A method and system for desensitizing oral medical privacy data based on federated learning

Abstract

The present invention relates to a method and system for desensitizing oral healthcare privacy data based on federated learning. The method comprises the following steps: acquiring multi-source oral healthcare-related data; classifying, cleaning, and desensitizing different types of medical data; constructing a federated learning system based on the processed, high-quality data to enable collaborative modeling across oral healthcare institutions; and improving the security and model performance of the federated learning system through intelligent dynamic adjustment of privacy parameters and attack defense. This invention enables the coordinated optimization of privacy protection and diagnostic model performance.

Description

Technical Field

[0001] The present invention relates to the field of data management, and in particular to a method and system for desensitizing oral medical privacy data based on federated learning. Background Art

[0002] With the rapid development of modern medical technology, particularly the widespread application of artificial intelligence (AI) in healthcare, the dental healthcare industry is gradually moving towards digitalization and intelligentization. However, as highly sensitive data with high privacy and specificity, the processing and sharing of dental medical data present numerous technical and legal challenges. Firstly, this data, which includes diverse forms such as dental images (such as dental X-rays and 3D models), clinical records, and treatment plans, is highly correlated with patient identities and can easily expose privacy. Secondly, when sharing data between medical institutions, or between research institutes and medical institutions for model training or research, strict precautions must be taken to prevent data leakage. Therefore, how to fully realize the value of this data while protecting data privacy has become a pressing issue. Summary of the Invention

[0003] In order to solve the above problems, the purpose of the present invention is to provide a method and system for desensitizing oral medical privacy data based on federated learning, which can achieve the coordinated optimization of privacy protection and diagnostic model performance, and provide a safe, efficient and sustainable solution for cross-regional and multi-institutional medical collaboration.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for desensitizing oral medical privacy data based on federated learning, comprising the following steps:

[0006] Acquire multi-source data related to oral healthcare, and classify, clean, and desensitize different types of medical data;

[0007] Based on the processed high-quality data, a federated learning system is built to achieve collaborative modeling across dental medical institutions;

[0008] Improve the security and model performance of federated learning systems by intelligently and dynamically adjusting privacy parameters and attack defenses, as follows:

[0009] A reinforcement learning algorithm is introduced to dynamically adjust differential privacy parameters, controlling the noise amplitude based on the current training round and model accuracy to minimize the impact on model performance.

[0010] An anomaly detection mechanism is introduced in the transmission gradient to monitor whether the uploaded updates are abnormal and defend against data poisoning or gradient backpropagation attacks.

[0011] Furthermore, we acquire multi-source data related to oral healthcare and classify, clean, and desensitize different types of medical data, as follows:

[0012] The medical data includes oral imaging data, structured data and unstructured text data;

[0013] Check and standardize data formats, remove redundant fields and useless data;

[0014] Oral image data uses a deep learning segmentation algorithm to separate the dental area and facial information, removing unnecessary privacy features;

[0015] Structured data processing involves hashing or replacing sensitive fields;

[0016] Unstructured text data uses NLP models to automatically detect and mask sensitive information such as patient names and ID numbers.

[0017] Furthermore, deep learning segmentation algorithms are used to separate dental areas and facial information from oral imaging data, removing unnecessary privacy features, as follows:

[0018] Acquire oral imaging data, including dental X-rays and CT slices, and standardize the oral imaging data to a fixed resolution to ensure consistent support for size by the deep learning model:

[0019] ;

[0020] in, The height and width corresponding to the target resolution; is the original image; is the resized image; is a pixel; is the size adjustment function;

[0021] Using professional dental annotation tools, manually annotate the segmentation mask of the tooth area in the image to form the training label M(x,y);

[0022] Build a segmentation model based on the UNet model, consisting of an encoder and a decoder, supporting high-precision pixel-level segmentation;

[0023] Encoder: progressively downsamples the image and extracts multi-scale features using multiple layers of convolution and pooling;

[0024] Decoder: progressively upsamples and uses skip connections to pass encoder features to the decoder, combining high- and low-level feature information;

[0025] According to the training dataset, the input image and the true label mask M(x,y) are used to calculate the loss L through forward propagation, and the model parameters are updated using the Adam optimizer;

[0026] Input a new unlabeled oral image and output the tooth area segmentation result after the segmentation model predicts it. ,

[0027] Based on the tooth region segmentation results , the non-tooth areas are set to black or pseudo-data noise to generate a desensitized image.

