the image data in the image super matrix are row normalized so that the sum of the weight coefficients of each row of image data is 1, and the weight coefficient of each image data is not greater than W maxThat is, after normalization, the maximum weight coefficient of each row of the image cannot exceed W max , for example, cannot exceed 0.55.

[0118] At 1050, the weighted sum of each row of image data of the image super matrix is performed to obtain a mixed image super matrix of size m*1, and the image in the obtained mixed image super matrix is the third desensitized image data after image mixing processing.

[0119] In addition, when performing image mixing processing, the weight coefficient of each desensitized image data is also used to perform label mixing processing on the corresponding labeled label of each desensitized image data, so as to obtain the labeled label after label mixing processing for each desensitized image data.

[0120] In some embodiments, when performing label mixing processing, if the image data participating in the image mixing processing is from the same category, the label of each image data participating in the image mixing processing is kept unchanged. If at least part of the image data participating in the image mixing processing is from different categories, one of the labels of each image data is adjusted to k non-zero values, and each non-zero value corresponds to a weight coefficient of a sub-image data.

[0121] Returning to Figure 8 At 830, the first loss function Loss1 is determined using the third desensitized image data, the corresponding labeled label data after label mixing processing, and the current image recognition model.

[0122] For example, the third desensitized image data can be provided to the current image recognition model to obtain the first predicted label data of the third desensitized image data. Then, the first loss function Loss1 is determined according to the first predicted label data and the corresponding labeled label data after label mixing processing.

[0123] In this specification, the image recognition model can be any suitable machine learning model.

[0124] Figure 11 An example structure diagram of an image recognition model 1100 according to an embodiment of the present specification is shown. As Figure 11 shown, the image recognition model 1100 can include an input layer 1110, a feature extraction layer 1120, and an output layer 1130.

[0125] The input layer 1110 can be used to receive the desensitized image data after image mixing processing obtained after the image data processing described in the Figure 2 above.

[0126] In some embodiments, the input layer 1110 may have multiple input channels, the number of which may be the same as the number of feature maps (e.g., frequency component channel feature maps) in the desensitized image data, with each channel corresponding to one feature map.

[0127] In some embodiments, the number of input channels of the initially created image recognition model can be adjusted to match the number of feature maps obtained by the image processing method described above.

[0128] In some embodiments, the number of feature maps in the obtained desensitized image data can be set to match the number of channels in the original image recognition model by setting the number of target feature maps after filtering. This allows the model architecture of the original image recognition model to be used without any adjustments.

[0129] The feature extraction layer 1120 can be used to process the input desensitized image data to obtain the feature map representation (or prediction vector) of the desensitized image data.

[0130] In some embodiments, the feature extraction layer can be a deep neural network, such as a CNN or RNN. The feature extraction layer can process each feature map (e.g., convolution, pooling) to obtain a more abstract feature map representation.

[0131] The output layer 1130 can convert the feature map representation into the identification result of the target object corresponding to the desensitized image data.

[0132] The target object can refer to a living organism, an object, or a part thereof in the image. For example, a person, a face, an animal, or a building.

[0133] The identity recognition result can refer to the corresponding identity of the target object in the image, such as the identity of a person, the category of an animal, the name of a building, etc.

[0134] The output layer 1130 can transform the feature map representation of the desensitized image data to obtain a predicted value, which can indicate the identity information of the task in the image, i.e. the identity recognition result of the target object.

[0135] In some embodiments, the output layer 1130 may be a multilayer perceptron, a fully connected layer, etc., and this embodiment does not limit it.

[0136] At step 840, the current image recognition model is updated based on the first loss function. For example, the gradient information of each model parameter of the image recognition model can be determined based on the first loss function. Then, the parameter update amount of each model parameter can be determined based on the gradient information, and the image recognition model can be updated according to the determined parameter update amount.

[0137] At 850, a second loss function Loss2 is determined using the third desensitized image data, the corresponding labeled label data after label mixing processing, and the updated image recognition model.

[0138] For example, the third desensitized image data is provided to the updated current image recognition model to obtain second predicted label data of the third desensitized image data, and the second loss function Loss2 is determined according to the second predicted label data and the corresponding labeled label data after label mixing processing.

[0139] At 860, a third loss function Loss3 is determined based on the first loss function Loss1 and the second loss function Loss2. For example, the formula Loss3 = (a*Loss1+b*Loss2)*Pro k may be used, where a and b are weights corresponding to the first loss function and the second loss function, and Pro k is the selection probability of the second hyperparameter k.

