Federal learning aggregation method and system capable of resisting dimension collapse and medical image application

Abstract

The invention discloses a federated learning aggregation method and system capable of resisting dimension collapse and medical image application, and the federated learning aggregation method is non-independent identically distributed data federated learning based on a tracking kernel and robust aggregation, and aims at dimension collapse in federated learning. And a regularization item is introduced to punish covariance between singular values in a local model training process so as to inhibit a dimension collapse phenomenon. Based on the application of the method, the invention provides a federated learning system for medical image classification, and provides a solution combining singular value variance suppression and robust aggregation for the problems of high privacy of medical data, strong distribution isomerism (Non-IID) and model dimension collapse. The client equalizes a medical image feature representation space and relieves dimension collapse through trace-kernel normal form approximate singular value distribution; and the server dynamically allocates the weight of the client model based on the geometric median to resist abnormal data interference, and is suitable for cross-mechanism medical image cooperative analysis.

Description

Technical Field

[0001] The present invention belongs to the technical fields of medical image processing and federated learning, and particularly relates to a federated learning method for cross-institutional medical image collaboration, which improves classification accuracy and data privacy by suppressing model dimensional collapse and robust aggregation. Background Art

[0002] In recent years, although the rapid development of medical image intelligent diagnosis technology has significantly improved the disease recognition efficiency, it faces multiple severe challenges in the federated learning framework. First, there is a fundamental contradiction between the privacy protection and collaboration needs of medical data: restricted by regulations such as HIPAA and GDPR, cross-institutional data sharing has risks of leakage and compliance obstacles, making it difficult to implement the traditional centralized training mode. More complexly, medical data naturally exhibits the characteristic of non-independent and identically distributed (Non-IID), which is manifested as the skewed disease distribution between specialized hospitals and general institutions (such as a 35% difference in the proportion of melanoma), and also results from the feature space offset caused by differences in equipment parameters and patient groups. This data heterogeneity easily leads to the dimensional collapse of the federated model - the feature expression degenerates into a low-dimensional subspace, making it difficult to distinguish subtle pathological features (such as the pigment distribution pattern in dermoscopic images), and the misdiagnosis rate of similar diseases between classes (such as the X-ray signs of viral and bacterial pneumonia) increases significantly. Existing solutions have obvious defects: basic federated algorithms (such as FedAvg) are sensitive to Non-IID data, and the cross-institutional pneumonia classification accuracy drops by more than 15%; personalized methods (such as FedPer) do not decouple general / specific features, and the communication overhead increases by 40%; although the anti-collapse mechanism (FedDecorr) introduces decorrelation regularization, the computational cost increases by 30%, and it lacks effective defense against data poisoning attacks (such as a 20% label flip resulting in a 0.2 decrease in AUC), making it difficult to meet the high-precision and real-time requirements of clinical diagnosis.

[0003] The technological evolution process reveals the core contradiction: from centralized training (restricted by privacy barriers) to the basic federated framework (FedAvg proposed in 2017), then to personalized adaptation (FedBN in 2020) and anti-collapse exploration (FedDecorr in 2022), the "privacy-performance-efficiency" trilemma has never been broken through. Existing medical federated solutions (such as the model decoupling strategy in CN patents) have partially alleviated the feature coupling problem, but have not integrated a dynamic anti-interference mechanism and are still vulnerable under malicious attacks. Therefore, there is an urgent need to build a new federated learning framework that, while ensuring data privacy, enhances the medical feature expression ability by suppressing dimensional collapse, integrates lightweight and robust aggregation to resist abnormal interference, and at the same time adapts to the computing power constraints of edge medical devices. This is the technical breakthrough direction of the FedTK method of the present invention. Summary of the Invention

[0004] Objective of the Invention: The present invention aims to provide a federated learning aggregation method, system, and medical imaging application for anti-dimensional collapse, which uses the anti-dimensional collapse mechanism and robust aggregation technology of federated learning to collaboratively optimize medical image diagnosis models, improving the accuracy of pathological feature expression and the reliability of abnormal attack defense.

