Image classification model training method and device

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an image classification model training method and device. The method comprises the following steps: decrypting a first feature matrix convolution value set sent by at least one device to obtain a second feature matrix convolution value set; screening the second feature matrix convolution value set by using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set; according to at least one second feature matrix convolution value in the third feature matrix convolution value set and the at least one feature matrix, acquiring variance information corresponding to the third feature matrix convolution value set, and acquiring a target convolution kernel corresponding to the third feature matrix convolution value set; and training an initial image classification model according to the target convolution kernel to obtain a target image classification model. According to the invention, the accuracy of model training can be improved, and the image classification accuracy of the image classification model is improved.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to a training method and device for an image classification model. Background Art

[0002] With the development of science and technology, image recognition plays an increasingly important role in users' lives. Image recognition is a technology that uses computers to process, analyze, and understand images to identify targets and objects of various different patterns. It is an important application practice of artificial intelligence. Image classification is an image processing method that distinguishes targets of different categories based on their different features reflected in the image information. For example, an image classification algorithm can be used to train a model to classify images. However, this method has restrictions on image classification scenarios, resulting in poor accuracy in image classification. Summary of the invention

[0003] The present disclosure provides a method and device for training an image classification model, which can improve the accuracy of model training and improve the accuracy of image classification performed by the image classification model. The technical solution of the present disclosure is as follows:

[0004] According to a first aspect of an embodiment of the present disclosure, a method for training an image classification model is provided, comprising:

[0005] Decrypting a first feature matrix convolution value set sent by at least one device to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for the feature matrix corresponding to the image with the classification label;

[0006] Using a feature matrix convolution value anomaly recognition algorithm to screen the second feature matrix convolution value set to obtain a third feature matrix convolution value set;

[0007] According to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix, obtaining variance information corresponding to the third characteristic matrix convolution value set, and obtaining a target convolution kernel corresponding to the third characteristic matrix convolution value set;

[0008] The initial image classification model is trained according to the target convolution kernel to obtain a target image classification model.

[0009] According to some embodiments, the adopting a feature matrix convolution value anomaly recognition algorithm to screen the second feature matrix convolution value set to obtain a third feature matrix convolution value set includes:

[0010] Using a feature matrix convolution value anomaly identification algorithm, obtain an anomaly coefficient corresponding to any first feature convolution value in the second feature matrix convolution value set;

[0011] When the abnormal coefficient is greater than the first coefficient threshold, determining any one of the characteristic matrix convolution values ​​as an abnormal coefficient, and removing any one of the first characteristic convolution values;

[0012] or

[0013] In the case where the abnormal coefficient is less than or equal to a second coefficient threshold, any one of the feature matrix convolution values ​​is added to a third feature matrix convolution value set, wherein the second coefficient threshold is less than or equal to the first coefficient threshold.

[0014] According to some embodiments, obtaining an abnormal coefficient corresponding to any first feature convolution value in the second feature matrix convolution value set includes:

[0015] Obtaining a first mean value of the feature matrix convolution values ​​corresponding to the second feature matrix convolution value set;

[0016] Obtaining the variance of the feature matrix convolution value corresponding to the second feature matrix convolution value set;

[0017] According to any feature matrix convolution value in the second feature matrix convolution value set, the mean value and the variance, an abnormal coefficient of the any feature matrix convolution value is obtained.

[0018] According to some embodiments, obtaining variance information corresponding to the third characteristic matrix convolution value set and obtaining a target convolution kernel corresponding to the third characteristic matrix convolution value set according to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set includes:

[0019] Obtain a first feature matrix corresponding to a target feature convolution kernel from the at least one feature matrix, and obtain an eigenvalue corresponding to the first feature matrix, wherein the eigenvalue is an eigenvalue to be optimized;

[0020] Obtaining a second feature matrix of each image relative to the target feature convolution kernel, wherein the second feature matrix is ​​any feature matrix of the at least one feature matrix except the first feature matrix;

[0021] Obtaining variance information corresponding to the convolution value set of the third characteristic matrix according to the first characteristic matrix and the second characteristic matrix;

[0022] According to the variance information, the square function and the derivative information corresponding to each parameter, a target convolution kernel corresponding to the third feature matrix convolution value set is obtained.

[0023] According to some embodiments, obtaining variance information corresponding to the convolution value set of the third characteristic matrix according to the first characteristic matrix and the second characteristic matrix includes:

[0024] Obtaining characteristic matrix convolution values ​​corresponding to each second characteristic matrix in at least one second characteristic matrix;

[0025] Obtaining the square function corresponding to each parameter in the first feature matrix and the derivative information corresponding to each parameter;

[0026] According to the second mean values ​​corresponding to the characteristic matrix convolution values ​​corresponding to the second characteristic matrices;

[0027] The feature matrix convolution values ​​corresponding to the second feature matrices, the second mean value, and the number of received images with classification labels are obtained, and the variance information corresponding to the third feature matrix convolution value set is obtained.

[0028] According to a second aspect of an embodiment of the present disclosure, a method for training an image classification model is provided, comprising:

[0029] Get a collection of images with classification labels;

[0030] Performing feature extraction on each image in the image set to obtain a matrix corresponding to each image;

[0031] Performing a convolution operation on a convolution kernel corresponding to the classification information of each image and a matrix corresponding to the image to obtain a convolution matrix corresponding to each image;

[0032] According to the convolution matrix, the first feature matrix convolution value and the feature matrix corresponding to each image are obtained, and the encrypted first feature matrix convolution value and the encrypted feature matrix are sent to the network device, wherein the encrypted first feature matrix convolution value and the encrypted feature matrix are used to instruct the network device to train an image classification model.

[0033] According to some embodiments, obtaining the first feature matrix convolution value and the feature matrix corresponding to each image according to the convolution matrix includes:

[0034] Taking the maximum element value corresponding to the convolution matrix corresponding to each image as the first feature matrix convolution value of each image;

[0035] A convolution matrix participating in the convolution operation and corresponding to the convolution value of the first feature matrix is ​​used as the feature matrix corresponding to each image.

[0036] According to a third aspect of an embodiment of the present disclosure, there is provided a training device for an image classification model, comprising:

[0037] A first set acquisition unit is used to decrypt a first feature matrix convolution value set sent by at least one device to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for the feature matrix corresponding to the image with the classification label;

[0038] The first set acquisition unit is further used to screen the second feature matrix convolution value set by using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set;

[0039] An information acquisition unit, configured to acquire variance information corresponding to the third characteristic matrix convolution value set and a target convolution kernel corresponding to the third characteristic matrix convolution value set according to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix;

[0040] The model training unit is used to train the initial image classification model according to the target convolution kernel to obtain the target image classification model.