[0140] At 870, the model parameters of the hyperparameter selection model are updated based on the third loss function Loss3. For example, the gradient information of each model parameter of the hyperparameter selection model can be determined based on the third loss function Loss3, then the parameter update amount of each model parameter is determined based on the gradient information, and the hyperparameter selection model is updated according to the determined parameter update amount.

[0141] Returning to Figure 2 At 250, the first desensitized image data is provided to the hyperparameter selection model to select a first hyperparameter for indicating the number of images participating in the image mixing processing from the candidate hyperparameter set. It should be noted that if the hyperparameter selection model is updated, the first hyperparameter is selected using the updated hyperparameter selection model at 250. If the hyperparameter selection model is not updated, the first hyperparameter is selected using the original hyperparameter selection model at 250.

[0142] At 260, the first desensitized image data is subjected to image mixing processing based on Mixup data enhancement using the first hyperparameter to obtain second desensitized image data and corresponding labeled label data after label mixing processing.

[0143] At 270, the current image recognition model is trained using the second desensitized image data and the corresponding labeled label data after label mixing processing to obtain a model training result of the current image recognition model. The model training result may, for example, include gradient information, model parameter update amount, or updated model parameters.

[0144] For example, the second desensitized image data can be provided to the current image recognition model to predict third predicted label data of the second desensitized image data, and a fourth loss function is determined according to the third predicted label data and the mixed label data after the label data is mixed. Then, the model training result of the current image recognition model is determined according to the fourth loss function.

[0145] At 280, it is determined whether a second predetermined condition is met. Examples of the second predetermined condition can include, but are not limited to, for example, that the interval of the training rounds between the current training round of the image recognition model and the training round when the model training result is last sent reaches a specified round number (i.e., a second specified round number).

[0146] If the second predetermined condition is not met, return to 210 to perform the next loop training process using the updated image recognition model. If the second predetermined condition is met, at 290, the model training result of the current image recognition model is provided to the second member device. In response to receiving the model training results from the plurality of first member devices, the second member device updates the image recognition model using the model training results from the plurality of first member devices, and provides the updated image recognition model to each first member device for the next round of image recognition model training

[0147] It should be noted that the above description of each flow is merely for example and illustration, and does not limit the scope of application of the embodiments of the present specification. Various modifications and changes can be made to the flow under the guidance of the present specification. However, these modifications and changes are still within the scope defined by the embodiments of the present specification. For example, in other embodiments, part of the steps in Figure 2 may be removed, such as part or all of the operations of 230, 240 and 280.

[0148] Figure 12 An example block diagram of an image recognition model training apparatus 1200 according to an embodiment of the present specification is shown. As shown in Figure 12 , the image recognition training apparatus 1200 includes an image recognition model receiving unit 1210, a training sample data obtaining unit 1220, a data desensitization processing unit 1230, a hyperparameter selection model updating unit 1240, a hyperparameter selection unit 1250, an image mixing processing unit 1260, a model training unit 1270 and a model training result sending unit 1280.

[0149] The image recognition model receiving unit 1210, the training sample data obtaining unit 1220, the data de-sensitization processing unit 1230, the hyper-parameter selection model updating unit 1240, the hyper-parameter selection unit 1250, the image mixing processing unit 1260, the model training unit 1270, and the model training result sending unit 1280 perform operations cyclically until a training end condition is met. Examples of the training end condition include, but are not limited to, reaching a training round number, or an image recognition result meeting a predetermined requirement, such as an image recognition rate reaching a predetermined value, an image recognition difference being within a predetermined range, and the like.

[0150] In each round of model training, the image recognition model receiving unit 1210 is configured to receive a current image recognition model from a second member device for maintaining the image recognition model. It is to be noted that in some embodiments, the image recognition model receiving unit 1210 does not receive a current image recognition model from a second member device for maintaining the image recognition model in each round of model training. For example, in the case where the first member device sends a model training result of the image recognition model to the second member device only every t rounds, the image recognition model receiving unit 1210 receives an updated image recognition model from the second member device for maintaining the image recognition model only when the second member device receives the model training result and updates the image recognition model based on the received model training result, as the current image recognition model for the next round of model training.