[0005] Technical Solution: An anti-dimensional collapse federated learning aggregation method, which is a federated learning of non-independent and identically distributed data based on a tracking kernel and robust aggregation. To address dimensional collapse in federated learning, a regularization term is introduced to penalize the covariance between singular values during the local model training process, thereby suppressing the phenomenon of dimensional collapse. Specifically, it includes:

[0006] (1) Calculate the local loss function f i (w), and update the local model parameters based on dimensional descent;

[0007] (2) Combine the trace and nuclear norm to approximately represent the singular value variance, and introduce a regularization term to constrain the singular value variance of the feature representation. Its expression is:

[0008]

[0009] where w represents the client model parameters, D represents the training dataset, d represents the data dimension, K represents the feature matrix, and ||K|| * represents the nuclear norm, which is the sum of singular values;

[0010] The overall objective of each local model is: min f i (w)+μf singular (w,D), where μ is a hyperparameter;

[0011] (3) The server calculates the weights for the updates of each local model based on the geometric median. The expression is:

[0012]

[0013] where α i is the initial weight, δ is the tolerance parameter, ∨ represents the maximization operation, and ||w i || is the Euclidean norm of the model parameters;

[0014] The server performs an aggregation operation based on the received weights β i and the local model update vector of the client to calculate the global model update, expressed as:

[0015]

[0016] S tDenote as the set of clients selected by the central server to participate in the training in the $t$-th round of federated learning training;

[0017] (4) The server aggregates the local models of the clients to generate a global model, and distributes it to the corresponding local models of each client for iterative training until the model converges.

[0018] Furthermore, the calculation of the feature matrix $K$ is as follows:

[0019] First, calculate the output representations $z$ of the local models of $N$ clients i , and obtain the covariance matrix:

[0020]

[0021] where the $i$-th data point is denoted as $z$ i , is the average of all data point representations;

[0022] Then, perform vector normalization on the outputs of all local models, which is mathematically expressed as:

[0023]

[0024] Through the above formula, the covariance matrix $z$ i is made equal to its feature matrix $K$, resulting in:

[0025]

[0026] Furthermore, during the local training process, the regularization strength is dynamically adjusted to adapt to data scenarios with different degrees of non-independent and identically distributed.

[0027] On the second aspect, the present invention provides a federated learning system for non-independent and identically distributed data, including:

[0028] Client devices, configured to execute the local model training steps of the above-mentioned federated learning aggregation method for anti-dimensional collapse;

[0029] A central server, configured to execute the robust aggregation steps of the above-mentioned federated learning aggregation method for anti-dimensional collapse, including aggregating the local models of the clients to generate a global model, and distributing it to the corresponding local models of each client for iterative training;

[0030] A communication module, used to transmit encrypted model parameter updates between the client and the server.

[0031] On the third aspect, the present invention provides a computer-readable storage medium, storing a computer program, which when executed by a processor, implements the above-mentioned federated learning aggregation method for anti-dimensional collapse.

[0032] Fourthly, as an application, the present invention provides a federated learning system for medical image classification. This system implements the above-mentioned federated learning aggregation method against dimensional collapse for model training of clients and global model update, including:

[0033] Clients deployed in medical institutions, where the clients train local models through local medical image data and integrate the singular value equalization regularization term;

[0034] The central server calculates the weights of model updates of each client based on the geometric median, dynamically excludes the interference of abnormal clients, and updates the global model and distributes local model training parameters;

[0035] Communication encryption module: Ensures the transmission privacy of medical image data and model parameters.

[0036] Furthermore, in a federated learning system for medical image classification, the local loss function is an inter-class separability loss function optimized based on the hypersphere space, and its mathematical representation is:

[0037]

[0038] where is the angle between the feature and the class weight vector; s is the scaling factor, and v is the boundary penalty coefficient. By mapping the feature vector and the classification weight to the unit sphere, the similarity of different disease classes is dynamically calculated.