[0028] Furthermore, a segmentation model is constructed based on the UNet model, as follows:

[0029] In the encoder part, two convolution operations are performed consecutively:

[0030] ;

[0031] Among them, is the feature map, is the convolution kernel, is the bias, is a nonlinear activation function; For input data;

[0032] Use max pooling to progressively downsample the feature map:

[0033] ;

[0034] Where P(X) represents the output feature map after downsampling; K=k*l is the pooling window size; is the maximum value of all values ​​in the pooling window; X is the input feature map;

[0035] In the decoder part, the spatial resolution of the feature map is increased by upsampling:

[0036] ;

[0037] Among them, U(X) represents the output feature map after upsampling; Indicates upsampling; scale=2 is the upsampling multiple, which means that the resolution of the feature map in each dimension is increased by 2 times;

[0038] Each layer of decoder is concatenated with the corresponding encoder feature map through skip connections:

[0039] ;

[0040] Among them, C is the concatenated feature map; Concat represents the concatenation operation; are the outputs of the decoder and encoder respectively;

[0041] Output predicted segmentation mask , the predicted segmentation mask value is a probability distribution value between [0,1];

[0042] During training, a weighted combination of BCE and Dice is used as the loss function L:

[0043] ;

[0044] ;

[0045] ;

[0046] Among them, λ1,λ2 represent the weights of BCE and Dice Loss; L BCE is BCE loss; L Dice is the Dice loss, and N is the number of samples.

[0047] Furthermore, the federated learning system includes independent nodes and a central server. Specifically:

[0048] Each medical institution is treated as an independent node, using the cleaned data of the medical institution to participate in distributed training. Only model parameters are uploaded without sharing the original data. The node local model processes specific tasks.

[0049] Each node Using its local dataset , according to the current weight W t Perform local gradient updates :

[0050]

[0051] in, η is the learning rate; For nodes Loss function for medical tasks; represents the gradient function;

[0052] Central server: schedules training tasks and collects model updates from all nodes After that, perform global model aggregation to generate the optimized global model weight W t+1 :

[0053] ;

[0054] Where n is the number of nodes;

[0055] The optimized global model weights are distributed back to each node for the next round of iteration.

[0056] Furthermore, a reinforcement learning algorithm is introduced to dynamically adjust differential privacy parameters and control the noise amplitude based on the current training round and model accuracy to minimize the impact on model performance. The details are as follows:

[0057] a. Initialize the federated learning global model W0 and the initial privacy budget , noise intensity σ0; start the reinforcement learning agent and generate privacy parameter adjustment actions according to the initial strategy;

[0058] b. Each node Using the privacy budget at time t and noise intensity σ t , in the local dataset Model training is completed on the GitHub repository; the implementation formula for differential privacy is:

[0059] ;

[0060] in, is the gradient update; Gaussian noise generated by the current noise intensity; For nodes On local data The gradient of the loss function L on the model parameters; I is the identity matrix; represents differential privacy noise;

[0061] c. The central server aggregates the updates uploaded by the nodes to obtain the global model;

[0062] d. According to the current status , using reinforcement learning models to generate new privacy parameter adjustment actions ; Among them, L t is the loss function at the current moment; are the adjustment amounts for privacy budget and noise intensity respectively;

[0063] Execute the action to update the privacy parameters:

[0064] ;

[0065] Use RL algorithms to update reinforcement learning policy network weights;

[0066] e. Privacy parameters Model weight W t+1 Distribute to nodes and enter the next round of training, and loop until the termination condition is met.

