A Multi-Classifier Ensemble Training Method Based on Semi-Supervised Federated Learning

Abstract

This invention provides a method for ensemble training of multi-classifiers based on semi-supervised federated learning, mainly comprising: clients designing classification models and sending the resulting personalized local models to a trusted central server; the central server collecting models from multiple clients to form a "model pool" for supervised training; multiple clients using the "model pool" from the central server to classify their local data; multiple clients using local labeled data to perform staggered training of their models and uploading them to the trusted central server; the central server performing federated aggregation of the classifier models in the "model pool," fine-tuning the aggregated classifiers, and then sending them to the multiple clients. This invention, based on a federated learning framework, uses a small amount of labeled data and a large amount of unlabeled private data for training, rationally utilizes public data resources, integrates personalized user needs, and effectively protects the privacy information of multiple participating users.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more particularly to a method for ensemble training of multi-classifiers based on semi-supervised federated learning. Background Technology

[0002] With the widespread use of medical image acquisition devices such as X-ray imaging equipment, CT scanners, and ultrasound imaging devices, massive amounts of medical image data are generated and stored in medical institutions. Since medical data contains a large amount of disease characteristic data, in-depth analysis of the characteristics of medical data is of great significance for disease diagnosis.

[0003] The development of artificial intelligence (AI) technology has provided broader opportunities for the application of medical images. Specifically, AI technologies related to medical image classification can be broadly categorized into the following three types:

[0004] I. Data Automatic Classification Technology Based on Artificial Intelligence Methods. Most existing artificial intelligence methods aim to use publicly available labeled datasets for training and testing with machine learning models. For example, Supratak et al., in "DeepSleepNet: A Model for Automatic Sleep Stage Scoring Based on Raw Single-Channel EEG," used two convolutional kernels of different sizes to extract temporal and frequency features from sleep data and used a bidirectional Long Short-Term Memory (LSTM) module to learn time transition patterns. Convolutional Neural Networks (CNNs) and Deep Belief Networks (DBNs) have excellent performance in feature extraction, while Recurrent Neural Networks (RNNs) have good capabilities in time series signal processing. Currently, data automatic classification technology based on artificial intelligence methods has been widely promoted and applied in various fields such as politics, economics, and culture.

[0005] Second, semi-supervised learning techniques based on "divergence" utilize "multiple views," i.e., multiple "attribute sets" of data objects, to address the problem of using a large number of unlabeled samples to improve learning algorithm performance when only a small number of labeled samples are available. Specifically, for web page classification tasks, the description of each example can be divided into two different views. For example, the description of a web page can be divided into words on the web page and words in hyperlinks pointing to the web page. When there is insufficient labeled data, two views can be used simultaneously to supplement a smaller set of labeled examples with inexpensive unlabeled data. Blum et al. first proposed the Co-training algorithm in "Combining labeled and unlabeled data with co-training," which trains the model using two "views" of a large amount of unlabeled data when there is a small amount of labeled data, obtaining two models and completing the data labeling task. Currently, semi-supervised learning techniques have been widely adopted, especially in web page classification tasks where they have shown significant application value.

[0006] Third, federated learning-based multi-center big data modeling technology is a distributed learning technique that emerged to address the current situation where stricter personal data regulations restrict data collection, making it impossible to obtain better-performing models through centralized learning. For example, hospital patient information often includes Personal Health Information (PHI), which contains a large amount of sensitive personal information such as names, addresses, and phone numbers. Collecting and using this information for the learning process violates privacy laws worldwide, and sharing private data often carries the risk of data leakage. In deep learning, it is generally believed that more training data leads to better performance; however, due to patient privacy concerns, a large amount of data remains idle and cannot be fully utilized. Federated learning technology solves the data privacy problem.

[0007] However, for personalized tasks, such as personalized annotation of privacy-preserving medical image data, the data accumulated by multiple participating clients is unlabeled, so existing federated learning frameworks cannot be used directly. Summary of the Invention

[0008] In view of the shortcomings of existing technologies, this invention proposes a multi-classifier ensemble training method based on semi-supervised federated learning. Based on the federated learning framework, it uses a small amount of labeled data and a large amount of unlabeled private data for training, makes reasonable use of public data resources, integrates users' personalized needs, effectively protects the privacy information of multiple participating users, improves business speed while ensuring accurate classification, and increases the utilization rate of public resources.

[0009] The technical means employed in this invention are as follows:

[0010] A multi-classifier ensemble training method based on semi-supervised federated learning is applied to a distributed classification system. The system includes a central server and multiple participating clients. The central server has first training data, and the multiple clients each have second training data. The first training data includes labeled data, and the second training data includes labeled data and unlabeled data.

[0011] The model training method includes the following steps:

[0012] S1. The central server sets and sends fixed configuration parameters of the classifier to multiple participating clients. The fixed configuration parameters of the classifier include the input format and the output format.