[0041] According to a fourth aspect of an embodiment of the present disclosure, a training device for an image classification model is provided, comprising:

[0042] A second set acquisition unit is used to acquire an image set with classification labels;

[0043] A matrix acquisition unit, used to extract features from each image in the image set and acquire a matrix corresponding to each image;

[0044] The matrix acquisition unit is further used to perform a convolution operation on the convolution kernel corresponding to the classification information of each image and the matrix corresponding to the image to obtain the convolution matrix corresponding to each image;

[0045] A matrix sending unit is used to obtain the first feature matrix convolution value and the feature matrix corresponding to each image according to the convolution matrix, and send the encrypted first feature matrix convolution value and the encrypted feature matrix to the network device, wherein the encrypted first feature matrix convolution value and the encrypted feature matrix are used to instruct the network device to train an image classification model.

[0046] According to a fifth aspect of an embodiment of the present disclosure, a network device is provided, including:

[0047] processor;

[0048] a memory for storing instructions executable by the processor;

[0049] The processor is configured to execute the instructions to implement the training method of the image classification model described in any one of the aforementioned aspects.

[0050] According to a sixth aspect of an embodiment of the present disclosure, a training system for an image classification model is provided, comprising: a multi-party secure image classification server, a convolution kernel optimization distribution server, and at least one device, wherein:

[0051] Any one of the at least one device is used to obtain a set of images with classification labels;

[0052] Any of the devices is further used to extract features from each image in the image set to obtain a matrix corresponding to each image;

[0053] Any of the devices is further used to perform a convolution operation on a convolution kernel corresponding to the classification information of each image and a matrix corresponding to the image to obtain a convolution matrix corresponding to each image;

[0054] Any of the devices is further used to obtain the first feature matrix convolution value and the feature matrix corresponding to each image according to the convolution matrix, and send the encrypted first feature matrix convolution value and the encrypted feature matrix to the multi-party secure image classification server;

[0055] The multi-party secure image classification server is used to forward the encrypted first feature matrix convolution value and the encrypted feature matrix to the convolution kernel optimization distribution server;

[0056] The convolution kernel optimization distribution server is used to decrypt the first feature matrix convolution value set to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for the feature matrix corresponding to the image with the classification label;

[0057] The convolution kernel optimization distribution server is used to screen the second feature matrix convolution value set by using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set;

[0058] The convolution kernel optimization distribution server is used to obtain variance information corresponding to the third characteristic matrix convolution value set and obtain a target convolution kernel corresponding to the third characteristic matrix convolution value set according to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix;

[0059] The convolution kernel optimization distribution server is used to send the target convolution kernel to the multi-party secure image classification server; the multi-party secure image classification server is used to train the initial image classification model according to the target convolution kernel to obtain the target image classification model.

[0060] According to the seventh aspect of an embodiment of the present disclosure, a storage medium is provided. When instructions in the storage medium are executed by a processor of a network device, the network device is enabled to execute the training method of the image classification model described in any one of the preceding aspects.

[0061] According to an eighth aspect of an embodiment of the present disclosure, a computer program product is provided, including a computer program, wherein the computer program implements any one of the methods described in the preceding aspects when executed by a processor.

[0062] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0063] In some or related embodiments, a second feature matrix convolution value set is obtained by decrypting a first feature matrix convolution value set sent by at least one device, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for a feature matrix corresponding to an image with a classification label; the second feature matrix convolution value set is screened using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set; based on at least one second feature matrix convolution value in the third feature matrix convolution value set and the at least one feature matrix, variance information corresponding to the third feature matrix convolution value set is obtained, and a target convolution kernel corresponding to the third feature matrix convolution value set is obtained; based on the target convolution kernel, an initial image classification model is trained to obtain a target image classification model. Therefore, by receiving encrypted convolution values, the accuracy of convolution value transmission can be improved without leaking image details, and the image details cannot be inferred through the convolution values, which can improve the security of image transmission and reduce the probability of image leakage. Model training can be performed without sharing image details, and the convolution kernel can be dynamically iterated and optimized by calculation. The obtained target convolution kernel can be used for model training, which can improve the accuracy of model training and improve the accuracy of image classification by image classification models.

[0064] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0066] Figure 1 is a flowchart of a method for training an image classification model according to an exemplary embodiment;

[0067] Figure 2 is a flowchart of a method for training an image classification model according to an exemplary embodiment;

[0068] Figure 3 is a flowchart of a method for training an image classification model according to an exemplary embodiment;

[0069] Figure 4 is an exemplary schematic diagram of a training system for an image classification model according to an exemplary embodiment;

[0070] Figure 5 is a block diagram of a training device for an image classification model according to an exemplary embodiment;

[0071] Figure 6 is a block diagram of a training device for an image classification model according to an exemplary embodiment;

[0072] Figure 7 The invention is a block diagram of a network device according to an exemplary embodiment. DETAILED DESCRIPTION

[0073] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0074] The embodiments of the present disclosure propose a training method and device for an image classification model. In some embodiments, the training method for an image classification model can be replaced with the terms such as an information processing method and a communication method, the training device for an image classification model can be replaced with the terms such as an information processing device and a communication device, and the terms such as an information processing system and a communication system can be replaced with each other.

[0075] The embodiments of the present disclosure are not exhaustive, but are only illustrative of some embodiments, and are not intended to be a specific limitation on the scope of protection of the present disclosure. In the absence of contradiction, each step in a certain embodiment can be implemented as an independent embodiment, and the steps can be arbitrarily combined. For example, a solution after removing some steps in a certain embodiment can also be implemented as an independent embodiment, and the order of the steps in a certain embodiment can be arbitrarily exchanged. In addition, the optional implementation methods in a certain embodiment can be arbitrarily combined; in addition, the embodiments can be arbitrarily combined, for example, some or all of the steps of different embodiments can be arbitrarily combined, and a certain embodiment can be arbitrarily combined with the optional implementation methods of other embodiments.

[0076] In each embodiment of the present disclosure, unless otherwise specified or there is a logical conflict, the terms and / or descriptions between the embodiments are consistent and can be referenced to each other, and the technical features in different embodiments can be combined to form a new embodiment based on their internal logical relationships.

[0077] The terms used in the embodiments of the present disclosure are only for the purpose of describing specific embodiments and are not intended to limit the present disclosure.

[0078] In the embodiments of the present disclosure, unless otherwise specified, elements expressed in the singular form, such as "a", "an", "the", "above", "said", "aforementioned", "this", etc., may mean "one and only one", or "one or more", "at least one", etc. For example, when using articles such as "a", "an", "the" in English in translation, the noun after the article may be understood as a singular expression or a plural expression.