[0067] Furthermore, an anomaly detection mechanism is introduced in the transmission gradient to monitor whether the uploaded updates are abnormal and defend against data poisoning or gradient back-propagation attacks. The details are as follows:

[0068] In the first few rounds of federated learning training, the gradient updates uploaded by normal nodes are collected. ;

[0069] Gradient updates uploaded to normal nodes , normalized to meet the requirements of the input neural network:

[0070] ;

[0071] in, represents the L2 norm of the gradient; is the normalized input;

[0072] Using normal gradient as training data set, optimizing reconstruction error, the autoencoder contains encoder and decoder ;The optimization goal is to minimize the reconstruction error, and the loss function is MSE;

[0073] In each round of federated learning training, the uploaded gradient is checked before the gradient is updated. Perform standardized preprocessing;

[0074] Use the trained autoencoder to reconstruct the gradient and calculate the reconstruction error :

[0075] ;

[0076] Compare with the detection threshold δ:

[0077] if >δ, the gradient is judged to be abnormal and may come from a malicious node;

[0078] if ≤δ, the gradient is judged to be normal.

[0079] A system for desensitizing oral medical privacy data based on federated learning includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of the above-mentioned method for desensitizing oral medical privacy data based on federated learning.

[0080] The present invention has the following beneficial effects:

[0081] 1. This invention can achieve the coordinated optimization of privacy protection and diagnostic model performance, providing a safe, efficient, and sustainable solution for cross-regional and multi-institutional medical collaboration;

[0082] 2. This invention uses federated learning to allow oral medical data to be stored locally while participating in collaborative training, eliminating the need for data to be discharged from the hospital. This fundamentally reduces the risk of privacy leaks caused by data flow. By dynamically adjusting privacy parameters through reinforcement learning, the noise scale can be dynamically increased or decreased based on the current training round and model accuracy, achieving a refined balance between privacy protection and model performance.

[0083] 3. This invention uses differential privacy technology, intelligent adjustment, and gradient anomaly detection mechanism for multiple encryption protections to effectively curb malicious attacks and ensure data security. By combining high-quality data with dynamic privacy parameter adjustment, it not only ensures the efficient convergence of federated learning, but also improves the overall diagnostic capabilities of the model, defends against malicious updates, and significantly reduces the impact of unstable factors on global model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0085] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0086] refer to Figure 1 In this embodiment, a method for desensitizing oral medical privacy data based on federated learning is provided, comprising the following steps:

[0087] Acquire multi-source data related to oral healthcare, and classify, clean, and desensitize different types of medical data;

[0088] Based on the processed high-quality data, a federated learning system is built to achieve collaborative modeling across dental medical institutions;

[0089] Improve the security and model performance of federated learning systems by intelligently and dynamically adjusting privacy parameters and attack defenses, as follows:

[0090] A reinforcement learning algorithm is introduced to dynamically adjust differential privacy parameters (such as noise intensity and privacy budget ε). The noise amplitude is controlled according to the current training round and model accuracy to minimize the impact on model performance.

[0091] An anomaly detection mechanism is introduced in the transmission gradient to monitor whether the uploaded updates are abnormal and defend against data poisoning or gradient backpropagation attacks.

[0092] In this embodiment, multi-source data related to oral healthcare is obtained, and different types of medical data are classified, cleaned, and desensitized as follows:

[0093] The medical data includes oral imaging data (such as X-rays, 3D models, dental photos), structured data (patient medical records, dental scoring sheets, etc.) and unstructured text data (doctor's diagnosis records, patient descriptions);

[0094] Check and standardize data formats (such as image resolution and field content consistency), and remove redundant fields and useless data;

[0095] Oral image data uses a deep learning segmentation algorithm to separate the dental area and facial information, removing unnecessary privacy features;

[0096] Structured data processing involves hashing or replacing sensitive fields (such as address and contact information);

[0097] Unstructured text data uses NLP models (such as the BERT-based NER model) to automatically detect and mask sensitive information such as patient names and ID numbers.

[0098] In this embodiment, the oral image data uses a deep learning segmentation algorithm to separate the tooth area and facial information and remove unnecessary privacy features, as follows:

[0099] Acquire oral imaging data, including dental X-rays and CT slices, and standardize the oral imaging data to a fixed resolution to ensure consistent support for size by the deep learning model:

[0100] ;

[0101] in, The height and width corresponding to the target resolution; is the original image; is the resized image; is a pixel; is the size adjustment function;

[0102] Using professional dental annotation tools, manually annotate the segmentation mask of the tooth area in the image to form the training label M(x,y);

[0103] Build a segmentation model based on the UNet model, consisting of an encoder and a decoder, supporting high-precision pixel-level segmentation;

[0104] Encoder: progressively downsamples the image and extracts multi-scale features using multiple layers of convolution and pooling;

[0105] Decoder: progressively upsamples and uses skip connections to pass encoder features to the decoder, combining high- and low-level feature information;

[0106] According to the training dataset, the input image and the true label mask M(x,y) are used to calculate the loss L through forward propagation, and the model parameters are updated using the Adam optimizer;

[0107] Input a new unlabeled oral image and output the tooth area segmentation result after the segmentation model predicts it. ,

[0108] Based on the tooth region segmentation results , the non-tooth areas are set to black or pseudo-data noise to generate a desensitized image.