[0151] The training sample data obtaining unit 1220 is configured to obtain current training sample image data and its labeled label data. The operation of the training sample data obtaining unit 1220 can refer to the operation described above with reference to 210. Figure 2

[0152] The data de-sensitization processing unit 1230 is configured to perform frequency domain transformation-based data de-sensitization processing on the current training sample image data to obtain first de-sensitized image data of the current training sample image data. The operation of the data de-sensitization processing unit 1230 can refer to the operation described above with reference to 220. Figure 2

[0153] The hyper-parameter selection model updating unit 1240 is configured to update the model parameters of the hyper-parameter selection model based on the first de-sensitized image data. In some embodiments, the hyper-parameter selection model updating unit 1240 is configured to update the model parameters of the hyper-parameter selection model based on the first de-sensitized image data in response to a first predetermined condition being met. The operation of the hyper-parameter selection model updating unit 1240 can refer to the operation described above with reference to 240. Figure 2

[0154] ​​​The hyperparameter selection unit 1250 is configured to provide the first desensitized image data to a hyperparameter selection model to select a first hyperparameter for indicating the number of images participating in the image mixing processing from a candidate hyperparameter set. The operation of the hyperparameter selection unit 1250 can refer to the operation described above with reference to 250. Figure 2

[0155] The image mixing processing unit 1260 is configured to perform the Mixup data augmentation based image mixing processing on the first desensitized image data using the first hyperparameter to obtain second desensitized image data and corresponding label mixed labeled data. The operation of the image mixing processing unit 1260 can refer to the operation described above with reference to 260. Figure 2

[0156] The model training unit 1270 is configured to train the current image recognition model using the second desensitized image data and corresponding label mixed labeled data. The operation of the model training unit 1270 can refer to the operation described above with reference to 270. Figure 2

[0157] The model training result sending unit 1280 is configured to send the model training result of the current image recognition model to the second member device for the second member device to use to update the current image recognition model. In some embodiments, the model training result sending unit 1280 is configured to send the model training result of the current image recognition model to the second member device for the second member device to use to update the current image recognition model in response to that a second predetermined condition is satisfied.

[0158] Figure 13 An example block diagram of a data desensitization processing unit 1300 according to an embodiment of the present specification is shown. As shown, the data desensitization processing unit 1300 includes a local frequency domain transformation module 1310, a channel feature map construction module 1320, and a feature map screening module 1330. Figure 13

[0159] The local frequency domain transformation module 1310 is configured to perform local frequency domain transformation processing on the image data to obtain at least one feature map, each feature map including a plurality of elements and corresponding to a data block in the image data, each element corresponding to a frequency in the frequency domain. The operation of the local frequency domain transformation module 1310 can refer to the operation described above with reference to 310. Figure 3

[0160] The channel feature map construction module 1320 is configured to construct a frequency component channel feature map corresponding to each frequency using the elements corresponding to the frequency in the at least one feature map, respectively. The operation of the channel feature map construction module 1320 can refer to the operation described above with reference to 320. Figure 3 ​​​​​the operations described above with reference to 320.

[0161] The feature map screening module 1330 is configured to screen at least one target frequency component channel feature map from the constructed frequency component channel feature maps, so as to obtain the desensitized image data of the image data, and the screened target frequency component channel feature map is a key channel feature of image recognition. The operations of the feature map screening module 1330 can refer to the operations described above with reference to 330. Figure 3

[0162] Figure 14 An example block diagram of the data desensitization processing unit 1400 according to another embodiment of the present specification is shown. As shown, the data desensitization processing unit 1400 includes a local frequency domain transformation module 1410, a channel feature map construction module 1420, a feature map screening module 1430, a first shuffling module 1440, a first normalization processing module 1450, a channel mixing processing module 1460, a second shuffling module 1470, and a second normalization processing module 1480. Figure 14

[0163] The local frequency domain transformation module 1410 is configured to perform local frequency domain transformation processing on the image data to obtain at least one feature map, each feature map including a plurality of elements and corresponding to a data block in the image data, and each element corresponding to a frequency in the frequency domain. The operations of the local frequency domain transformation module 1410 can refer to the operations described above with reference to 710. Figure 7

[0164] The channel feature map construction module 1420 is configured to construct a frequency component channel feature map corresponding to each frequency using the elements corresponding to the frequency in the at least one feature map, respectively. The operations of the channel feature map construction module 1420 can refer to the operations described above with reference to 720. Figure 7

[0165] The feature map screening module 1430 is configured to screen at least one target frequency component channel feature map from the constructed frequency component channel feature maps, and the screened target frequency component channel feature map is a key channel feature of image recognition. In some embodiments, the feature map screening module 1430 can screen at least one target frequency component channel feature map from the constructed frequency component channel feature maps based on channel importance or based on a preset screening rule. The operations of the feature map screening module 1430 can refer to the operations described above with reference to 730. Figure 7