[0079] In the embodiments of the present disclosure, “plurality” refers to two or more.

[0080] In some embodiments, the terms “at least one,” “one or more,” “a plurality of,” “multiple,” etc. may be used interchangeably.

[0081] The prefixes such as "first" and "second" in the embodiments of the present disclosure are only used to distinguish different description objects, and do not constitute restrictions on the position, order, priority, quantity or content of the description objects. The statement of the description object refers to the description in the context of the claims or embodiments, and should not constitute unnecessary restrictions due to the use of prefixes. For example, if the description object is a "field", the ordinal number before the "field" in the "first field" and the "second field" does not limit the position or order between the "fields", and the "first" and "second" do not limit whether the "fields" they modify are in the same message, nor do they limit the order of the "first field" and the "second field". For another example, if the description object is a "level", the ordinal number before the "level" in the "first level" and the "second level" does not limit the priority between the "levels". For another example, the number of description objects is not limited by the ordinal number, and can be one or more. Taking the "first device" as an example, the number of "devices" can be one or more. In addition, the objects modified by different prefixes may be the same or different. For example, if the description object is "device", then the "first device" and the "second device" may be the same device or different devices, and their types may be the same or different. For another example, if the description object is "information", then the "first information" and the "second information" may be the same information or different information, and their contents may be the same or different.

[0082] In some embodiments, "terminal" or "terminal device" may be referred to as "user equipment (UE)", "user terminal" "mobile station (MS)", "mobile terminal (MT)", subscriber station, mobile unit, subscriber unit, wireless unit, remote unit, mobile device, wireless device, wireless communication device, remote device, mobile subscriber station, access terminal, mobile terminal, wireless terminal, remote terminal, handset, user agent, mobile client, client, etc.

[0083] In some embodiments, data, information, etc. may be obtained with the user's consent.

[0084] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0085] According to some embodiments, image classification can be, for example, an image processing method that distinguishes different categories of objects based on different features reflected in the image information. It uses a computer to perform qualitative analysis on the image and classify the image into one of several categories to replace human visual interpretation. Among them, the image classification method can be implemented with CNN and RESNET algorithms as the core, and specifically has the following steps:

[0086] 1. Obtain sample sets: Obtain image samples through web crawlers, system docking, etc., transmit image samples to image recognition servers, and mark sample data categories manually or semi-automatically.

[0087] 2. Training model: Input sample data into the centralized image classification server and train it with CNN and RESNET algorithms to obtain the image classification model.

[0088] 3. Sample classification: Input the sample images to be classified into the model trained in step 2 to obtain the classification results. Among them, some technologies can, for example, centrally store sample images and then use image classification algorithms to train the model. However, this model training method cannot fully meet all image classification scenarios. For example, if the sample images belong to different institutions and there are privacy and security restrictions between institutions, the sample images cannot be transmitted to the centralized image classification server. In this scenario, the image classification method of some embodiments will be powerless and unable to improve the accuracy of image classification.

[0089] Figure 1 is a flowchart of a method for training an image classification model according to an exemplary embodiment. Figure 1 As shown, the training method of the image classification model can be used in the training scenario of the image classification model, including the following steps:

[0090] In step S11, a first feature matrix convolution value set sent by at least one device is decrypted to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for the feature matrix corresponding to the image with the classification label;

[0091] According to some embodiments, the execution subject of the embodiment of the present disclosure may be, for example, a network device. The network device may include, for example, at least one device. The network device may also be referred to as a server, or may be referred to as a server cluster. For example, the network device may include a device of two devices. The network device does not specifically refer to a fixed device. When the number of devices included in the network device changes, the network device may also change accordingly. The network device of the embodiment of the present disclosure may include, for example, a multi-party secure image classification server and a convolution kernel optimization distribution server.

[0092] In some embodiments, at least one device may be used to indicate a device that sends an image, for example. Among them, at least one device is used to indicate the number of devices, for example. The at least one device does not specifically refer to a fixed device. For example, when any device identifier of at least one device changes, the device may also change accordingly. For example, when the type of image changes, the at least one device may also change accordingly. The device may be, for example, a terminal, and may also be a server that receives images sent by the terminal. This is not limited in the embodiments of the present disclosure. The device name is not limited. For example, the device may also be referred to as an image ownership device. Among them, the image in the embodiments of the present disclosure may also be referred to as an image.

[0093] According to some embodiments, the first one of the first feature matrix convolution value set is used to distinguish from the remaining feature matrix convolution value sets. The first feature matrix convolution value set does not specifically refer to a fixed set. For example, when the number of convolution values ​​corresponding to the first feature matrix convolution value set changes, the first feature matrix convolution value set may also change accordingly. For example, when any feature matrix convolution value in the first feature matrix convolution value set changes, the first feature matrix convolution value set may also change accordingly.

[0094] According to some embodiments, the first characteristic matrix convolution value may be, for example, a convolution value included in a first characteristic matrix convolution value set.

[0095] The feature matrix convolution value may be, for example, a convolution value obtained by a device calculating a feature matrix corresponding to an image with a classification label. Different devices may obtain different convolution values.

[0096] In some embodiments, the second of the second characteristic matrix convolution value set is used to distinguish from the remaining characteristic matrix convolution value sets. The second characteristic matrix convolution value set does not specifically refer to a fixed set. For example, when the number of convolution values ​​corresponding to the second characteristic matrix convolution value set changes, the second characteristic matrix convolution value set may also change accordingly. For example, when any characteristic matrix convolution value in the second characteristic matrix convolution value set changes, the second characteristic matrix convolution value set may also change accordingly.

[0097] According to some embodiments, the feature matrix may be, for example, an n-order matrix. The n may be, for example, greater than or equal to 1. The feature matrix may be, for example, obtained by extracting an image by a device. When the convolution kernel used by the device changes, the feature matrix may also change accordingly. The convolution kernel obtained by the device may be, for example, issued by a network device. The convolution kernel may be, for example, updated when the network device obtains the convolution value. In other words, the convolution kernel does not specifically refer to a fixed convolution kernel. For example, when the time point at which the convolution kernel is sent changes, the convolution kernel may also change accordingly.

[0098] In some embodiments, the image with the classification label may be, for example, an image obtained by pre-identifying the image. The classification label does not specifically refer to a fixed label. For example, when the label generation method changes, the classification label of the image may also change accordingly. For example, when the image changes, the classification label of the image may also change accordingly.

[0099] According to some embodiments, the convolution value may be, for example, a convolution value obtained by performing a convolution value generation operation on a feature matrix corresponding to the image. The convolution value does not specifically refer to a fixed convolution value. For example, when the convolution value generation method changes, the convolution value may also change accordingly.