[0109] In this embodiment, a segmentation model is constructed based on the UNet model, as follows:

[0110] In the encoder part, two convolution operations are performed consecutively:

[0111] ;

[0112] Among them, is the feature map, is the convolution kernel, is the bias, is a nonlinear activation function; For input data;

[0113] Use Max-Pooling to gradually downsample the feature map:

[0114] ;

[0115] Where P(X) represents the output feature map after downsampling; K=k*l is the pooling window size; is the maximum value of all values ​​in the pooling window; X is the input feature map;

[0116] In the decoder part, the spatial resolution of the feature map is increased by upsampling:

[0117] ;

[0118] Among them, U(X) represents the output feature map after upsampling; Indicates upsampling; scale=2 is the upsampling multiple, which means that the resolution of the feature map in each dimension is increased by 2 times;

[0119] Each layer of decoder is concatenated with the corresponding encoder feature map through skip connections:

[0120] ;

[0121] Among them, C is the concatenated feature map; Concat represents the concatenation operation; are the outputs of the decoder and encoder respectively;

[0122] Output predicted segmentation mask , whose value is a probability distribution value between [0,1];

[0123] During training, a weighted combination of BCE and Dice is used as the loss function L:

[0124] ;

[0125] ;

[0126] ;

[0127] Among them, λ1,λ2 represent the weights of BCE and Dice Loss; L BCE is the BCE loss (for pixel-level prediction accuracy); L Dice is the Dice loss (optimizing the segmentation performance of the target area), and N is the number of samples.

[0128] In this embodiment, the federated learning system includes independent nodes and a central server. Specifically:

[0129] Each medical institution is treated as an independent node, using its local cleaned data to participate in distributed training. Only model parameters (such as weights or gradients) are uploaded without sharing the original data. The node's local model processes specific tasks (such as tooth region segmentation, caries detection, case prediction, etc.).

[0130] Each node Using its local dataset , according to the current weight W t Perform local gradient updates :

[0131]

[0132] in, η is the learning rate; For nodes Loss functions for medical tasks such as segmentation or classification; represents the gradient function;

[0133] Central server: schedules training tasks and collects model updates from all nodes After that, perform global model aggregation (such as FedAvg algorithm) to generate the optimized global model weight W t+1 :

[0134] ;

[0135] Where n is the number of nodes;

[0136] The optimized global model weights are distributed back to each node for the next round of iteration.

[0137] In this embodiment, a reinforcement learning algorithm is introduced to dynamically adjust differential privacy parameters and control the noise amplitude based on the current training round and model accuracy to minimize the impact on model performance. The details are as follows:

[0138] a. Initialize the federated learning global model W0 and the initial privacy budget , noise intensity σ0; start the reinforcement learning agent and generate privacy parameter adjustment actions according to the initial strategy;

[0139] b. Each node Using the privacy budget at the current time t and noise intensity σ t , in the local dataset Model training is completed on the GitHub repository; the implementation formula for differential privacy is:

[0140] ;

[0141] in, is the gradient update; Gaussian noise generated by the current noise intensity; For nodes On local data The gradient of the loss function L on the model parameters; I is the identity matrix; represents differential privacy noise;

[0142] c. The central server aggregates the updates uploaded by the nodes to obtain the global model;

[0143] d. According to the current status , using reinforcement learning models to generate new privacy parameter adjustment actions ; Among them, L t is the loss function at the current moment; are the adjustment amounts for privacy budget and noise intensity respectively;

[0144] Execute the action to update the privacy parameters:

[0145] ;

[0146] Use RL algorithms to update reinforcement learning policy network weights;

[0147] e. Privacy parameters Model weight W t+1Distribute to nodes and enter the next round of training, and loop until the termination condition is met (such as reaching the target model accuracy or the maximum number of training rounds).

