A method and system for integrating live broadcast platform information

By freezing messaging, the model passes information and weighted average calculations on the client and server side, the low generalization and privacy and security problems in the information integration of live broadcast platforms are solved, and the stable and efficient integration of information is achieved.

CN118646907BActive Publication Date: 2024-11-15HANGZHOU DIMENSION ISLAND TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411110933.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-14
Publication Date
2024-11-15
Estimated Expiration
2044-08-14

AI Technical Summary

Technical Problem

In the process of information integration, existing live broadcast platforms face problems such as low generalization of model, heterogeneity of data, heterogeneity of system, and privacy and security, which leads to high difficulty and low efficiency of information integration.

Method used

The frozen messenger model is used to pass information on the client and server side, feature extraction and distillation are performed through the receiver and transmitter, and parameter weighted average calculation is performed until the model converges to achieve stable integration of information.

Benefits of technology

Effectively integrating client information reduces local training costs, reduces client drift, and improves the stability and efficiency of information integration.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118646907B_ABST
    Figure CN118646907B_ABST
Patent Text Reader

Abstract

A method and system for integrating information on a live broadcast platform, the method is initiated by a client, and the client transmits information to a server through a receiver and a transmitter, including freezing the first messenger model of the client, injecting the first messenger model into the first local model of the client through a receiver, and obtaining a second local model located at the client; freezing the second local model, distilling the information extracted by the second local model into the first messenger model through a transmitter, and obtaining a second messenger model located at the client; sending the second messenger model to the server; performing parameter weighted average calculation on the messenger model and the local model at the server; feeding back the result of the parameter weighted average calculation to the client, and repeating the above steps until the model converges to integrate the information. The traditional public data set is replaced by a lightweight model, information is transmitted between the client and the server, and a model is established to integrate information.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of information integration, and in particular relates to a live broadcast platform information integration method and system. Background Art

[0002] Live streaming has gradually become a mainstream lifestyle today. For example, live streaming can help users interact, compare products and shop on e-commerce live streaming platforms through portable terminals. Figure 1 As shown in the figure, when the e-commerce live broadcast platform communicates with multiple portable terminals, it is particularly important to protect the security of the user's privacy data. In order to improve the live broadcast skills of live broadcast experts and integrate the information of various platforms for convenient management, it is inevitable to involve the privacy data of multiple users, such as which live broadcasts a certain type of user prefers to watch and at what time; which live broadcasts have users purchased items in, which interactive methods are more effective in attracting traffic, etc. The integration of this information can help different live broadcast experts to increase their live broadcast traffic, and it is also helpful for live broadcast companies to integrate information management on various live broadcast platforms. Most of the existing methods for protecting user privacy data encrypt user privacy data by hashing values, but this method has poor reliability and is easily cracked by hackers, resulting in the leakage of user privacy data. In addition, the data obtained by each platform is different, the hardware equipment is also different, and the protocol is also different, which makes it difficult to integrate platform information.

[0003] Federated learning (FL) can effectively reduce privacy risks on the one hand, and make full use of the computing resources of edge devices on the other hand, which has very practical significance. However, unlike centralized training, federated learning training may need to consider the following limitations: data heterogeneity, system heterogeneity, communication overhead, privacy and security, etc. The two most important aspects are the heterogeneity of client models and the differences in local data. Since each client is not necessarily the same and is in a different space, this leads to great differences in the communication volume, computing power and data owned by each client, and these differences will seriously affect the quality of the model jointly trained by each client.

[0004] Knowledge Distillation (KD) is a technique used for model compression and knowledge transfer in deep learning. This method is widely used to reduce model complexity and reduce computing resource requirements. It is achieved by transferring the knowledge of a large complex model (teacher model) to a small simple model (student model) and using soft labels as the training target of the student model, so that the performance of the student model can be optimized. Distillation can be optimized by minimizing the cross entropy loss between the soft labels and the output of the student model. This method has achieved remarkable success on various tasks and datasets, including image classification, natural language processing, and speech recognition.