[0100] In some embodiments, when the device obtains the feature matrix convolution value, the feature matrix convolution value can be encrypted and the encrypted feature matrix convolution value can be sent to the network device. The encryption method is not limited. The encryption method of the embodiment of the present disclosure can be, for example, homomorphic encryption.

[0101] According to some embodiments, a first feature matrix convolution value set sent by at least one device is decrypted to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for the feature matrix corresponding to an image with a classification label.

[0102] In step S12, the second feature matrix convolution value set is screened using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set;

[0103] In some embodiments, the feature matrix convolution value anomaly identification algorithm may be, for example, an algorithm for removing anomalies in feature matrix convolution values. The feature matrix convolution value anomaly identification algorithm may be, for example, an algorithm for identifying convolution values ​​that do not meet convolution value requirements in a feature matrix convolution value set. The feature matrix convolution value anomaly identification algorithm does not specifically refer to a fixed algorithm. For example, when the parameters corresponding to the feature matrix convolution value anomaly identification algorithm change, the feature matrix convolution value anomaly identification algorithm may also change accordingly.

[0104] According to some embodiments, the third feature matrix convolution value set may be, for example, a set obtained after screening the second feature matrix convolution value set. The third feature matrix convolution value set does not specifically refer to a fixed set. For example, when the feature matrix convolution value abnormality recognition algorithm changes, the third feature matrix convolution value may also change accordingly. For example, when the number of convolution values ​​corresponding to the third feature matrix convolution value set changes, the third feature matrix convolution value set may also change accordingly.

[0105] In some embodiments, a feature matrix convolution value anomaly identification algorithm may be used to screen the second feature matrix convolution value set to obtain a third feature matrix convolution value set.

[0106] In step S13, according to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix, variance information corresponding to the third characteristic matrix convolution value set is obtained, and a target convolution kernel corresponding to the third characteristic matrix convolution value set is obtained;

[0107] According to some embodiments, the second characteristic matrix convolution value may be, for example, a convolution value included in a third characteristic matrix convolution value set. The second of the second characteristic matrix convolution values ​​is used to distinguish from the remaining characteristic matrix convolution values ​​and does not specifically refer to a fixed characteristic matrix convolution value. For example, when the third characteristic matrix convolution value set changes, the second characteristic matrix convolution value may also change accordingly.

[0108] In some embodiments, the variance information may refer to the variance corresponding to the third feature matrix convolution value set, for example. The variance information does not specifically refer to a fixed information. For example, when the third feature matrix convolution value set changes, the variance information may also change accordingly.

[0109] According to some embodiments, the target convolution kernel may be, for example, a convolution kernel obtained based on variance information and used to train the initial image classification model. The target convolution kernel does not specifically refer to a fixed convolution kernel. For example, when the variance information changes, the target convolution kernel may also change accordingly.

[0110] In some embodiments, for example, based on at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix, the variance information corresponding to the third characteristic matrix convolution value set can be obtained, and the target convolution kernel corresponding to the third characteristic matrix convolution value set can be obtained.

[0111] In step S14, the initial image classification model is trained according to the target convolution kernel to obtain a target image classification model.

[0112] According to some embodiments, the initial image classification model may refer to an untrained image classification model, for example. The initial image classification model does not specifically refer to a fixed model. For example, when the model classification information corresponding to the initial image classification model changes, the initial image classification model may also change accordingly.

[0113] In some embodiments, the target image classification model can be, for example, a model that can train the initial image classification model and meet the image classification requirements. The target image classification model does not specifically refer to a fixed model. For example, when the image classification requirements change, the target image classification model can also change accordingly.

[0114] In some or related embodiments, a second feature matrix convolution value set is obtained by decrypting a first feature matrix convolution value set sent by at least one device, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for a feature matrix corresponding to an image with a classification label; the second feature matrix convolution value set is screened using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set; based on at least one second feature matrix convolution value in the third feature matrix convolution value set and the at least one feature matrix, variance information corresponding to the third feature matrix convolution value set is obtained, and a target convolution kernel corresponding to the third feature matrix convolution value set is obtained; based on the target convolution kernel, an initial image classification model is trained to obtain a target image classification model. Therefore, by receiving encrypted convolution values, the accuracy of convolution value transmission can be improved without leaking image details, thereby improving the security of image transmission and reducing the probability of image leakage. Model training can be performed without sharing image details, and the target convolution kernel can be determined by calculation. The target convolution kernel can be used for model training, thereby improving the accuracy of model training and improving the accuracy of image classification by image classification models.

[0115] Figure 2 is a flowchart of a method for training an image classification model according to an exemplary embodiment. Figure 2 As shown, the training method of the image classification model can be used in the training scenario of the image classification model, including the following steps:

[0116] In step S21, a first feature matrix convolution value set sent by at least one device is decrypted to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for the feature matrix corresponding to the image with the classification label;

[0117] The specific process is as described above and will not be repeated here.

[0118] In some embodiments, the first characteristic matrix convolution value includes at least one first characteristic matrix convolution value, and the at least one first characteristic matrix convolution value may be, for example, a convolution value obtained after the device is encrypted. The encryption may be, for example, homomorphic encryption.

[0119] Among them, when at least one device can, for example, send a first feature matrix convolution value set to a multi-party security classification server in an embodiment of the present disclosure.

[0120] In step S22, the second feature matrix convolution value set is screened using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set;

[0121] The specific process is as described above and will not be repeated here.

[0122] In order to ensure the accuracy of the image training model, the multi-party secure image classification server will perform anomaly detection on the reported convolution feature matrix values, thereby reducing the impact of abnormal data on the training model and improving the accuracy of the image classification model.

[0123] According to some embodiments, the adopting a feature matrix convolution value anomaly recognition algorithm to screen the second feature matrix convolution value set to obtain a third feature matrix convolution value set includes:

[0124] Using a feature matrix convolution value anomaly identification algorithm, obtain an anomaly coefficient corresponding to any first feature convolution value in the second feature matrix convolution value set;

[0125] When the abnormal coefficient is greater than the first coefficient threshold, determining any one of the characteristic matrix convolution values ​​as an abnormal coefficient, and removing any one of the first characteristic convolution values;

[0126] or

[0127] In the case where the abnormal coefficient is less than or equal to a second coefficient threshold, any one of the feature matrix convolution values ​​is added to a third feature matrix convolution value set, wherein the second coefficient threshold is less than or equal to the first coefficient threshold.