[0166] ​​​​​The first shuffling module 1440 is configured to perform a first shuffling process on the target frequency component channel feature map to obtain a first shuffled feature map. The first normalization processing module 1450 is configured to perform normalization processing on the first shuffled feature map. The operations of the first shuffling module 1440 and the first normalization processing module 1450 can refer to the operations described above with reference to 740. Figure 7

[0167] The channel mixing processing module 1460 is configured to perform channel mixing processing on the first shuffled feature map after normalization processing. The operations of the channel mixing processing module 1460 can refer to the operations described above with reference to 750. Figure 7

[0168] The second shuffling module 1470 is configured to perform a second shuffling process on the first shuffled feature map after channel mixing processing to obtain a second shuffled feature map. The second normalization processing module 1480 is configured to perform normalization processing on the second shuffled feature map. The operations of the first shuffling module 1470 and the second normalization processing module 1480 can refer to the operations described above with reference to 760. Figure 7

[0169] Figure 15 An example block diagram of the image mixing processing unit 1500 according to an embodiment of the present specification is shown. As shown, the image mixing processing unit 1500 includes an image shuffling processing module 1510, an image super matrix construction module 1520, a weight coefficient generation module 1530, a weight coefficient normalization module 1540, and an image mixing processing module 1550. Figure 15

[0170] The image shuffling processing module 1510 is configured to perform k-1 times of shuffling processing on the image data set of the desensitized image data to obtain k image data sets. The operations of the image shuffling processing module 1510 can refer to the operations described above with reference to 1010. Figure 10

[0171] The image super matrix construction module 1520 is configured to construct an image super matrix with a size of m*k based on the obtained k image data sets, a first column in the constructed image super matrix corresponds to the original image data set, and m is the number of image data in the original subgraph data set. The operations of the image super matrix construction module 1520 can refer to the operations described above with reference to 1020. Figure 10

[0172] The weight coefficient generation module 1530 is configured to randomly generate a weight coefficient for each image data in the image super matrix. The operations of the weight coefficient generation module 1530 can refer to the operations described above with reference to 1030. Figure 10 ​​​​​​​

[0173] The weight coefficient normalization module 1540 is configured to perform row normalization on the weight coefficients of the image data in the image super matrix, so that the sum of the weight coefficients of each row of image data is 1, and the maximum coefficient of each image is not greater than W max The operation of the weight coefficient normalization module 1540 can refer to the operation described above with reference to 1040 of the image data normalization module 1040. Figure 10

[0174] The image mixing processing module 1550 is configured to perform weighted summation on each row of image data of the image super matrix to obtain a mixed image super matrix of size m*1, and the images in the obtained mixed image super matrix are the second desensitized image data after data enhancement. The operation of the image mixing processing module 1550 can refer to the operation described above with reference to 1050 of the image mixing processing module 1050. Figure 10

[0175] Figure 16 An example block diagram of the hyperparameter selection unit 1600 according to an embodiment of the present specification is shown. As shown, the hyperparameter selection unit 1600 includes an image data scaling module 1610, a feature extraction module 1620, a pooling processing module 1630, a selection probability determination module 1640, and a hyperparameter selection module 1650. Figure 16

[0176] The image data scaling module 1610 is configured to scale the image size of the first desensitized image data to the original image size of the training sample image data. The operation of the image data scaling module 1610 can refer to the operation described above with reference to 910 of the image data scaling module 910. Figure 9

[0177] The feature extraction module 1620 is configured to provide the scaled first desensitized image data to the feature extraction layer of the hyperparameter selection model to extract the feature map representation of the first desensitized image data. The operation of the feature extraction module 1620 can refer to the operation described above with reference to 920 of the feature extraction module 920. Figure 9

[0178] The pooling processing module 1630 is configured to perform pooling processing on the obtained feature map representation of the first desensitized image data, thereby pooling the image number dimension b of the feature map representation from batchsize dimension to 1 dimension. The operation of the pooling processing module 1630 can refer to the operation described above with reference to 930 of the pooling processing module 930. Figure 9

[0179] The selection probability determination module 1640 is configured to provide the pooling result to the fully connected layer of the hyperparameter selection model to obtain the selection probability of each candidate hyperparameter in the candidate hyperparameter set. The operation of the selection probability determination module 1640 can refer to the operation described above with reference to 1040 of the image data normalization module 1040. Figure 9 ​​​​​​the operations described above with reference to 940.