[0039] In the model aggregation stage, by combining geometric median calculation and proximal gradient optimization, and analyzing the deviation degree of the parameters uploaded by each client, the Euclidean norm deviation degree of the client model parameter w i is:

[0040]

[0041] If score i > 2, it is determined as an abnormal client and its weight is reduced, automatically identifying and reducing the weight impact of abnormal clients, and combining inertia term constraints to ensure the stability of model update.

[0042] Beneficial effects: From an algorithmic perspective, the present invention approximates the singular value variance by combining the trace and the nuclear norm, thereby alleviating the model dimensionality collapse problem caused by non-independent and identically distributed (non-iid) data. At the same time, a robust aggregation strategy based on the geometric median is introduced to reduce the impact of abnormal clients on the global model and enhance the stability and reliability of the model. From an application perspective, the present invention can efficiently address the problems of high medical data privacy, strong distribution heterogeneity (Non-IID), and model dimensionality collapse. Overall, the present invention proposes singular value variance suppression and robust aggregation. At the same time, two key challenges are considered and solved: one is that some client models are overly simplified, resulting in the underutilization of the representation space to distinguish diverse data of different categories, leading to dimensionality collapse; the other is that the server is interfered by abnormal data due to the client model weight problem. Brief Description of the Drawings

[0043] Figure 1 is the system framework diagram of MedFedTK described in the present invention;

[0044] Figure 2 is the federated training flowchart of MedFedTK described in the present invention;

[0045] Figure 3 is the model accuracy obtained when different proportions of damaged clients participate in training on the CIFAR10 dataset with a data heterogeneity of 0.1 in the present invention;

[0046] Figure 4 is the comparison of the singular value distributions of dermoscopic image features before and after balancing;

[0047] Figure 5 is the classification accuracy comparison curve of different methods under poisoning attacks. Detailed Embodiments

[0048] With the increasing prominence of data privacy issues, federated learning has emerged as a distributed machine learning method. It enables clients to use their own local data for training and jointly train the model by uploading model parameters. In this way, while ensuring privacy, the model performance is maintained. However, federated learning faces problems such as model performance degradation caused by non-independent and identically distributed (NON-IID) data and malicious attacks, which affect the performance of the global model. To address the data heterogeneity problem, existing methods face challenges such as high computational costs, limited applicability, or the inability to fully capture data features; in the model aggregation stage, traditional weighted average aggregation methods are sensitive to abnormal clients, and although there are multiple improved algorithms, they still face challenges.

[0049] To address this problem, the present invention provides a federated learning aggregation method against dimensionality collapse, combined with Figure 1and Figure 2 which we call the FedTK method, and this method is also a method for non-independent and identically distributed federated learning through trace kernels and robust aggregation. This federated learning framework includes the following core modules:

[0050] (1) Model decoupling strategy based on singular value decomposition (SVB-DLMS): Decompose the local model into a universal component (UC) and a personalized component (PC), and control the knowledge sharing granularity through singular value thresholds;

[0051] (2) Inter-class separability loss function (IS Loss): Dynamically adjust the decision boundary angle in the feature space to enhance the model's ability to distinguish similar pathological features;

[0052] (3) Geometric median robust aggregation (GeoMed Aggregation): Dynamically allocate weights based on the model parameter norm to suppress the interference of abnormal clients.

[0053] Specifically, the implementation of the FedTK method is as follows:

[0054]

[0055] Example 1: For the FedTK algorithm, we conducted the following experiments for analysis and evaluation.