[0005] Chinese patent CN202110844739.0 discloses a heterogeneous model aggregation method and system based on federated learning, including initializing a neural network model; each client contributes part of the local data to upload to the server to form a shared data set, and trains a CGAN model; the client uses the local data set and the data set generated by the CGAN model to train the local model, predict each data in the shared data set and upload the prediction score to the server; the server calculates the degree of deviation of the prediction score of each client and uses the inverse of the calculation result as the weight to calculate the global prediction score, and uses the global prediction score to perform knowledge distillation on the server model; the client downloads the prediction scores of other client models from the server for cooperative training; the model converges after multiple iterations. However, each client needs to contribute part of the local data to upload to the server to form a shared data set, and this part of the local data involves privacy data, which makes this method difficult to implement, and the security protocols of each live broadcast platform are different, and if it is attacked during the data upload process, it will lead to the leakage of privacy data.

[0006] Chinese patent CN202111189306.2 provides a model training method and device based on a privacy data set. The method includes: training a server-side model based on a public data set and a real label corresponding to the public data set; obtaining a first model output sent by each client; the first model output is obtained by inputting the public data set into a local learning model; the local learning model is obtained by training based on the privacy data set and the corresponding label; training the server-side model based on the corresponding public data of each first model output; inputting the public data set into the server-side model to obtain a second model output; sending the second model output to each client, so that each client can retrain the local learning model based on the second model output and the public data set. In this way, under the premise of avoiding the leakage of the privacy data set, the model training is carried out based on knowledge distillation and knowledge fusion with the privacy data set as part of the training sample. Label knowledge refers to the label information obtained by predicting the data set through the trained model, which is also commonly referred to as "dark knowledge". The label knowledge method is simple and universal, easy to implement, and suitable for almost all tasks such as classification, recognition, and segmentation. However, label knowledge also has the disadvantages of single knowledge, dependence on the design of loss function, and sensitivity to parameters. In addition, label knowledge also contains a lot of uncertain information, which usually reflects the similarity or interference between samples and the difficulty of sample prediction. Therefore, label knowledge usually provides very limited information and has relative uncertainty, but it is still one of the key points and hot spots in the research of basic distillation methods, because it is closely related to traditional pseudo-label learning or self-training methods, which actually opens up a new path for semi-supervised learning. Label knowledge is one of the foundations of knowledge distillation in various tasks and is suitable for scenarios with relatively low security and privacy requirements. At present, in order to improve the effect of lightweight models in different scenarios, it is usually used to optimize the model training using the label data in the corresponding scenarios, so that each model optimization training will be subject to the label data in a specific scenario, which not only has high labeling costs, but also has very low model optimization efficiency. Summary of the invention

[0007] In view of the above-mentioned prior art, the purpose of the present invention is to provide a more stable live broadcast platform information integration method and system, aiming to solve the problems of information integration in existing live broadcast scenarios, low model generalization, and difficulty and low efficiency of information integration due to data heterogeneity, system heterogeneity, data privacy and security on various platforms.

[0008] In order to solve the above technical problems, the present application provides a method for integrating live broadcast platform information, including:

[0009] The method is initiated by a client, and the client transmits information with a server through a receiver and a transmitter, and is characterized in that:

[0010] Freeze the first messenger model of the client, and inject the first messenger model into the first local model of the client through a receiver to obtain a second local model located at the client;

[0011] Freeze the second local model, and distill the information extracted from the second local model to the first messenger model through a transmitter to obtain a second messenger model located at the client;

[0012] Sending the second messenger model to the server;

[0013] Performing weighted average calculation of parameters of the messenger model and the local model on the server side;

[0014] The result of the weighted average calculation of the parameters is fed back to the client, and the above steps are repeated until the model converges to integrate the information.