[0128] According to some embodiments, obtaining an abnormal coefficient corresponding to any first feature convolution value in the second feature matrix convolution value set includes:

[0129] Obtaining a first mean value of the feature matrix convolution values ​​corresponding to the second feature matrix convolution value set;

[0130] Obtaining the variance of the feature matrix convolution value corresponding to the second feature matrix convolution value set;

[0131] According to any feature matrix convolution value in the second feature matrix convolution value set, the mean value and the variance, an abnormal coefficient of the any feature matrix convolution value is obtained.

[0132] According to some embodiments, for example, all convolution feature matrices conform to the normal distribution, and outliers are data points distributed at the tails on both sides. The farther away from the center of the normal distribution, the higher the degree of data abnormality. The calculation method of the abnormal coefficient measurement indicator z is as shown in Formula 7:

[0133]

[0134] Among them, z i is the abnormal coefficient of the convolution value of each image feature matrix;

[0135] e i is the convolution value of the image feature matrix;

[0136] E all is the mean of the convolution values ​​of all image feature matrices;

[0137] σ all is the variance of all image feature matrix convolution values. Assuming that there are g image feature matrix convolution values ​​in the system as shown in Formula 8:

[0138]

[0139] When |z|>β (β can be set according to actual conditions), it is judged as abnormal data and this part of the data needs to be deleted; when the incremental gradient of |z|≤β is considered to be a valid feature matrix convolution value, it can be added to the third feature matrix convolution value set.

[0140] In step S23, a first feature matrix corresponding to the target feature convolution kernel in the at least one feature matrix is ​​obtained, and an eigenvalue corresponding to the first feature matrix is ​​obtained, wherein the eigenvalue is an eigenvalue to be optimized;

[0141] The specific process is as described above and will not be repeated here.

[0142] In step S24, a second feature matrix of each image relative to the target feature convolution kernel is obtained, wherein the second feature matrix is ​​any feature matrix of the at least one feature matrix except the first feature matrix;

[0143] The specific process is as described above and will not be repeated here.

[0144] In step S25, according to the first characteristic matrix and the second characteristic matrix, the variance information corresponding to the convolution value set of the third characteristic matrix is ​​obtained;

[0145] The specific process is as described above and will not be repeated here.

[0146] According to some embodiments, obtaining variance information corresponding to the convolution value set of the third characteristic matrix according to the first characteristic matrix and the second characteristic matrix includes:

[0147] Obtaining characteristic matrix convolution values ​​corresponding to each second characteristic matrix in at least one second characteristic matrix;

[0148] Obtaining the square function corresponding to each parameter in the first feature matrix and the derivative information corresponding to each parameter;

[0149] According to the second mean values ​​corresponding to the characteristic matrix convolution values ​​corresponding to the second characteristic matrices;

[0150] The feature matrix convolution values ​​corresponding to the second feature matrices, the second mean value, and the number of received images with classification labels are obtained, and the variance information corresponding to the third feature matrix convolution value set is obtained.

[0151] In step S26, a target convolution kernel corresponding to the third characteristic matrix convolution value set is obtained according to the variance information, the square function and the derivative information corresponding to each parameter;

[0152] According to some embodiments, in order to improve the accuracy of the classification model, the convolution kernel can be dynamically iterated without leaking the privacy of multiple image owners. The convolution kernel with the smallest variance information between all feature matrix values ​​is used as the objective function to calculate the optimal convolution kernel, i.e., the target convolution kernel. Since all devices of each party's image only upload the eigenvalues ​​of the convolution feature matrix when the multi-party image classification server collects data, and does not upload specific image details, the multi-party image classification server cannot infer the specific image.

[0153] According to some embodiments, a certain feature convolution kernel is c and is a matrix value to be optimized, as shown in Formula 9:

[0154]

[0155] Among them, all 9 parameters in c are dynamic values ​​to be optimized.

[0156] The feature matrix of each image is assumed to be K. The feature matrix of any image relative to the convolution kernel c can be expressed as shown in Formula 10:

[0157]

[0158] The objective function calculates C and K of each image according to Formula 9 i Convolution value E i , assuming there are g images in total,

[0159] E i=k i11 *c 11 +k i12 *c 12 +k i13 *c 13 +...+k i33 *c 33 ..............(11) Make all E i The variance between e The square is the smallest.

[0160] Min(σ e ) 2 ................................(12)

[0161] J=Min(σ e ) 2 ................................(13)

[0162] The calculation method of variance is based on formula 6:

[0163]

[0164] in is the mean of the convolution values ​​of g images, σ e 2 Here is the square function of the 9 parameters of C. The derivative of each parameter is calculated in turn, and the extreme value is obtained when the derivative is equal to zero:

[0165]

[0166] Through the above method, the optimal convolution kernel, i.e. the target convolution kernel, is calculated, and the new convolution kernel C is updated to participate in model training.

[0167] In step S27, the initial image classification model is trained according to the target convolution kernel to obtain a target image classification model.

[0168] The specific process is as described above and will not be repeated here.

[0169] According to some embodiments, based on the dynamically iteratively updated convolution kernel, i.e., the target convolution kernel, the initial image classification model can be trained using a support vector machine or a convolution network, or a classification algorithm. Since the optimized convolution kernel is used, the accuracy of the model can be improved when training the classification model.

[0170] According to some embodiments, the feature matrix convolution value abnormality identification algorithm is used to screen the second feature matrix convolution value set to obtain the third feature matrix convolution value set SVM, and the feature matrix convolution values ​​whose abnormal coefficients do not meet the requirements can be discarded, which can improve the accuracy of the convolution value set acquisition. In addition, according to the first feature matrix and the second feature matrix, the variance information corresponding to the third feature matrix convolution value set is obtained; according to the variance information, the square function and the derivative information corresponding to each parameter, the target convolution kernel corresponding to the third feature matrix convolution value set is obtained; according to the target convolution kernel, the initial image classification model is trained to obtain the target image classification model. Therefore, the target convolution kernel can be determined according to the variance information, the accuracy of convolution kernel acquisition can be improved, the accuracy of image classification model acquisition can be improved, and the accuracy of image classification can be improved.

[0171] Figure 3 is a flowchart of a method for training an image classification model according to an exemplary embodiment. Figure 3 As shown, the training method of the image classification model can be used in the training scenario of the image classification model, including the following steps:

[0172] In step S31, a set of images with classification labels is obtained;

[0173] According to some embodiments, the execution subject of the embodiment of the present disclosure may be, for example, a device. The device may be, for example, any one of the at least one device mentioned above.

[0174] In some embodiments, for example, an image set may be obtained, and each image in the image set may be classified to obtain an image set with classification labels. For example, an image set with classification labels may also be directly obtained. The image set does not specifically refer to a fixed set. For example, when the number of images corresponding to the image set changes, the image set may also change accordingly.

[0175] In some embodiments, for example, a set of images with classification labels may be obtained.