[0013] S2. The client designs a classification model based on the characteristics of local private data, local data classification preferences, and central server configuration parameter requirements, and sends the resulting personalized local model to the trusted central server.

[0014] S3. The central server collects models from multiple clients to form a "model pool". It uses the publicly available first training data to perform supervised training on the classifiers in the "model pool" to obtain an initialized global "model pool" and sends it to the multiple clients.

[0015] S4. Multiple clients use the "model pool" sent by the central server to classify local data, use an integration function on multiple classification outputs to obtain classification results; expand the labeled dataset based on confidence and update the unlabeled dataset;

[0016] S5. Multiple clients use local labeled data to perform staggered training on the model, obtain a local training "model pool", and upload it to the trusted center server.

[0017] S6. The Trusted Center server performs federated aggregation of classifier models in the "model pool", and uses a batch of server data to fine-tune the aggregated classifiers before sending them to multiple clients.

[0018] S7. Repeat steps S4-S6 until the global model converges and the client data update task is completed, then stop communication.

[0019] Furthermore, the integration function is:

[0020]

[0021] Where, λ jThe model decision weights designed for the client are used to adjust the influence of the classifiers in the "model pool" on the classification decision of local data. f1(x) and f2(x) are sub-functions of the ensemble method, f1(x) = Softmax(x) and f2(x) = Sigmoid(x).

[0022] Furthermore, based on the characteristics of local private data, local data classification preferences, and the configuration parameter requirements of the central server, the client independently designs a classification model and sends the resulting personalized local model to the trusted central server, including:

[0023] The client examines local data and categorizes it into labeled and unlabeled data.

[0024] Determine local classification preferences;

[0025] Design one or more classifier model frameworks and internal parameters according to local needs, and send the local model to the trusted central server for centralized initialization.

[0026] Furthermore, the local classification preference is: obtaining a target category labeling accuracy rate higher than a threshold, wherein the target category and the labeling accuracy threshold are preset.

[0027] Furthermore, multiple clients utilize the "model pool" sent by the central server to classify their local data, apply an ensemble function to the multiple classification outputs, and obtain the classification results; the labeled dataset is expanded based on confidence scores, and the unlabeled dataset is updated, including:

[0028] The client receives a global "model pool" from the trusted server. The client then uses the classifiers in the "model pool" to iterate through the local data once and obtain the classification result of each classifier.

[0029] The client uses the classifier and the obtained classification results for interleaved training, that is, it uses data with classification results from other classes to train the classifier, obtains an updated "model pool", and sends it to the trusted center server.

[0030] The client uses the updated "model pool" in conjunction with the ensemble function to obtain classification results, selects classification labels with high confidence, and updates local labeled and unlabeled data.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] 1. This invention proposes a novel semi-supervised learning and federated learning classification model training method based on "divergence", which uses only a portion of labeled data to achieve annotation and personalized model training of a large amount of unlabeled data.

[0033] 2. This patent addresses the common problems of insufficient datasets and low accuracy in supervised deep learning models based on federated learning. Currently, most horizontal federated learning frameworks are supervised learning, but obtaining labeled data in real-world applications is expensive and inconsistent. Existing federated learning uses a single model setting standard, which cannot fully meet the personalized needs of multiple participating clients. To date, semi-supervised learning based on "divergence" often uses a fixed number of participating classifiers, failing to dynamically adapt to the personalized learning requirements of the data. This patent utilizes a "model pool" of multiple classifiers based on "divergence" to implement a semi-supervised federated learning framework, allowing clients to label data while keeping the data separate, meeting the personalized learning needs of clients, protecting data privacy while ensuring good classification accuracy, and achieving high consistency in data labeling. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of a multi-classifier ensemble training method based on semi-supervised federated learning according to the present invention.

[0036] Figure 2 This is a flowchart of a semi-supervised learning process in an embodiment of the present invention.

[0037] Figure 3 This is a schematic diagram illustrating the application of the federated learning framework in the medical field in an embodiment of the present invention. Detailed Implementation

[0038] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0039] This patent invention discloses a method for automatic classification of private, unlabeled data based on semi-supervised learning and federated learning. The algorithm combines a "dynamic" classifier combination, receiving a variable number of personalized models to form a "model pool," and uses interactive training and ensemble classification to update data and models. Publicly available data can be obtained through reasonable use, and a large amount of unlabeled data can be used for model training and data labeling. This algorithm uses federated learning for modeling, but is not limited to the previous single-model standard federated learning. It combines classifier ensemble methods to aggregate multiple models from clients, leveraging the characteristic of "models moving, data remaining stationary" to establish a shared "model pool" while ensuring data privacy. It learns the features of all participating client models, fine-tunes personalized private models, and generates a more robust classifier "model pool."