[0015] Preferably, injecting the first messenger model into the first local model of the client through a receiver comprises:

[0016] Obtain a first feature extractor and a second feature extractor;

[0017] Extracting a feature vector of an input in the first messenger model according to the first feature extractor to obtain a first classifier for a downstream task;

[0018] Instructing the second feature extractor to extract the feature vector of the weights in the first local model through the input feature vector in the first messenger model to obtain a second classifier for the downstream task;

[0019] Preferably, the step of distilling the information extracted from the second local model to the first messenger model through a transmitter comprises:

[0020] Obtaining a third feature extractor and a fourth feature extractor;

[0021] Extracting a feature vector of an input in the second local model according to the third feature extractor to obtain a third classifier for a downstream task;

[0022] The fourth feature extractor is guided by the input feature vector in the second local model to extract the feature vector of the weights in the first messenger model, so as to obtain a fourth classifier for the downstream task.

[0023] Preferably, the output logic of establishing the first messenger model to guide the first local model is:

[0024]

[0025] Where E represents the expectation, (X, Y) is the joint distribution of samples and features, and are the predictions of the first local model and the first messenger model for sample x, is the cross entropy function, is the divergence function, λ is a hot label y and soft label Hyperparameters that balance between .

[0026] Preferably, the output logic of establishing the second local model to guide the first messenger model is:

[0027]

[0028] Among them, E represents the expectation, (X, Y) is the joint distribution of samples and features, and are the features of the second local model and the first messenger model at the i-th layer, respectively. ψ is a feature projector that maps the features of the first messenger model to match the features of the second local model. ||XY|| represents the square root of the sum of the squares of the components after the vector difference. i is an integer greater than 0.

[0029] As a preference, a parameter weighted average function is established as:

[0030]

[0031] in , , They represent the initial feature vector, the feature vector and weight output by the feature extractor or classifier of the i-th client respectively, ||XY|| represents the square root of the sum of the squares of the components after the vector is subtracted, and i is an integer greater than 0.

[0032] Preferably, sending the second messenger model to the server side includes uploading the feature extractor and classifier in the second messenger model to the server side, wherein the feature extractor and classifier in the second messenger model contain feature vectors and weights required for parameter weight averaging.

[0033] In order to solve the above technical problems, the present application also provides a live broadcast platform information integration system, including:

[0034] Injection module: freezes the first messenger model of the client, and injects the first messenger model into the first local model of the client through a receiver to obtain a second local model located at the client;

[0035] Distillation module: freezing the second local model, and distilling the information extracted from the second local model to the first messenger model through the transmitter to obtain a second messenger model located at the client;

[0036] Upload module: send the second messenger model to the server;

[0037] Calculation module: performing parameter weighted average calculation on the messenger model and the local model on the server side;

[0038] Download module: Feedback the result of the weighted average calculation of the parameters to the client, repeat the above steps until the model converges, so as to integrate the information. Information integration includes image classification, natural language processing and speech recognition.

[0039] In order to solve the above technical problems, the present application also provides a live broadcast platform information integration device, including:

[0040] Memory for storing computer programs;

[0041] The processor is used to point to a computer program to implement the steps based on a live broadcast platform information integration processing method.

[0042] In order to solve the above technical problems, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, all the steps of the above-mentioned live broadcast platform information integration method are implemented.

[0043] The beneficial effects of this application are as follows:

[0044] 1. This application designs a messenger model as a lightweight information transfer model to transmit information between clients and servers, servers and clients, and clients and clients in multiple rounds. The messenger model can effectively integrate information from each client without increasing local training costs. The client transfers its knowledge and capabilities to the server through the local model and the messenger model through lightweight continuous training, and vice versa, the server transfers its knowledge and capabilities to each client through the messenger model through lightweight continuous training.

[0045] 2. This application designs a set of transmitter and receiver modules for each client to separate local deviations from general information. This separation effectively reduces the deviation information collected from each client, thereby solving and reducing the problem of client drift. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 To indicate the information source of each client;

[0048] Figure 2 Flowchart provided for this application;

[0049] Figure 3 A schematic diagram of the model training guidance provided for this application;

[0050] Figure 4 A schematic diagram of another model training guide provided for this application;

[0051] Figure 5 This is the system structure diagram provided for this application. DETAILED DESCRIPTION

[0052] As required, detailed embodiments of the present invention are disclosed herein, but it should be understood that the disclosed embodiments are merely exemplary of the present invention, and the present invention may be implemented in different and alternative forms. The accompanying drawings may not be drawn to scale, and certain features may be exaggerated or reduced to show the details of specific components. Therefore, the specific structural and functional details disclosed herein should not be understood as having a limiting meaning, but only as a representative basis to teach those skilled in the art to adopt the present invention differently. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present invention.