[0056] Dataset: Three typical and classic datasets were used, namely MNIST, CIFAR10, and CIFAR100. These datasets have different attributes, dimensions, and numbers of classes, so as to comprehensively and efficiently evaluate the performance of the proposed method. To simulate a real federated learning environment, the Dirichlet distribution was used to simulate the non-IID data scenario, and the data was assigned to each client. Specifically, the parameters of the Dirichlet distribution were set to 0.1, 0.5, and 1 to generate different levels of non-IID characteristics, so as to test and verify the performance of the method under different data conditions.

[0057] Model: A simple-structured CNN model was adopted for the classification task. For the MNIST dataset, a CNN model containing 2 convolutional layers and 2 fully connected layers was used. For the more complex CIFAR10 and CIFAR100 datasets, a CNN model containing 2 convolutional layers and 3 fully connected layers was adopted.

[0058] Hyperparameters: All training was completed within 100 rounds. For CIFAR10 and CIFAR100, the learning rate was 0.01; for MNIST, the learning rate was 0.05. The SGD optimizer (momentum was 0.9, batch size was 32) was used. There were 100 clients, and 10 clients were randomly selected for training in each round of iteration.

[0059] Baseline methods: FedAvg, FedProx, Scaffold, Moon, and FedDecorr represent several typical algorithms in the field of federated learning. By comparing with these algorithms, the performance of the method proposed in the present invention can be effectively evaluated.

[0060] Update poisoning and omniscient poisoning settings: The attacker will modify the parameters returned by the selected clients participating in the training, such that the model parameters returned by these clients become the negative values of the aggregated results of the uncompromised model parameters, i.e., set to. To explore the impact of different proportions of compromised clients on the model performance, in this experiment, 10%, 20%, and 40% of the clients are set as omniscient compromised clients for the experiment, so as to further test the stability and anti-interference ability of the model when suffering from high-intensity attacks.

[0061] We conducted experiments on the performance of each algorithm under different datasets and different Dirichlet distribution parameters, and the results are shown in Tables I, II, and III. We also conducted experiments under data poisoning and omniscient poisoning (with different severities), and the data poisoning results on the CIFAR10 dataset are shown in Tables IV and V. The results show that the method can effectively prevent the degradation of the global model performance due to attacks on some clients. On the CIFAR10 dataset, the model accuracy under omniscient poisoning when the data heterogeneity is 0.1 is as Figure 3 shown. Omniscient poisoning has a negative impact on the performance of all methods. The FedTK method proposed by us shows the strongest robustness when facing omniscient poisoning and can maintain a high test accuracy under different levels of poisoning. Other methods show varying degrees of performance degradation under omniscient poisoning, among which FedDecorr shows a more significant performance degradation under a higher level of poisoning.

[0062] Table I. Model accuracies under different heterogeneity degrees on the MNIST dataset

[0063] Method Dir(0.1) Dir(0.5) Dir(1) FedAvg 0.964 0.979 0.980 FedProx 0.963 0.978 0.980 Scaffold 0.967 0.985 0.986 Moon 0.964 0.977 0.980 FedDecorr 0.964 0.978 0.979 FedTK 0.986 0.991 0.991

[0064] Table II. Model accuracies under different heterogeneity degrees on the CIFAR10 dataset

[0065] Method Dir(0.1) Dir(0.5) Dir(1) FedAvg 0.473 0.499 0.515 FedProx 0.296 0.487 0.516 Scaffold 0.307 0.430 0.492 Moon 0.435 0.492 0.510 FedDecorr 0.478 0.479 0.519 FedTK 0.466 0.505 0.535

[0066] Table III. Model accuracies under different heterogeneity degrees on the CIFAR100 dataset

[0067] Method Dir(0.1) Dir(0.5) Dir(1) FedAvg 0.160 0.173 0.174 FedProx 0.121 0.167 0.166 Scaffold 0.0646 0.0784 0.0979 Moon 0.160 0.178 0.166 FedDecorr 0.164 0.173 0.179 FedTK 0.168 0.193 0.174