[0040] The method disclosed in this invention includes the following execution flow: First, a trusted central server sets fixed configuration parameters for the classifier, such as the input and output formats of the classifier; then, it sends these fixed configuration parameters to all participating clients. Next, each participating client constructs a personalized local classifier based on the characteristics of its local private data, local data management needs and preferences, and the central server's configuration parameter requirements; then, each client sends its self-designed classifier model to the trusted central server, forming a "model pool". The trusted central server loads the public dataset and performs supervised training on the classifier models in the "model pool" to obtain an initialized global "model pool," after which it sends all classifiers to the clients. Upon receiving the "model pool," each client performs semi-supervised training on its local unlabeled data to obtain an updated "model pool" and a smaller labeled dataset, then uploads its local "model pool" to the central server. The central server aggregates the classifiers in the "model pool," makes minor adjustments using labeled data, and resends them to the clients for training. This initialization process is repeated until the model converges and the client data update task is completed.

[0041] Detailed steps are explained below;

[0042] S1: The central server sets and sends fixed configuration parameters of the classifier to multiple participating clients: input format and output format.

[0043] S2: The client designs a classification model based on the characteristics of its local private data, local data classification preferences, and the configuration parameters of the central server, and sends the personalized local model to the trusted central server.

[0044] S3: The central server collects models from multiple client sources to form a "model pool." It then uses the loaded public dataset to perform supervised training on the classifiers in the "model pool" to obtain an initialized global "model pool." Right now And send to multiple clients.

[0045] S4: Multiple clients utilize the "model pool" sent by the central server. Right now Classify local data and use an ensemble function for multiple classification outputs, for example... Obtain the category.

[0046]

[0047] Where, λ j The model decision weights are designed for the client, allowing users to adjust the influence of the classifiers in the "model pool" on the classification decisions of local data according to actual needs. f1(x) and f2(x) are sub-functions of the ensemble method; for example, f1(x) = Softmax(x) and f2(x) = Sigmoid(x).

[0048] Expanding the labeled dataset based on confidence levels Right now And update the unlabeled dataset. Right now

[0049] S5: Multiple clients utilize local tag data Interleaved training of the model, i.e., using For the model Training is performed, where a≠b, to obtain a local training "model pool" and upload it to the Trusted Center server.

[0050] S6: The Trusted Center server performs federated aggregation of classifier models in the "model pool", fine-tunes the aggregated classifiers using a batch of server data, and then sends them to multiple clients.

[0051] S7: Repeat process S4 to S6 for t rounds until the global model converges and the client data update task is completed, then stop communication.

[0052] In this process, Algorithm 1 is used for local training of the model on the client and server sides. In round t, firstly, all clients load the model. Perform local model initialization, and then perform personalized model training based on local data.

[0053]

[0054]

[0055] In the Trusted Center server, Algorithm 2 is used for model initialization, global aggregation, and fine-tuning.

[0056]

[0057]

[0058] The initialization and fine-tuning process of the central server's global classifier "model pool" is combined with the integration function. Complete interactive training to obtain an updated "model pool".

[0059] User Guide:

[0060] Step 1: The client requests and obtains the model's input and output settings from the Trusted Center server;

[0061] Step 2: The client reviews the local data, dividing it into labeled and unlabeled data; determines the local classification preferences, such as needing to obtain a higher labeling accuracy for the first stage of non-rapid eye movement; designs one or more classifier model frameworks and internal parameters based on local requirements, and sends the local model to the Trust Center server for centralized initialization;

[0062] Step 3: The client receives the global "model pool" from the trusted server. The client uses the classifiers in the "model pool" to traverse the local data once and obtain the classification result of each classifier.

[0063] Step 4: The client uses the classifier and the obtained classification results for interleaved training, that is, it uses data with classification results from other classes to train the classifier, obtains an updated "model pool", and sends it to the trusted center server.

[0064] Step 5: The client uses the updated "model pool" in conjunction with the ensemble function to obtain the classification results, selects the classification labels with high confidence, and updates the local labeled and unlabeled data;

[0065] Step 6: Repeat steps 3 to 5 until the local data classification is completed and a convergent model is obtained.

[0066] The following specific application examples will further illustrate the solution and effects of the present invention.

[0067] Figure 1This is the algorithm execution flowchart in this embodiment. First, fixed parameters are set on the trusted central server, and clients set personalized local models. Second, the central server collects client models to form a "model pool" and initializes the "model pool" using public data. Then, it is sent to all participating clients for local personalized unlabeled learning. During each round of learning, all models need to perform a calculation on all local data, selecting data with high confidence and adding it to the labeled dataset. The number of labeled data in the dataset continuously increases, while the number of unlabeled data gradually decreases. Afterward, the client sends the classifier "model pool" to the trusted server. The server aggregates each classifier in the "model pool" separately and performs supervised fine-tuning of the whole dataset before sending it back to the client. After multiple aggregations and updates, a converged classifier is output, and the data of each client has been updated to labeled data. The entire process is completed in a way that allows private data to be used but not seen. The trained classification model can be used for data classification in various domains and can be integrated with an organization's AI evaluation system.