[0053] The following is an explanation of the prior art used in this application:

[0054] Attention mechanism: The task of the attention mechanism is to obtain local attention information. The introduction of the attention mechanism allows the task to know which parts of the input data are more worthy of attention. It is usually expressed as Q, K, V, namely Query, Key and Value. All three are derived from the input features themselves and are vectors generated based on the input features. Q can be regarded as a vector of a single input feature. When a set of Q is directly input into the network for training, the network is a network without the introduction of the attention mechanism. However, if the attention mechanism is introduced, the set of Q needs to be multiplied by a set of weights W (K, V) respectively, so that the local input features can be paid attention to. W (K, V) is used to calculate the similarity between K and V. The common one is the dot-product attention mechanism that uses dot product calculation, that is, Q is the vector representing the input features, K and V are the feature vectors for calculating the weights of the attention mechanism, and all three are obtained from the input features. It means that the similarity of the current Key and Value is calculated, and this similarity value is calculated through the Softmax layer to obtain a set of weights. The Value value under the attention mechanism is obtained by summing the product of this set of weights and the corresponding Value.

[0055] Figure 2 A flowchart of a live broadcast platform information integration method provided in an embodiment of the present application:

[0056] The first step is to freeze the first messenger model. At this time, the first messenger model contains the global information obtained in the previous round. The information is injected into the first local model, that is, the first messenger model is used to guide the training of the first local model. In this way, the information transmission is completed and the second local model is obtained. Freezing refers to deleting the feature extractor and classifier of the first messenger model and the receiver of the client so that it does not affect the training of the second local model. Instead, the training of the second local model is guided by the loss generated in the process of generating the classifier when the first local model and the first messenger model perform feature similarity calculations with the receiver of the client. The process of information injection is to extract features from the first messenger model through the first feature extractor, and store the target features required for this round of training in the receiver of the client, while other features are stored in the first classifier of the downstream task. This process is as follows: Figure 3 As shown, all feature dimensions are unified, and then the features in the first local model are defined as V and K. The features are also extracted by the second feature extractor, and the feature dimensions are unified. Other features are stored in the classifier of the downstream task. The result generated by the classifier here is the classification result of the first local model and the first messenger model when calculating their own similarity, which plays a certain role in fine-tuning the training of the second local model. Among them, the output logic of establishing the first messenger model to guide the first local model is:

[0057]

[0058] Where E represents the expectation, (X, Y) is the joint distribution of samples and features, and are the predictions of the first local model and the first messenger model for sample x, is the cross entropy function, is the divergence function, λ is a hot label y and soft label Hyperparameters that balance between .

[0059] Train a small first local model to learn knowledge from a large pre-trained first messenger model. Probability and features are the most commonly used types of knowledge in KD, describing the predictive distribution of the model.

[0060] The second step is to freeze the second local model. At this time, the second local model obtains the local information and the global information obtained in the previous step, and distills the information to the first messenger model, that is, the second local model is used to guide the training of the first messenger model to obtain the second messenger model. At this time, the second messenger model obtains the local information. Similarly, freezing refers to deleting the feature extractor and classifier of the second local model and the client's receiver so that they do not affect the training of the second messenger model. Instead, the training of the second messenger model is guided by the loss generated in the process of generating the classifier when the first local model and the first messenger model perform feature similarity calculations with the client's receiver. This process is as follows: Figure 4 As shown, all feature dimensions are unified. At this time, the features in the first messenger model are defined as V and K, and the features in the second local model are defined as Q, that is, the second local model is used to fine-tune the second messenger model. The output logic of establishing the second local model to guide the first messenger model is:

[0061]

[0062] Among them, E represents the expectation, (X, Y) is the joint distribution of samples and features, and are the features of the second local model and the first messenger model at the i-th layer, respectively. ψ is a feature projector that maps the features of the first messenger model to match the features of the second local model. ||XY|| represents the square root of the sum of the squares of the components after the vector difference. i is an integer greater than 0.