[0068] Table IV. Model accuracies of the models participating in the training after the client labels are flipped

[0069] Method Dir(0.1) Dir(0.5) Dir(1) FedAvg 0.434 0.455 0.473 FedProx 0.394 0.453 0.451 Scaffold 0.264 0.345 0.364 Moon 0.396 0.462 0.458 FedDecorr 0.428 0.448 0.454 FedTK 0.446 0.487 0.499

[0070] Table V. Model accuracy of the client after pixel flipping for training

[0071]

[0072]

[0073] Example 2: Dermoscopic image classification

[0074] First, the ISIC2017 and ISIC2018 datasets publicly available from the International Skin Imaging Collaboration (ISIC) were adopted, including a total of 10,490 dermoscopic images of three categories: melanoma (MEL), benign keratosis (BKL), and nevus (NV). To simulate the collaborative scenario of real medical institutions, the data was divided into 4 clients, and non-independent and identically distributed (Non-IID) data was generated through the Dirichlet distribution (α = 0.3) to ensure that the class distributions of each client were skewed but there was partial overlap (for example, the proportion of melanoma in a certain specialized hospital reached 40%, while in a general hospital it was only 5%). The data of each client was divided into a training set and a test set at a ratio of 8:2. The training set was used for local model updates, and the test set was used to evaluate the cross-institutional generalization ability. Image preprocessing included uniformly scaling to 224×224 pixels, mean-standard deviation normalization (parameters: Mean = [0.485, 0.456, 0.406], Std = [0.229, 0.224, 0.225]), and applying random horizontal flipping, rotation (±15°), and brightness adjustment (±10%) to enhance data diversity.

[0075] Secondly, the model used ResNet34 as the backbone network, removed the original fully connected layer, and replaced it with a classification layer with an input dimension of 512 and an output dimension of 3. Local training used the Stochastic Gradient Descent (SGD) optimizer with a momentum of 0.9, a weight decay of 0.001, an initial learning rate of 0.01 (cosine decay to 10% of the original value every 50 rounds), a batch size of 16, and 5 local training epochs. To suppress dimensional collapse, a singular value variance regularization term was introduced, and the total loss function was defined as:

[0076]

[0077] where K is the normalized feature matrix, is the cross-entropy loss. This design forces the singular value distribution in the feature space to be uniform, avoiding the collapse of tail singular values.

[0078] When the system is initialized, the server distributes the global model to 4 clients. In each round of communication, all clients participate in local training:

[0079] (1) Local training: The client calculates the feature z based on local data i Generate the correlation coefficient matrix K, and synchronously optimize the classification loss and the singular value regularization term during backpropagation.

[0080] (2) Anomaly detection: The server identifies abnormal clients through the Euclidean norm ||w i ||, and the model with a deviation from the mean exceeding 2 standard deviations is marked as abnormal.

[0081] (3) Dynamic weight allocation: Adjust the aggregation weights using the following formula:

[0082]

[0083] (4) Geometric median aggregation: The server updates the global model through the weighted median algorithm to reduce the impact of abnormal parameters.

[0084] Finally, after 100 rounds of federated training, the model performance is significantly better than the traditional method:

[0085] Evaluation Metrics FedAvg FedProx MedFedTK Accuracy (ACC) 84.8％ 85.2％ 92.3％ AUC 0.872 0.880 0.941 F1 Score 0.821 0.830 0.902

[0086] This embodiment shows significant technical advantages in actual verification: First, through the comparative analysis of the singular value distribution (as Figure 4 shown), the singular values at the tail of the model do not collapse to zero, and the utilization rate of the feature space is increased by 35% compared with the traditional method, effectively solving the problem of dimensional collapse caused by the degradation of medical image features; Second, in the face of the attack scenario where 10% of the clients are maliciously flipped with melanoma and benign keratosis labels, the system still maintains an accurate diagnosis ability of 89.7% (the FedAvg scheme drops to 72.1%), verifying its strong anti-interference ability; In addition, through dynamic parameter compression and abnormal client filtering, the amount of communication data per round is reduced by 32%, and the transmission time is shortened from 320 ms to 64 ms (the speed is increased by 5 times), providing efficient support for the real-time diagnosis of edge medical devices. The breakthroughs in these three core indicators mark the coordinated optimization of accuracy, security, and efficiency of this solution in the field of medical federated learning.