[0180] The hyperparameter selection module 1650 is configured to select the second hyperparameter based on the selection probability of each candidate hyperparameter. The operations of the hyperparameter selection module 1650 can refer to the operations described above with reference to 950. Figure 9

[0181] Figure 17 An example block diagram of the hyperparameter selection model updating unit 1700 according to an embodiment of the present specification is shown. As shown, the hyperparameter selection model updating unit 1700 includes a hyperparameter selection module 1710, an image mixing processing module 1720, a first loss function determination module 1730, a first model updating module 1740, a second loss function determination module 1750, a third loss function determination module 1760, and a second model updating module 1770. Figure 17

[0182] The hyperparameter selection module 1710 is configured to provide the first desensitized image data to the hyperparameter selection model to select the second hyperparameter for indicating the number of images participating in the image mixing processing from the candidate hyperparameter set. The operations of the hyperparameter selection module 1710 can refer to the operations described above with reference to 810. Figure 8

[0183] The image mixing processing module 1720 is configured to perform the data augmentation based image mixing processing on the first desensitized image data using the second hyperparameter to obtain the third desensitized image data and the corresponding label mixed labeled data. The operations of the image mixing processing module 1720 can refer to the operations described above with reference to 820. Figure 8

[0184] The first loss function determination module 1730 is configured to determine the first loss function using the third desensitized image data, the corresponding label mixed labeled data, and the current image recognition model. The operations of the first loss function determination module 1730 can refer to the operations described above with reference to 830. Figure 8

[0185] The first model updating module 1740 is configured to update the current image recognition model according to the first loss function. The operations of the first model updating module 1740 can refer to the operations described above with reference to 840. Figure 8

[0186] The second loss function determination module 1750 is configured to determine the second loss function using the third desensitized image data, the corresponding label mixed labeled data, and the updated image recognition model. The operations of the second loss function determination module 1750 can refer to the operations described above with reference to 950. Figure 8 ​​​​​​the operations described above with reference to 840 of FIG. 8.

[0187] The third loss function determination module 1760 is configured to determine the third loss function based on the first loss function and the second loss function. The operations of the third loss function determination module 1760 can refer to the operations described above with reference to 850 of FIG. 8. Figure 8

[0188] The second model update module 1770 is configured to update the model parameters of the hyperparameter selection model based on the third loss function. The operations of the second model update module 1770 can refer to the operations described above with reference to 850 of FIG. 8. Figure 8

[0189] Figure 18 An example block diagram of an image recognition model training apparatus 1800 of a second member device according to an embodiment of the present specification is shown. As Figure 18 indicated, the image recognition model training apparatus 1800 includes a model training result receiving unit 1810, a model updating unit 1820, and a model sending unit 1830.

[0190] The model training result receiving unit 1810 is configured to receive the model training result of the image recognition model from each first member device.

[0191] The model updating unit 1820 is configured to update the current image recognition model using the model training result of the current image recognition model received from each first member device.

[0192] The model sending unit 1830 is configured to send the updated image recognition model to each first member device for local model training.

[0193] As described above with reference to Figures 1 to 18 , the image recognition model training method and the image recognition model training apparatus according to an embodiment of the present specification are described. The image recognition model training apparatus above can be implemented by hardware, or by software or a combination of hardware and software.

[0194] Figure 19 An example schematic diagram of an image recognition model training apparatus 1900 implemented based on a computer system according to an embodiment of the present specification is shown. As Figure 19 ​​As shown, the image recognition model training apparatus 1900 can include at least one processor 1910, a memory (e.g., a non-volatile memory) 1920, an internal memory 1930, and a communication interface 1940, and the at least one processor 1910, the memory 1920, the internal memory 1930, and the communication interface 1940 are connected together via a bus 1960. The at least one processor 1910 executes at least one computer-readable instruction (i.e., the above-mentioned elements implemented in the form of software) stored or encoded in the memory.

[0195] In one embodiment, computer-executable instructions are stored in the memory, which when executed cause the at least one processor 1910 to: cyclically perform the following model training process until a model training end condition is met: obtain current training sample image data and its labeled label data; perform frequency domain transformation-based data desensitization processing on the current training sample image data to obtain first desensitized image data of the current training sample image data; provide the first desensitized image data to the hyperparameter selection model to select a first hyperparameter for indicating the number of images participating in the image mixing processing from the candidate hyperparameter set; perform Mixup data enhancement-based image mixing processing on the first desensitized image data using the first hyperparameter to obtain second desensitized image data and corresponding labeled label data after label mixing processing; train the current image recognition model using the second desensitized image data and the corresponding labeled label data after label mixing processing; and provide a model training result of the current image recognition model to a second member device for maintaining the image recognition model, so that the second member device uses the model training results from a plurality of first member devices to perform image recognition model updating, and receives an updated image recognition model from the second member device for use in the next round of image recognition model training.