[0063] At this time, the second messenger model acts as a small model to learn knowledge from the relatively larger second local model. That is, the prompt-based distillation uses a finer-grained second local model activation in the intermediate layer to train the second messenger model.

[0064] The third step is to upload the features and weights in the second messenger model to the server through the feature extractor and classifier. At this time, the messenger model, as a lightweight model, completes the feature training of the local model and also obtains the features in the local model, which are not available on the server.

[0065] The fourth step is to establish the parameter weighted average function:

[0066]

[0067] in , , They represent the initial feature vector, the feature vector and weight output by the feature extractor or classifier of the i-th client respectively, ||XY|| represents the square root of the sum of the squares of the components after the vector is subtracted, and i is an integer greater than 0.

[0068] This step requires weighted averaging of the features and weights obtained by the messenger model from each client to unify the information obtained and store the features of this information into a new messenger model.

[0069] The fifth step is to bring this new messenger model into the next round and once again act as the first messenger model to guide the training of the local model.

[0070] like Figure 5 As shown, a live broadcast platform information integration device includes:

[0071] A memory 30, for storing computer programs;

[0072] The processor 31 is used to implement the steps of the live broadcast platform information integration method mentioned in the above embodiment when executing the computer program.

[0073] The live broadcast platform information integration device provided in this embodiment may include but is not limited to a smart phone, a tablet computer, a laptop computer or a desktop computer.

[0074] Among them, the processor 31 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 31 can be implemented in at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), and programmable logic array (PLA). The processor 31 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a central processing unit (CPU); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 31 may be integrated with a graphics processing unit (GPU), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 31 may also include an artificial intelligence (AI) processor, which is used to process computing operations related to machine learning.

[0075] The memory 30 may include one or more computer-readable storage media, which may be non-transitory. The memory 30 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In this embodiment, the memory 30 is at least used to store the following computer program, wherein, after the computer program is loaded and executed by the processor 31, the relevant steps of the live broadcast platform information integration method disclosed in any of the aforementioned embodiments can be implemented. In addition, the resources stored in the memory 30 may also include an operating system and data, etc., and the storage method may be short-term storage or permanent storage. Among them, the operating system may include Windows, Unix, Linux, etc. The data may include but is not limited to the live broadcast platform information integration method, etc.

[0076] In some embodiments, the live broadcast platform information integration device may also include a display screen, an input and output interface, a communication interface, a power supply, and a communication bus.

[0077] Those skilled in the art will understand that Figure 3 The structure shown in does not constitute a limitation on the live broadcast platform information integration device, and may include more or fewer components than shown in the figure.

[0078] The live broadcast platform information integration device provided in the embodiment of the present application includes a memory 30 and a processor 31. When the processor 31 executes the program stored in the memory 30, the live broadcast platform information integration method can be implemented.

[0079] Finally, the present application also provides an embodiment corresponding to a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps recorded in the above method embodiment are implemented.

[0080] It is understandable that if the method in the above embodiment is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory), ROM, random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store program code.

[0081] The above is a detailed introduction to a live broadcast platform information integration method, device, equipment and medium provided by the present application. The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the embodiments can refer to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can refer to the method part description. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

[0082] It should also be noted that, in this specification, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