[0087] Example 3: Pneumonia detection in chest X-ray

[0088] First, a publicly available dataset containing 5,856 chest X-ray images (1,583 normal lungs, 2,780 bacterial pneumonias, and 1,493 viral pneumonias) was used. The data was distributed to 6 clients through the Dirichlet distribution (α = 0.5) to simulate the non-independent and identically distributed (Non-IID) characteristics among medical institutions. In a typical scenario, the proportion of viral pneumonia in a certain client was as high as 80%, while other clients mainly had bacterial pneumonia, fully reflecting the differences in the distribution of real medical data. The training set and the test set were divided at a ratio of 8:2, and two types of poisoning attacks were designed:

[0089] (1) Label flipping attack: Randomly select 20% of the clients and tamper with their training labels (e.g., mark normal images as pneumonia, and swap the viral and bacterial categories);

[0090] (2) Pixel perturbation attack: Add salt-and-pepper noise with a density of 5% to the images of the poisoned clients to simulate malicious data contamination.

[0091] Then, a classification model was constructed based on the lightweight network MobileNetV3, outputting three categories of results: normal / bacterial / viral pneumonia, to adapt to the deployment requirements of edge medical devices. Adam optimizer was used for local training (initial learning rate 0.001, decaying by 50% every 30 rounds), batch size 24, and the number of training rounds was 10. To enhance the anti-interference ability, the loss function integrated a singular value equalization regularization term:

[0092]

[0093] where is the cross-entropy loss, and K is the correlation coefficient matrix of the normalized features. This design effectively inhibits the feature space collapse caused by poisoned data.

[0094] Then, 4 clients were randomly selected in each round of communication to participate in the training, and the total number of training rounds was 100. The server adopted a two-stage robust aggregation strategy: (1) Anomaly detection: Calculate the cosine similarity between the client model parameter w i and the global model w globel . When it is lower than the threshold of 0.7, it is marked as abnormal; (2) Dynamic aggregation: Assign weights according to the following formula to suppress the influence of abnormal clients:

[0095]

[0096] Finally, the eiszfeld algorithm was used to iteratively calculate the geometric median to generate an anti-interference global model. The experimental results verified that:

[0097] (1) Poisoning Defense: When 20% of the clients are subjected to label flipping and noise attacks, the accuracy of MedFedTK only drops by 1.6% (89.1% vs 90.7% normal), while FedAvg drops by 12.1% (76.4% vs 88.5%).

[0098] (2) Feature Visualization ( Figure 5 ): t-SNE dimensionality reduction shows that there is severe class overlap in the feature space of FedAvg (confusing normal and pneumonia), while MedFedTK maintains a clear decision boundary.

[0099] The present invention suppresses dimensional collapse through singular value equilibrium and resists poisoning attacks through geometric median aggregation, achieving a double breakthrough in accuracy and robustness in X-ray pneumonia diagnosis. Combining lightweight models with dynamic communication optimization, it provides a secure and efficient federated learning solution for edge medical devices. It has currently been piloted and deployed in multiple tertiary hospitals to support cross-institutional intelligent pneumonia screening.