[0196] It should be understood that the computer-executable instructions stored in the memory, when executed, cause the at least one processor 1910 to perform the various operations and functions described above in conjunction with various embodiments of the present specification. Figures 1-18

[0197] According to one embodiment, a program product such as a machine-readable medium (e.g., a non-transitory machine-readable medium) is provided. The machine-readable medium can have instructions (i.e., the above-mentioned elements implemented in the form of software) stored thereon, which when executed by a machine, cause the machine to perform the various operations and functions described above in conjunction with various embodiments of the present specification. Specifically, a system or apparatus equipped with a readable storage medium on which software program codes implementing the functions of any of the above-mentioned embodiments are stored, and causing the computer or processor of the system or apparatus to read out and execute the instructions stored in the readable storage medium can be provided. Figures 1-18

[0198] ​​In this case, the program code itself read from the readable medium can implement the functionality of any of the above-mentioned embodiments, so that the machine-readable code and the readable storage medium storing the machine-readable code form part of the invention.

[0199] Embodiments of the readable storage medium include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROMs, CD-Rs, CD-RWs, DVD-ROMs, DVD-RAMs, DVD-RWs, DVD-RWs), magnetic tapes, non-volatile memory cards and ROMs. Alternatively, the program code can be downloaded from a server computer or cloud over a communication network.

[0200] According to an embodiment, a computer program product is provided, the computer program product comprising a computer program which, when executed by a processor, causes the processor to perform the various embodiments of the above-described operations and functions in conjunction with Figures 1-18 the various embodiments of the present specification.

[0201] Those skilled in the art will understand that the various embodiments disclosed above can be variously modified and altered without departing from the spirit of the invention. Therefore, the scope of the present invention should be defined by the appended claims.

[0202] It should be noted that not all steps and units in the above-described flowcharts and system block diagrams are necessary, and some steps or units can be omitted according to actual needs. The execution order of each step is not fixed and can be determined as needed. The device structure described in the above embodiments can be a physical structure or a logical structure, i.e., some units can be implemented by the same physical entity, or some units can be implemented by multiple physical entities, or some units can be implemented by some components in multiple independent devices.

[0203] In the above embodiments, a hardware unit or module can be implemented by mechanical or electrical means. For example, a hardware unit, module or processor can include permanent dedicated circuitry or logic (such as a dedicated processor, FPGA or ASIC) to complete the corresponding operation. The hardware unit or processor can also include programmable logic or circuitry (such as a general-purpose processor or other programmable processor), which can be temporarily set by software to complete the corresponding operation. The specific implementation method (mechanical method, or dedicated permanent circuit, or temporarily set circuit) can be determined based on cost and time considerations.

[0204] The detailed description set forth above describes exemplary embodiments but is not intended to represent the only embodiments in which the claimant can realize and use the present disclosure. The term "exemplary" is used herein to mean "serving as an example, instance, or illustration," and not to imply that a described embodiment is "preferred" or "having some other qualitative advantage" over other embodiments. Specific details are given in the description to provide a thorough understanding of the embodiments presented herein. However, well-known structures and devices, as well as that which is in accordance with the present embodiments, are shown in block diagram form in order to avoid obscuring the concepts of the described embodiments. The above description is provided in the context of the accompanying drawings, which show the embodiments by way of example, and are not meant to limit the scope of the disclosure to the specific embodiments described. The terms used throughout this description are for the purpose of describing particular embodiments and are not intended to limit this disclosure from other embodiments, which are within the scope of the appended claims. The description of the embodiments is not meant to be taken in a normative sense, nor is it intended to be bound by the specific examples and designs described herein. Rather, the present disclosure is intended to be construed as consistent with the broadest scope of the principles and novel features disclosed herein.