Claims

1. A method for integrating live broadcast platform information, the method is initiated by a client, the client transmits information with a server through a receiver and a transmitter, and is characterized in that: Freeze the first messenger model of the client, inject the first messenger model into the first local model of the client through the receiver, and obtain the second local model located at the client; including obtaining a first feature extractor and a second feature extractor; extracting the input feature vector in the first messenger model according to the first feature extractor to obtain a first classifier for the downstream task; guiding the second feature extractor to extract the feature vector of the weight in the first local model through the feature vector input in the first messenger model to obtain a second classifier for the downstream task; the freezing refers to deleting the feature extractor and classifier of the first messenger model and the receiver of the client, so as not to affect the training of the second local model; wherein, the output logic of establishing the first messenger model to guide the first local model is: , Where E represents the expectation, (X, Y) is the joint distribution of samples and features, and are the predictions of the first local model and the first messenger model for sample x, is the cross entropy function, is the divergence function, λ is a hot label y and soft label Hyperparameters that trade off between Freeze the second local model, distill the information extracted by the second local model to the first messenger model through the transmitter, and obtain the second messenger model located at the client; including obtaining a third feature extractor and a fourth feature extractor; extracting the input feature vector in the second local model according to the third feature extractor to obtain a third classifier for the downstream task; guide the fourth feature extractor to extract the feature vector of the weight in the first messenger model through the feature vector input in the second local model, and obtain a fourth classifier for the downstream task; wherein, the output logic of establishing the second local model to guide the first messenger model is: , Among them, E represents the expectation, (X, Y) is the joint distribution of samples and features, corresponding to and , and are the features of the second local model and the first messenger model at the i-th layer, respectively. ψ is a feature projector that maps the features of the first messenger model to match the features of the second local model. ||XY|| represents the square root of the sum of the squares of the components after the vector is subtracted. i is an integer greater than 0. Sending the second messenger model to the server; Performing weighted average calculation of parameters of the messenger model and the local model on the server side; The result of the weighted average calculation of the parameters is fed back to the client, and the above steps are repeated until the model converges to integrate the information, which includes image classification, natural language processing and speech recognition.

2. The live broadcast platform information integration method according to claim 1, characterized in that: The parameter weighted average function is established as: , in , , They represent the initial feature vector, the feature vector and weight output by the feature extractor or classifier of the i-th client respectively, ||XY|| represents the square root of the sum of the squares of the components after the vector difference is taken, i is an integer greater than 0, and N represents N clients.

3. The live broadcast platform information integration method according to claim 1, characterized in that: Sending the second messenger model to the server side includes uploading the feature extractor and classifier in the second messenger model to the server side, wherein the feature extractor and classifier in the second messenger model contain feature vectors and weights required for parameter weight averaging.

4. A live broadcast platform information integration system, used to implement the live broadcast platform information integration method according to any one of claims 1 to 3, characterized in that: include Injection module: freeze the first messenger model of the client, inject the first messenger model into the first local model of the client through the receiver, and obtain the second local model located at the client; including obtaining a first feature extractor and a second feature extractor; extracting the input feature vector in the first messenger model according to the first feature extractor to obtain the first classifier of the downstream task; guiding the second feature extractor to extract the feature vector of the weight in the first local model through the feature vector input in the first messenger model to obtain the second classifier of the downstream task; the freezing means deleting the feature extractor and classifier of the first messenger model and the receiver of the client, so as not to affect the training of the second local model; Distillation module: freeze the second local model, distill the information extracted by the second local model to the first messenger model through the transmitter, and obtain the second messenger model located at the client; including obtaining a third feature extractor and a fourth feature extractor; extracting the input feature vector in the second local model according to the third feature extractor to obtain a third classifier for the downstream task; guiding the fourth feature extractor to extract the feature vector of the weight in the first messenger model through the feature vector input in the second local model, and obtaining a fourth classifier for the downstream task; Upload module: send the second messenger model to the server; Calculation module: performing parameter weighted average calculation on the messenger model and the local model on the server side; Download module: Feedback the result of the weighted average calculation of the parameters to the client, repeat the functions of the above modules until the model converges to integrate the information, and the information integration includes image classification, natural language processing and speech recognition.

5. A live broadcast platform information integration device, characterized in that: include: Memory for storing computer programs; A processor, used to implement the steps of the live broadcast platform information integration method as described in any one of claims 1 to 3 when executing the computer program.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the live broadcast platform information integration method as described in any one of claims 1 to 3 are implemented.