[0068] Figure 2 This is a diagram of the algorithm modules for semi-supervised learning performed by the local client in this embodiment.

[0069] Figure 3 This diagram illustrates the application of the algorithm framework in the medical field. First, the trusted central server sets fixed input and output parameters. Clients design and upload classifier models according to these parameters and their individual needs. Then, the trusted central server loads a public dataset, such as the Sleep-EDF dataset from PhysioBank, converting the monitoring signals into medical image data in spectrogram form. Supervised training is then performed on all classifier models to form an initial "model pool." Next, the data is sent to the clients, each performing semi-supervised training and updating both labeled and unlabeled datasets and their classification models in their private data. Finally, multiple participating clients upload their locally updated classification models. The central server performs federated aggregation and fine-tuning of the models before sending them back to the clients. These initialization steps are repeated until all clients' private data has been labeled.

[0070] In summary, this invention designs a novel classification and labeling algorithm based on semi-supervised learning and federated learning, which can meet the personalized local needs of participating institutions, effectively complete the labeling of unlabeled data and the updating of classification models, and is of great significance, especially to the field of medical image processing.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-classifier ensemble training method based on semi-supervised federated learning, characterized in that, The application is applied to a distributed classification system, the system comprising a central server end and a plurality of participating clients, the central server end having first training data, and the plurality of clients each having second training data, the first training data comprising publicly collectible medical image data with labels, and the second training data comprising private medical image data with labels and medical image data without labels; The model training method comprises the following steps: S1, the central server end sets and sends classifier fixed configuration parameters to the plurality of participating clients, the classifier fixed configuration parameters comprising input formats and output formats; S2, the clients independently design classification models according to local private data characteristics, local data classification preferences and central server configuration parameter requirements, and send the obtained personalized local models to the trusted central server; S3, the central server end collects the models of the plurality of clients to form a "model pool", uses the public first training data to perform supervised training on the classifiers in the "model pool", obtains an initialized global "model pool", and sends it to the plurality of clients; S4, the plurality of clients use the "model pool" sent by the central server to classify local data, use an integration function on multiple classification outputs to obtain classification results, and expand a labeled data set based on confidence and update an unlabeled data set; S5, the plurality of clients use local label data to perform interleaved training on the models, obtain a local training "model pool", and upload it to the trusted central server; S6, the trusted central server performs federated aggregation on the classifier models in the "model pool", and performs fine adjustment on the aggregated classifiers using a batch of server data, and then sends them to the plurality of clients; S7, repeatedly perform S4-S6 until the global model converges and the client data completes the update task, and stop communication.

2. The multi-classifier ensemble training method based on semi-supervised federated learning according to claim 1, characterized in that, The integration function is: where λ j is the model decision weight designed for the client to adjust the influence of the classifier in the "model pool" on the classification decision of the local data, f1(x) and f2(x) are the sub-functions of the ensemble method, f1(x) = Softmax(x), and f2(x) = Sigmoid(x).

3. The method of claim 1, wherein the method further comprises: The clients independently design classification models according to local private data characteristics, local data classification preferences and central server configuration parameter requirements, and send the obtained personalized local models to the trusted central server, comprising: The clients review local data and divide them into labeled data and unlabeled data; Determine local classification preferences; Design one or more classifier model frameworks and model internal parameters according to local requirements, and send the local models to the trusted central server for centralized initialization, the classifier model framework being ResNet, SENet or DenseNet.

4. The multi-classifier ensemble training method based on semi-supervised federated learning according to claim 3, characterized in that, The local classification preference is that the annotation accuracy of a target class is higher than a threshold value, wherein the target class and the annotation accuracy threshold value are pre-set.

5. The multi-classifier ensemble training method based on semi-supervised federated learning according to claim 1, characterized in that, The plurality of clients use the "model pool" sent by the central server to classify local data, use an integration function on multiple classification outputs to obtain classification results; Expand a labeled data set based on confidence and update an unlabeled data set, comprising: Upon receiving the global "model pool" sent by the trusted server, the clients use the classifiers in the "model pool" to perform a single pass on local data to obtain classification results of each classifier; The client uses the classifier and the obtained classification result to perform interleaved training, that is, the classifier is trained using data with classification results of non-local classifiers, an updated "model pool" is obtained and sent to the trusted center server; The client uses the updated "model pool" in combination with an integration function to obtain a classification result, selects a classification label with high confidence, and updates the local labeled data and unlabeled data.

Images