[0205] The above description of the present disclosure is provided to enable any person skilled in the art to practice or use the present disclosure. Various modifications to the present disclosure will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other variations without departing from the scope of the disclosure. Thus, the present disclosure is not intended to be limited to the examples described herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for training an image recognition model, the method being performed by a first member device having local training data, the method comprising: cyclically performing a model training process until a model training end condition is met, the model training process comprising: obtaining current training sample image data and its labeled label data; performing frequency domain transformation based data sanitization on the current training sample image data to obtain first sanitized image data of the current training sample image data; providing the first sanitized image data to a hyperparameter selection model to select a first hyperparameter from a candidate hyperparameter set, the first hyperparameter being used to indicate a number of images participating in image mixing processing; performing Mixup data augmentation based image mixing processing on the first sanitized image data using the first hyperparameter to obtain second sanitized image data and corresponding label mixing processed labeled label data; training a current image recognition model using the second sanitized image data and the corresponding label mixing processed labeled label data; and providing a model training result of the current image recognition model to a second member device for maintaining image recognition models, for the second member device to perform image recognition model updating using model training results from a plurality of first member devices, and receiving the updated image recognition model from the second member device for next round image recognition model training. Before providing the first sanitized image data to the hyperparameter selection model to select a first hyperparameter from a candidate hyperparameter set, the first hyperparameter being used to indicate a number of images participating in image mixing processing, the method further comprises:

2. The method of claim 1, wherein, in response to a first predetermined condition being met, updating the hyperparameter selection model through a model updating process comprising: providing the first sanitized image data to the hyperparameter selection model to select a second hyperparameter from the candidate hyperparameter set, the second hyperparameter being used to indicate a number of images participating in image mixing processing; performing data augmentation based image mixing processing on the first sanitized image data using the second hyperparameter to obtain third sanitized image data and corresponding label mixing processed labeled label data; providing the third sanitized image data to the current image recognition model to obtain second predicted label data of the third sanitized image data, and determining a first loss function according to the second predicted label data and the corresponding label mixing processed labeled label data; updating the current image recognition model according to the first loss function; providing the third sanitized image data to the updated current image recognition model to obtain third predicted label data of the third sanitized image data, and determining a second loss function according to the third predicted label data and the corresponding label mixing processed labeled label data; determining a third loss function based on the first loss function and the second loss function; and updating model parameters of the hyperparameter selection model based on the third loss function. the first predetermined condition comprises:

3. The method of claim 2, wherein, a round number interval between a current training round number of the image recognition model and a training round number at a last updating processing time of the hyperparameter selection model reaches a first specified round number. ​ 4. The method of claim 1, wherein, providing the model training result of the current image recognition model to the second member device comprises: in response to a second predetermined condition being met, providing the model training result of the current image recognition model to the second member device.

5. The method of claim 4, wherein, the second predetermined condition comprises: a round number interval between a current training round number of the image recognition model and a training round number when the model training result was last sent reaches a second specified round number.

6. The method of claim 1, wherein, The first super parameter is k, and the maximum weight coefficient of the mixed image is W max , performing data augmentation-based image mixing processing on the first desensitized image data using the first hyperparameters comprises: performing k-1 times of shuffle processing on a set of image data of the first desensitized image data to obtain k sets of image data; based on the k sets of image data, constructing an image super matrix of size m*k, a first column in the image super matrix corresponding to an original set of image data of the first desensitized image data, and m being a number of image data in the original set of image data; randomly generating a weight coefficient for each image data in the image super matrix; weight coefficients of the image data in the image super matrix are row normalized such that a sum of the weight coefficients of each row of image data is 1 and the weight coefficient of each image data is not greater than W max ; and performing weighted summation on each row of image data of the image super matrix to obtain a mixed image super matrix of size m*1, the image data in the mixed image super matrix being desensitized image data after data augmentation.

7. The method of claim 1, wherein, performing frequency domain transformation-based data desensitization processing on the current training sample image data comprises: performing local frequency domain transformation processing on the current training sample image data to obtain at least one feature map, each feature map including a plurality of elements and corresponding to a data block in the current training sample image data, each element corresponding to a frequency in the frequency domain; using the elements corresponding to each frequency in the at least one feature map, respectively, to construct a frequency component channel feature map corresponding to each frequency; and selecting at least one target frequency component channel feature map from the constructed frequency component channel feature maps to obtain desensitized image data of the current training sample image data, the selected target frequency component channel feature map being a key channel feature for image recognition.

8. The method of claim 7, wherein, after selecting at least one target frequency component channel feature map from the constructed frequency component channel feature maps, the method further comprises: performing a first shuffling processing on the target frequency component channel feature map to obtain a first shuffled feature map; and performing normalization processing on the first shuffled feature map to obtain desensitized image data of the current training sample image data.