[0148] In this embodiment, an anomaly detection mechanism is introduced in the transmission gradient to monitor whether the uploaded updates are abnormal and defend against data poisoning or gradient back-propagation attacks. The details are as follows:

[0149] In the first few rounds of federated learning training (or from historical gradient distribution information), collect gradient updates uploaded by normal nodes ;

[0150] Gradient updates uploaded to normal nodes , normalized to meet the requirements of the input neural network:

[0151] ;

[0152] in, represents the L2 norm of the gradient; is the normalized input;

[0153] Using normal gradient as training data set, optimizing reconstruction error, the autoencoder contains encoder and decoder ;The optimization goal is to minimize the reconstruction error, and the loss function is MSE;

[0154] In each round of federated learning training, the uploaded gradient is checked before the gradient is updated. Perform standardized preprocessing;

[0155] Use the trained autoencoder to reconstruct the gradient and calculate the reconstruction error :

[0156] ;

[0157] Compare with the detection threshold δ:

[0158] if >δ, the gradient is judged to be abnormal and may come from a malicious node;

[0159] if ≤δ, the gradient is judged to be normal.

[0160] A system for desensitizing oral medical privacy data based on federated learning, comprising a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of the above-mentioned method for desensitizing oral medical privacy data based on federated learning.

[0161] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0162] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0163] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0164] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0165] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for desensitizing oral medical privacy data based on federated learning, characterized in that: The following steps are involved: Acquire multi-source data related to oral healthcare, and classify, clean, and desensitize different types of medical data; Based on the processed high-quality data, a federated learning system is built to achieve collaborative modeling across dental medical institutions; Improve the security and model performance of federated learning systems by intelligently and dynamically adjusting privacy parameters and attack defenses, as follows: A reinforcement learning algorithm is introduced to dynamically adjust differential privacy parameters, controlling the noise amplitude based on the current training round and model accuracy to minimize the impact on model performance. Introducing an anomaly detection mechanism in the transmission gradient to monitor uploaded updates for anomalies and prevent data poisoning or gradient backpropagation attacks. A reinforcement learning algorithm is introduced to dynamically adjust differential privacy parameters and control the noise amplitude based on the current training round and model accuracy to minimize the impact on model performance. The details are as follows: a. Initialize the federated learning global model W0 and the initial privacy budget , noise intensity σ0; Start the reinforcement learning agent and generate privacy parameter adjustment actions according to the initial policy; b. Each node Using the privacy budget at time t and noise intensity σ t , in the local dataset Model training is completed on the GitHub repository; the implementation formula for differential privacy is: ; in, is the gradient update; Gaussian noise generated by the current noise intensity; For nodes On local data The gradient of the loss function L on the model parameters; I is the identity matrix; represents differential privacy noise; c. The central server aggregates the updates uploaded by the nodes to obtain the global model; d. According to the current status , using reinforcement learning models to generate new privacy parameter adjustment actions ; Among them, L t is the loss function at the current moment; are the privacy budget and the adjustment amount of noise intensity respectively; Execute the action to update the privacy parameters: ; Use RL algorithms to update reinforcement learning policy network weights; e. Privacy parameters Model weight W t+1 Distribute to nodes and enter the next round of training, and loop until the termination condition is met.

2. The method for desensitizing oral medical privacy data based on federated learning according to claim 1 is characterized in that: The method of acquiring multi-source data related to oral healthcare and classifying, cleaning and desensitizing different types of medical data is as follows: The medical data includes oral imaging data, structured data and unstructured text data; Check and standardize data formats, remove redundant fields and useless data; Oral image data uses a deep learning segmentation algorithm to separate the dental area and facial information, removing unnecessary privacy features; Structured data processing involves hashing or replacing sensitive fields; Unstructured text data uses NLP models to automatically detect and mask sensitive information such as patient names and ID numbers.