[0176] In step S32, feature extraction is performed on each image in the image set to obtain a matrix corresponding to each image;

[0177] In some embodiments, the disclosed embodiments do not limit the feature extraction method. For example, the Scale-Invariant Feature Transform (SIFT) can be used to extract features from each image in the image set to obtain a matrix corresponding to each image, the Histogram of Oriented Gradients (HOG) can be used to extract features from each image in the image set to obtain a matrix corresponding to each image, and a convolutional neural network model can be used to extract features from each image in the image set to obtain a matrix corresponding to each image.

[0178] Herein, feature extraction is performed on each image in the image set to obtain a matrix corresponding to each image.

[0179] In step S33, a convolution operation is performed on the convolution kernel corresponding to the classification information of each image and the matrix corresponding to the image to obtain the convolution matrix corresponding to each image;

[0180] According to some embodiments, after extracting features from images, a convolution kernel corresponding to classification information of each image may be obtained.

[0181] In some embodiments, the convolution matrix refers to a matrix obtained by performing a convolution operation on a convolution kernel and a matrix.

[0182] Among them, different images can correspond to different convolution matrices.

[0183] According to some embodiments, a convolution operation may be performed on a convolution kernel corresponding to the classification information of each image and a matrix corresponding to the image to obtain a convolution matrix corresponding to each image.

[0184] In step S34, according to the convolution matrix, the first feature matrix convolution value and the feature matrix corresponding to each image are obtained, and the encrypted first feature matrix convolution value and the encrypted feature matrix are sent to the network device, wherein the encrypted first feature matrix convolution value and the encrypted feature matrix are used to instruct the network device to train an image classification model.

[0185] According to some embodiments, based on the convolution matrix, the first feature matrix convolution value and the feature matrix corresponding to each image are obtained, and the encrypted first feature matrix convolution value and the encrypted feature matrix are sent to the network device, wherein the encrypted first feature matrix convolution value and the encrypted feature matrix are used to instruct the network device to train an image classification model.

[0186] According to some embodiments, obtaining the first feature matrix convolution value and the feature matrix corresponding to each image according to the convolution matrix includes:

[0187] Taking the maximum element value corresponding to the convolution matrix corresponding to each image as the first feature matrix convolution value of each image;

[0188] A convolution matrix participating in the convolution operation and corresponding to the convolution value of the first feature matrix is ​​used as the feature matrix corresponding to each image.

[0189] In some embodiments, for example, an image sample A with a classification label is abstracted as an n*m matrix, specifically as formula (1):

[0190]

[0191] Classify the image into convolution kernels Perform convolution operation with A to obtain the convolved matrix E, where E is a (n-2)*(m-2) dimensional matrix. The specific calculation method is formula 2 to 4:

[0192] e 11 =a 11 *c 11 +a 12 *c 12 +a 13 *c 13 +...+a 33 *c 33 ..........(2)

[0193] e 12 =a 12 *c 11 +a 13 *c 12 +a 14 *c 13 +...+a 34 *c 33 ..........(3) ...............

[0195] For any e ij have: e ij =a ij *c 11 +a i(j+1) *c 12 +a i(j+2) *c 13 +...+a (i+2)(j+2) *c 33 ...(4) The final convolution matrix formula for each image is 5:

[0196] Select the maximum element in the convolution matrix E of each image as the image feature convolution value, assuming e ij is the maximum element value in E, indicating that e ij The higher the similarity between the matrix involved in the convolution in the corresponding image A and c1, the more it is the feature matrix of the image. The objective function is formula 6:

[0197] e max =max(E)........................(6)

[0198] When the value is e max , the matrix corresponding to A participating in the convolution is the feature matrix of the image.

[0199] According to some embodiments, each device may calculate the image feature matrix and feature matrix convolution value of each image on its own server in turn. Assuming there are p devices in the system, and a server has q labeled images, the feature matrix convolution value e of each image is calculated in turn through the above steps.

[0200] Figure 4 is a flowchart of a training system for an image classification model according to an exemplary embodiment. Figure 4 As shown, the training system of the image classification model includes at least one device, a multi-party security image classifier, and a convolution kernel optimization distributor, wherein:

[0201] Any one of the at least one device is used to obtain a set of images with classification labels;

[0202] Any of the devices is further used to extract features from each image in the image set to obtain a matrix corresponding to each image;

[0203] Any of the devices is further used to perform a convolution operation on a convolution kernel corresponding to the classification information of each image and a matrix corresponding to the image to obtain a convolution matrix corresponding to each image;

[0204] Any of the devices is further used to obtain the first feature matrix convolution value and the feature matrix corresponding to each image according to the convolution matrix, and send the encrypted first feature matrix convolution value and the encrypted feature matrix to the multi-party secure image classification server;

[0205] The multi-party secure image classification server is used to forward the encrypted first feature matrix convolution value and the encrypted feature matrix to the convolution kernel optimization distribution server;

[0206] The convolution kernel optimization distribution server is used to decrypt the first feature matrix convolution value set to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for the feature matrix corresponding to the image with the classification label;

[0207] The convolution kernel optimization distribution server is used to screen the second feature matrix convolution value set by using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set;

[0208] The convolution kernel optimization distribution server is used to obtain variance information corresponding to the third characteristic matrix convolution value set and obtain a target convolution kernel corresponding to the third characteristic matrix convolution value set according to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix;

[0209] The convolution kernel optimization distribution server is used to send the target convolution kernel to the multi-party secure image classification server;

[0210] The multi-party secure image classification server is used to train the initial image classification model according to the target convolution kernel to obtain a target image classification model.

[0211] Among them, the specific execution process of each server or device is as described above and will not be repeated here.

[0212] According to some embodiments, the system of the disclosed embodiment further includes a secure unified authentication server, which can uniformly generate and manage the secret keys of the multi-party secure image recognition system to ensure the security of the entire system.

[0213] A block diagram of a training device for an image classification model according to an exemplary embodiment. Figure 5 , the device 500 comprises:

[0214] A first set acquisition unit 501 is configured to decrypt a first feature matrix convolution value set sent by at least one device to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for a feature matrix corresponding to an image with a classification label;

[0215] The first set acquisition unit 501 is further configured to screen the second feature matrix convolution value set using a feature matrix convolution value anomaly recognition algorithm to acquire a third feature matrix convolution value set;

[0216] An information acquisition unit 502 is used to acquire variance information corresponding to the third characteristic matrix convolution value set and acquire a target convolution kernel corresponding to the third characteristic matrix convolution value set according to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix;

[0217] The model training unit 503 is used to train the initial image classification model according to the target convolution kernel to obtain a target image classification model.