9. The method of claim 8, wherein, after performing normalization processing on the first shuffled feature map, the method further comprises: performing channel mixing processing on the first shuffled feature map after normalization processing; performing a second shuffling processing on the first shuffled feature map after channel mixing processing to obtain a second shuffled feature map; and performing normalization processing on the second shuffled feature map to obtain desensitized image data of the current training sample image data.

10. A method for training an image recognition model, the method being performed by at least two first member devices having local training data and a second member device for maintaining an image recognition model, the method comprising: recursively performing a model training process until a model training end condition is met: At each first member device, a current image recognition model is trained locally using local training sample image data according to the method of any one of claims 1 to 9; and At the second member device, model training results of the current image recognition model received from each first member device are aggregated to update the current image recognition model, and the updated image recognition model is sent to each first member device for local model training.

11. An apparatus for training an image recognition model, the apparatus being applied to a first member device having local training data for execution, the apparatus comprising: an image recognition model receiving unit configured to obtain a current image recognition model from a second member device for maintaining an image recognition model; a training sample data obtaining unit configured to obtain current training sample image data and its labeled label data; a data de-sensitization processing unit configured to perform data de-sensitization processing on the current training sample image data based on frequency domain transformation to obtain first de-sensitized image data of the current training sample image data; a hyper-parameter selection unit configured to provide the first de-sensitized image data to a hyper-parameter selection model to select a first hyper-parameter for indicating a number of images participating in image mixing processing from a candidate hyper-parameter set; an image mixing processing unit configured to perform image mixing processing on the first de-sensitized image data based on Mixup data augmentation using the first hyper-parameter to obtain second de-sensitized image data and corresponding labeled label data after label mixing processing; a model training unit configured to train the current image recognition model using the second de-sensitized image data and the corresponding labeled label data after label mixing processing; and a model training result sending unit configured to send model training results of the current image recognition model to the second member device for updating the current image recognition model by the second member device, wherein the image recognition model receiving unit, the training sample data obtaining unit, the data de-sensitization processing unit, the hyper-parameter selection unit, the image mixing processing unit, the model training unit, and the model training result sending unit perform the following operations in a loop until a model training end condition is met.

12. The apparatus of claim 11, further comprising: a hyper-parameter selection model updating unit configured to update the hyper-parameter selection model through the following model updating process in response to a first predetermined condition being met: providing the first de-sensitized image data to the hyper-parameter selection model to select a second hyper-parameter for indicating a number of images participating in image mixing processing from the candidate hyper-parameter set; performing image mixing processing on the first de-sensitized image data based on data augmentation using the second hyper-parameter to obtain third de-sensitized image data and corresponding labeled label data after label mixing processing; providing the third de-sensitized image data to the current image recognition model to obtain second predicted label data of the third de-sensitized image data, and determining a first loss function according to the second predicted label data and the corresponding labeled label data after label mixing processing; and providing the first loss function to the hyper-parameter selection model to update the hyper-parameter selection model. updating the current image recognition model according to the first loss function; providing the third desensitized image data to the updated image recognition model to obtain third predicted label data of the third desensitized image data, and determining a second loss function according to the third predicted label data and the corresponding labeled label data after the label mixing processing; determining a third loss function based on the first loss function and the second loss function; and updating the model parameters of the hyperparameter selection model based on the third loss function.

13. A system for training an image recognition model, comprising: at least two first member devices, each first member device having local training sample image data and comprising the apparatus for training an image recognition model according to claim 11 or 12; and a second member device, the second member device maintaining an image recognition model, and the second member device comprising: a model training result receiving unit receiving model training results of the image recognition model from each first member device; a model updating unit updating the image recognition model using the model training results of the image recognition model received from each first member device; and a model sending unit sending the updated image recognition model to each first member device for local model training.

14. An apparatus for training an image recognition model, comprising: at least one processor, a memory coupled with the at least one processor, and a computer program stored in the memory, the at least one processor executing the computer program to implement the method for training an image recognition model according to any one of claims 1 to 9.

15. An apparatus for training an image recognition model, comprising: at least one processor, a memory coupled with the at least one processor, and a computer program stored in the memory, the at least one processor executing the computer program to implement the method for training an image recognition model according to claim 10.

16. A computer-readable storage medium storing executable instructions, the instructions, when executed, causing a processor to perform the method for training an image recognition model according to any one of claims 1 to 9 or perform the method for training an image recognition model according to claim 10.

17. A computer program product comprising a computer program, the computer program being executed by a processor to implement the method for training an image recognition model according to any one of claims 1 to 9 or implement the method for training an image recognition model according to claim 10.

Images