[0100] It should be noted that whether it is Example 2 or Example 3, for the misjudgment problem caused by the high inter-class similarity of medical images, the present invention provides an inter-class separability loss function (IS Loss) based on hypersphere space optimization:

[0101]

[0102] Among them, is the angle between the feature and the class weight vector; s is the scaling factor (default 30), and v is the boundary penalty coefficient (default 0.2). By mapping the feature vector and classification weights to the unit sphere, the similarity of different disease classes is dynamically calculated, and the decision boundary interval of easily confused classes (such as melanoma and seborrheic keratosis) is automatically enlarged. Compared with the traditional method, the classification spacing is expanded from 25° to 35°, and the inter-class misjudgment rate on the ISIC2018 dataset is reduced from 18.7% to 4.3%. This mechanism is particularly suitable for distinguishing pneumonia subtypes in chest X-rays and achieves accurate identification with an AUC of 0.93 in the classification task of viral and bacterial pneumonia.

[0103] In the model aggregation stage, this solution innovatively integrates geometric median calculation and proximal gradient optimization techniques. By analyzing the deviation degree of the parameters uploaded by each client (similar to abnormal body temperature detection), the Euclidean norm deviation of the client model parameter w i is calculated:

[0104]

[0105] If score iIf it is >2, it is determined as an abnormal client and its weight is reduced. Automatically identify and reduce the weight impact of abnormal clients, and combine the inertia term constraint to ensure the stability of model update. Experiments show that in the extreme case where 20% of the clients are subjected to label flipping attacks, the model accuracy retention rate can reach 89.1%, which is significantly higher than 76.4% of the traditional federated learning method. This mechanism, combined with the edge-cloud collaborative deployment architecture, realizes lightweight operation with a single-round training time of 15 seconds and a memory occupancy of 1.2GB on edge devices such as Raspberry Pi 4B, meeting the requirements of real-time analysis of emergency images.

[0106] In summary, the present invention demonstrates excellent scalability in multi-modal medical analysis scenarios. In the cross-modal brain tumor segmentation task of CT-MRI, the general component learns the morphological features of tumors, and the personalized component captures modality-specific information (such as calcification points in CT and edema signals in MRI), increasing the Dice coefficient by 8.3%. Actual deployment cases show that for a skin disease diagnosis system jointly trained by 4 medical institutions, the detection rate of melanoma has increased to 93.6%, and the misdiagnosis of benign lesions has decreased by 65%. On the premise of ensuring data privacy, it provides reliable technical support for multi-center medical research.

[0107] A federated learning optimization application for medical images provided by the present invention breaks through the model performance bottleneck under non-independent and identically distributed data through three core technologies. First, the dynamic singular value decoupling strategy (SVB-DLMS) is adopted to expand the convolutional layer parameters in the neural network into a two-dimensional matrix and perform principal component analysis. More than 90% of the core features are automatically selected according to the energy retention criterion to form a general component, and the remaining features are retained locally as personalized components. This process is realized through a dynamic masking mechanism protected by differential privacy, and the feature retention ratio is adaptively adjusted according to the data differences between medical institutions (such as the disease distribution differences between specialized hospitals and general hospitals, and the imaging characteristics differences of equipment). It not only protects privacy but also takes into account the knowledge sharing efficiency, reducing the amount of transmitted data by 40% compared with traditional methods, and increasing the recognition accuracy of melanoma from 84.8% to 92.3% in the dermoscopic image classification task.

[0108] Based on the above embodiments, the present invention analyzes cross-institutional medical data through a dynamic feature decoupling algorithm, can effectively capture the spatial correlation of non-independent and identically distributed features such as pigment distribution in dermoscopic images, and analyze the differential features of X-ray images of viral / bacterial pneumonia. This technology improves the cross-institutional pneumonia classification accuracy by 18%; real-time monitors gradient parameters and the elastic aggregation mechanism, responds to the risk of model dimensional collapse and data poisoning attacks (the AUC decline amplitude ≤0.05 under 20% label flipping attacks), enhances the security and stability of the federated system, and ensures the efficient convergence and accurate decision-making of the medical diagnosis model under privacy protection constraints.