[0218] According to some embodiments, the first set acquisition unit 501 is used to screen the second feature matrix convolution value set using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set, specifically for:

[0219] Using a feature matrix convolution value anomaly identification algorithm, obtain an anomaly coefficient corresponding to any first feature convolution value in the second feature matrix convolution value set;

[0220] When the abnormal coefficient is greater than the first coefficient threshold, determining any one of the characteristic matrix convolution values ​​as an abnormal coefficient, and removing any one of the first characteristic convolution values;

[0221] or

[0222] In the case where the abnormal coefficient is less than or equal to a second coefficient threshold, any one of the feature matrix convolution values ​​is added to a third feature matrix convolution value set, wherein the second coefficient threshold is less than or equal to the first coefficient threshold.

[0223] According to some embodiments, when the first set acquisition unit 501 is used to acquire an abnormal coefficient corresponding to any first feature convolution value in the second feature matrix convolution value set, it is specifically used to:

[0224] Obtaining a first mean value of the feature matrix convolution values ​​corresponding to the second feature matrix convolution value set;

[0225] Obtaining the variance of the feature matrix convolution value corresponding to the second feature matrix convolution value set;

[0226] According to any feature matrix convolution value in the second feature matrix convolution value set, the mean value and the variance, an abnormal coefficient of the any feature matrix convolution value is obtained.

[0227] According to some embodiments, the information acquisition unit 502 is used to obtain variance information corresponding to the third characteristic matrix convolution value set according to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix, and when obtaining the target convolution kernel corresponding to the third characteristic matrix convolution value set, specifically for:

[0228] Obtain a first feature matrix corresponding to a target feature convolution kernel from the at least one feature matrix, and obtain an eigenvalue corresponding to the first feature matrix, wherein the eigenvalue is an eigenvalue to be optimized;

[0229] Obtaining a second feature matrix of each image relative to the target feature convolution kernel, wherein the second feature matrix is ​​any feature matrix of the at least one feature matrix except the first feature matrix;

[0230] Obtaining variance information corresponding to the convolution value set of the third characteristic matrix according to the first characteristic matrix and the second characteristic matrix;

[0231] According to the variance information, the square function and the derivative information corresponding to each parameter, a target convolution kernel corresponding to the third feature matrix convolution value set is obtained.

[0232] According to some embodiments, the information acquisition unit 502 is used to acquire the variance information corresponding to the convolution value set of the third characteristic matrix according to the first characteristic matrix and the second characteristic matrix, specifically to:

[0233] Obtaining characteristic matrix convolution values ​​corresponding to each second characteristic matrix in at least one second characteristic matrix;

[0234] Obtaining the square function corresponding to each parameter in the first feature matrix and the derivative information corresponding to each parameter;

[0235] According to the second mean values ​​corresponding to the characteristic matrix convolution values ​​corresponding to the second characteristic matrices;

[0236] The feature matrix convolution values ​​corresponding to the second feature matrices, the second mean value, and the number of received images with classification labels are obtained, and the variance information corresponding to the third feature matrix convolution value set is obtained.

[0237] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0238] In some or related embodiments, a first set acquisition unit is used to decrypt a first feature matrix convolution value set sent by at least one device to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for a feature matrix corresponding to an image with a classification label; the first set acquisition unit is also used to screen the second feature matrix convolution value set using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set; an information acquisition unit is used to obtain variance information corresponding to the third feature matrix convolution value set based on at least one second feature matrix convolution value in the third feature matrix convolution value set and the at least one feature matrix, and obtain a target convolution kernel corresponding to the third feature matrix convolution value set; a model training unit is used to train an initial image classification model based on the target convolution kernel to obtain a target image classification model. Therefore, by receiving encrypted convolution values, the accuracy of convolution value transmission can be improved without leaking image details, thereby improving the security of image transmission and reducing the probability of image leakage. Model training can be performed without sharing image details, and the target convolution kernel can be determined by calculation. The target convolution kernel can be used for model training, thereby improving the accuracy of model training and improving the accuracy of image classification by image classification models.

[0239] A block diagram of a training device for an image classification model according to an exemplary embodiment. Figure 6 , the device 600 comprises:

[0240] The second set acquisition unit 601 is used to acquire an image set with classification labels;

[0241] A matrix acquisition unit 602 is used to extract features from each image in the image set and acquire a matrix corresponding to each image;

[0242] The matrix acquisition unit 602 is further used to perform a convolution operation on the convolution kernel corresponding to the classification information of each image and the matrix corresponding to the image to obtain the convolution matrix corresponding to each image;

[0243] The matrix sending unit 603 is used to obtain the first feature matrix convolution value and the feature matrix corresponding to each image according to the convolution matrix, and send the encrypted first feature matrix convolution value and the encrypted feature matrix to the network device, wherein the encrypted first feature matrix convolution value and the encrypted feature matrix are used to instruct the network device to train the image classification model.

[0244] According to some embodiments, the matrix acquisition unit 602 is used to acquire the first feature matrix convolution value and the feature matrix corresponding to each image according to the convolution matrix, specifically to:

[0245] Taking the maximum element value corresponding to the convolution matrix corresponding to each image as the first feature matrix convolution value of each image;

[0246] A convolution matrix participating in the convolution operation and corresponding to the convolution value of the first feature matrix is ​​used as the feature matrix corresponding to each image.

[0247] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0248] Figure 7 700 is a block diagram of a network device 700 provided in an embodiment of the present disclosure. For example, the network device 700 may be provided as a network device. Figure 7 The network device 700 includes a processing component 722, which further includes at least one processor, and a memory resource represented by a memory 732 for storing instructions executable by the processing component 722, such as an application. The application stored in the memory 732 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 722 is configured to execute instructions to perform any of the methods described above in the network device.

[0249] The network device 700 may also include a power supply component 727 configured to perform power management of the network device 700, a wired or wireless network interface 750 configured to connect the network device 700 to a network, and an input / output (I / O) interface 758. The network device 700 may operate based on an operating system stored in the memory 732, such as Windows Server TM, Mac OS X TM, Unix TM, Linux TM, Free BSD TM or the like.

[0250] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0251] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0252] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0253] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0254] The systems and techniques described herein can be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), the Internet, and a blockchain network.

[0255] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services ("Virtual Private Server", or "VPS" for short). The server may also be a server of a distributed system, or a server combined with a blockchain.

[0256] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0257] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A training method for an image classification model, characterized in that: include: Decrypting a first feature matrix convolution value set sent by at least one device to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for the feature matrix corresponding to the image with the classification label; Using a feature matrix convolution value anomaly recognition algorithm to screen the second feature matrix convolution value set to obtain a third feature matrix convolution value set; According to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix, obtaining variance information corresponding to the third characteristic matrix convolution value set, and obtaining a target convolution kernel corresponding to the third characteristic matrix convolution value set; The initial image classification model is trained according to the target convolution kernel to obtain a target image classification model.