3. The method for desensitizing oral medical privacy data based on federated learning according to claim 2 is characterized in that: The oral image data uses a deep learning segmentation algorithm to separate the dental area and facial information and remove unnecessary privacy features, as follows: Acquire oral imaging data, including dental X-rays and CT slices, and standardize the oral imaging data to a fixed resolution to ensure consistent support for size by the deep learning model: ； in, The height and width corresponding to the target resolution; is the original image; is the resized image; is a pixel; is the size adjustment function; Using professional dental annotation tools, manually annotate the segmentation mask of the tooth area in the image to form the training label M(x,y); Build a segmentation model based on the UNet model, which consists of an encoder and a decoder, supporting high-precision pixel-level segmentation. Encoder: progressively downsamples the image and extracts multi-scale features using multiple layers of convolution and pooling; Decoder: progressively upsamples and uses skip connections to pass encoder features to the decoder, combining high- and low-level feature information; According to the training dataset, the input image and the true label mask M(x,y) are used to calculate the loss L through forward propagation, and the model parameters are updated using the Adam optimizer; Input a new unlabeled oral image and output the tooth area segmentation result after the segmentation model predicts it. , Based on the tooth region segmentation results , the non-tooth areas are set to black or pseudo-data noise to generate a desensitized image.

4. The method for desensitizing oral medical privacy data based on federated learning according to claim 3 is characterized in that: The segmentation model is constructed based on the UNet model, as follows: In the encoder part, two convolution operations are performed consecutively: ； Among them, is the feature map, is the convolution kernel, is the bias, is a nonlinear activation function; For input data; Use max pooling to progressively downsample the feature map: ； Among them, K=k*l is the pooling window size; is the maximum value of all values ​​in the pooling window; X is the input feature map; In the decoder part, the spatial resolution of the feature map is increased by upsampling: ； Among them, U(X) represents the output feature map after upsampling; Indicates upsampling; scale=2 is the upsampling multiple, which means that the resolution of the feature map in each dimension is increased by 2 times; Each layer of decoder is concatenated with the corresponding encoder feature map through skip connections: ； Among them, C is the concatenated feature map; Concat represents the concatenation operation; are the outputs of the decoder and encoder respectively; Output predicted segmentation mask , the predicted segmentation mask value is a probability distribution value between [0,1]; During training, a weighted combination of BCE and Dice is used as the loss function L: ； ； ； Among them, λ1,λ2 represent the weights of BCE and Dice Loss; L BCE is BCE loss; L Dice is the Dice loss, and N is the number of samples.

5. The method for desensitizing oral medical privacy data based on federated learning according to claim 1 is characterized in that: The federated learning system includes independent nodes and a central server. Specifically: Each medical institution is treated as an independent node, using the cleaned data of the medical institution to participate in distributed training. Only model parameters are uploaded without sharing the original data. The node local model processes specific tasks. Each node Using its local dataset , according to the current weight W t Perform local gradient updates : ； in, η is the learning rate; For nodes Loss function for medical tasks; represents the gradient function; Central server: schedules training tasks and collects model updates from all nodes After that, perform global model aggregation to generate the optimized global model weight W t+1 : ； Where n is the number of nodes; The optimized global model weights are distributed back to each node for the next round of iteration.

6. The method for desensitizing oral medical privacy data based on federated learning according to claim 1, characterized in that: The above introduces an anomaly detection mechanism in the transmission gradient to monitor whether the uploaded updates are abnormal and defend against data poisoning or gradient backpropagation attacks. The details are as follows: In the first few rounds of federated learning training, the gradient updates uploaded by normal nodes are collected. ; Gradient updates uploaded to normal nodes , normalized to meet the requirements of the input neural network: ； in, represents the L2 norm of the gradient; is the normalized input; Using normal gradient as training data set, optimizing reconstruction error, the autoencoder contains encoder and decoder ;The optimization goal is to minimize the reconstruction error, and the loss function is MSE; In each round of federated learning training, the uploaded gradient is checked before the gradient is updated. Perform standardized preprocessing; Use the trained autoencoder to reconstruct the gradient and calculate the reconstruction error : ； and detection threshold Compare: if , it is determined that the gradient is abnormal and may come from a malicious node; if The gradient is judged to be normal.

7. A dental medical privacy data desensitization system based on federated learning, characterized by: The method comprises a processor, a memory and a computer program stored in the memory. When the processor executes the computer program, the method specifically performs the steps of the oral medical privacy data desensitization method based on federated learning as described in any one of claims 1 to 6.

Images