Claims

1. A federated learning aggregation method against dimensional collapse, which is a federated learning of non-independent and identically distributed data based on tracking kernels and robust aggregation, and is characterized in that, The method aims at the dimensional collapse in federated learning and introduces a regularization term to penalize the covariance between singular values during the training process of local models, so as to suppress the phenomenon of dimensional collapse. Specifically, it includes: (1) Calculate the local loss function f i (w), and update the local model parameters based on dimensionality descent; (2) Combining the trace and nuclear norm to approximately represent the singular value variance, and introducing a regularization term to constrain the singular value variance of feature representation. Its expression is: where \(w\) represents the client model parameters, \(D\) represents the training dataset, \(d\) represents the data dimension, \(K\) represents the feature matrix, and \(\|K\|\) * represents the nuclear norm, that is, the sum of singular values; The overall objective of each local model is: min f i (w) + μf singular (w, D), where μ is a hyperparameter; (3) The server calculates the weights of the updates of each local model based on the geometric median. The expression is: where α i is the initial weight, δ is the tolerance parameter, ∨ represents the maximization operation, and ∥w i ∥ is the Euclidean norm of the model parameters; The server performs an aggregation operation based on the received weight β i and the local model update vector of the client to calculate the global model update, expressed as: S t denotes the set of clients selected by the central server to participate in the training in the t-th round of federated learning training; (4) The server aggregates the local models of the clients to generate a global model, and distributes it to the corresponding local models of each client for iterative training until the model converges.

2. The federated learning aggregation method for anti-dimensional collapse according to claim 1, wherein The calculation of the feature matrix K is as follows: First, calculate the output representations z of the N client local models i , and obtain the covariance matrix: where the i-th data point is denoted as z i , is the average value represented by all data points; Then, the outputs of all local models are subjected to vector normalization processing. The mathematical representation is: The covariance matrix z is obtained by the above formula i equal to its eigenmatrix K, resulting in:

3. The federated learning aggregation method for anti-dimensional collapse according to claim 1, wherein During the local training process, the regularization strength is dynamically adjusted to adapt to data scenarios with different degrees of non-independent and identically distributed.

4. A federated learning system for non-independent and identically distributed data, characterized in that, It includes: A client device configured to execute the local model training steps described in any one of claims 1-4; A central server configured to execute the robust aggregation steps described in any one of claims 1-4, including aggregating the local models of the clients to generate a global model and distributing it to the corresponding local models of each client for iterative training; A communication module for transmitting encrypted model parameter updates between the client and the server.

5. A computer-readable storage medium, characterized in that, A computer program is stored, and when the program is executed by a processor, the method described in any one of claims 1-3 is implemented.

6. A federated learning system for medical image classification, characterized in that, The system implements the method described in any one of claims 1-3 for client model training and global model update; it includes: A client deployed in a medical institution, and the client trains a local model through local medical image data and integrates a singular value equalization regularization term; The central server calculates the weights of the updates of each client model based on the geometric median, dynamically excludes the interference of abnormal clients, and updates the global model and distributes the local model training parameters; A communication encryption module: ensuring the transmission privacy of medical image data and model parameters.

7. The federated learning system for medical image classification according to claim 6, wherein The local loss function in this system is an inter-class separability loss function optimized based on the hypersphere space. The mathematical representation is: Among them, is the angle between the feature and the class weight vector; s is the scaling factor, and b is the boundary penalty coefficient. By mapping the feature vector and the classification weight to the unit sphere, the similarity of different disease classes is dynamically calculated.

8. The federated learning system for medical image classification according to claim 6, wherein In the model aggregation stage, by combining geometric median calculation and proximal gradient optimization, the deviation degree of the parameters uploaded by each client is analyzed to calculate the client model parameter w i The Euclidean norm deviation degree of is: If score i > 2, it is determined as an abnormal client and its weight is reduced. Automatically identify and reduce the weight impact of abnormal clients, and combine inertial term constraints to ensure the stability of model updates.

Images