2. The method according to claim 1, characterized in that The adopting of the feature matrix convolution value abnormality recognition algorithm to screen the second feature matrix convolution value set to obtain a third feature matrix convolution value set includes: Using a feature matrix convolution value anomaly identification algorithm, obtain an anomaly coefficient corresponding to any first feature convolution value in the second feature matrix convolution value set; When the abnormal coefficient is greater than the first coefficient threshold, determining any one of the characteristic matrix convolution values ​​as an abnormal coefficient, and removing any one of the first characteristic convolution values; or In the case where the abnormal coefficient is less than or equal to a second coefficient threshold, any one of the feature matrix convolution values ​​is added to a third feature matrix convolution value set, wherein the second coefficient threshold is less than or equal to the first coefficient threshold.

3. The method according to claim 2, characterized in that The obtaining of an abnormal coefficient corresponding to any first characteristic convolution value in the second characteristic matrix convolution value set includes: Obtaining a first mean value of the feature matrix convolution values ​​corresponding to the second feature matrix convolution value set; Obtaining the variance of the feature matrix convolution value corresponding to the second feature matrix convolution value set; According to any feature matrix convolution value in the second feature matrix convolution value set, the mean value and the variance, an abnormal coefficient of the any feature matrix convolution value is obtained.

4. The method according to claim 1, characterized in that: The obtaining, according to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix, variance information corresponding to the third characteristic matrix convolution value set, and the obtaining a target convolution kernel corresponding to the third characteristic matrix convolution value set include: Obtain a first feature matrix corresponding to a target feature convolution kernel from the at least one feature matrix, and obtain an eigenvalue corresponding to the first feature matrix, wherein the eigenvalue is an eigenvalue to be optimized; Obtaining a second feature matrix of each image relative to the target feature convolution kernel, wherein the second feature matrix is ​​any feature matrix of the at least one feature matrix except the first feature matrix; Obtaining variance information corresponding to the convolution value set of the third characteristic matrix according to the first characteristic matrix and the second characteristic matrix; According to the variance information, the square function and the derivative information corresponding to each parameter, a target convolution kernel corresponding to the third feature matrix convolution value set is obtained.

5. The method according to claim 4, characterized in that The obtaining, according to the first characteristic matrix and the second characteristic matrix, variance information corresponding to the convolution value set of the third characteristic matrix includes: Obtaining characteristic matrix convolution values ​​corresponding to each second characteristic matrix in at least one second characteristic matrix; Obtaining the square function corresponding to each parameter in the first feature matrix and the derivative information corresponding to each parameter; According to the second mean values ​​corresponding to the characteristic matrix convolution values ​​corresponding to the second characteristic matrices; The feature matrix convolution values ​​corresponding to the second feature matrices, the second mean value, and the number of received images with classification labels are obtained, and the variance information corresponding to the third feature matrix convolution value set is obtained.

6. A training method for an image classification model, characterized in that: include: Get a collection of images with classification labels; Performing feature extraction on each image in the image set to obtain a matrix corresponding to each image; Performing a convolution operation on a convolution kernel corresponding to the classification information of each image and a matrix corresponding to the image to obtain a convolution matrix corresponding to each image; According to the convolution matrix, the first feature matrix convolution value and the feature matrix corresponding to each image are obtained, and the encrypted first feature matrix convolution value and the encrypted feature matrix are sent to the network device, wherein the encrypted first feature matrix convolution value and the encrypted feature matrix are used to instruct the network device to train an image classification model.

7. The method according to claim 6, characterized in that The step of obtaining the first feature matrix convolution value and the feature matrix corresponding to each image according to the convolution matrix includes: Taking the maximum element value corresponding to the convolution matrix corresponding to each image as the first feature matrix convolution value of each image; A convolution matrix participating in the convolution operation and corresponding to the convolution value of the first feature matrix is ​​used as the feature matrix corresponding to each image.

8. A training device for an image classification model, characterized in that: include: A first set acquisition unit is used to decrypt a first feature matrix convolution value set sent by at least one device to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for the feature matrix corresponding to the image with the classification label; The first set acquisition unit is further used to screen the second feature matrix convolution value set by using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set; An information acquisition unit, configured to acquire variance information corresponding to the third characteristic matrix convolution value set and a target convolution kernel corresponding to the third characteristic matrix convolution value set according to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix; The model training unit is used to train the initial image classification model according to the target convolution kernel to obtain the target image classification model.

9. A network device, characterized in that: include: A processor and a memory storing a computer program, wherein when the computer program is executed by the processor, the method according to any one of claims 1 to 5 or any one of claims 6 to 7 is executed.

10. A training system for an image classification model, characterized in that: include: A multi-party secure image classification server, a convolution kernel optimization distribution server, and at least one device, wherein: Any one of the at least one device is used to obtain a set of images with classification labels; Any of the devices is further used to extract features from each image in the image set to obtain a matrix corresponding to each image; Any of the devices is further used to perform a convolution operation on a convolution kernel corresponding to the classification information of each image and a matrix corresponding to the image to obtain a convolution matrix corresponding to each image; Any of the devices is further used to obtain the first feature matrix convolution value and the feature matrix corresponding to each image according to the convolution matrix, and send the encrypted first feature matrix convolution value and the encrypted feature matrix to the multi-party secure image classification server; The multi-party secure image classification server is used to forward the encrypted first feature matrix convolution value and the encrypted feature matrix to the convolution kernel optimization distribution server; The convolution kernel optimization distribution server is used to decrypt the first feature matrix convolution value set to obtain a second feature matrix convolution value set, wherein the first feature matrix convolution value set includes at least one first feature matrix convolution value and at least one feature matrix, and the at least one first feature matrix convolution value is a convolution value calculated by each device in the at least one device for the feature matrix corresponding to the image with the classification label; The convolution kernel optimization distribution server is used to screen the second feature matrix convolution value set by using a feature matrix convolution value anomaly recognition algorithm to obtain a third feature matrix convolution value set; The convolution kernel optimization distribution server is used to obtain variance information corresponding to the third characteristic matrix convolution value set and obtain a target convolution kernel corresponding to the third characteristic matrix convolution value set according to at least one second characteristic matrix convolution value in the third characteristic matrix convolution value set and the at least one characteristic matrix; The convolution kernel optimization distribution server is used to send the target convolution kernel to the multi-party secure image classification server; The multi-party secure image classification server is used to train the initial image classification model according to the target convolution kernel to obtain a target image classification model.

Images