Point cloud data classification method for federal few-sample learning based on privacy protection

By adopting a federated few-sample learning method based on privacy protection in vehicle autonomous driving, using attention mechanism and loss function of federated learning, the problem of model performance improvement and insufficient adaptability in heterogeneous data processing is solved, and point cloud classification is realized The high performance and robustness of the model ensures data privacy and learning efficiency.

CN118918448APending Publication Date: 2024-11-08YANGTZE DELTA REGION INST OF UNIV OF ELECTRONICS SCI & TECH OF CHINE (HUZHOU)
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410904311.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-08
Publication Date
2024-11-08

Smart Images

  • Figure CN118918448A_ABST
    Figure CN118918448A_ABST
Patent Text Reader

Abstract

The invention discloses a point cloud data classification method for federal few-sample learning based on privacy protection, and the method comprises the steps: constructing a local model of point cloud data classification, carrying out the classification training based on a group of obtained point cloud data, transmitting the obtained local model parameters to a global server, carrying out the aggregation through the global server in combination with the parameters of other nodes, and carrying out the classification of the local model. Global model parameters are obtained, and local model parameters are updated based on the global model parameters; wherein the local model performs feature extraction on a point cloud data support set and a query set through an embedded layer, performs adaptive data information enhancement on local prototype features and local query features which are obtained respectively through an attention mechanism to obtain attention parameters, updates the local features based on the attention parameters, and performs data information enhancement on the local prototype features and the local query features based on the updated local prototype features. And the global server aggregates and distributes the updated local features of the local nodes, and the local nodes perform classification prediction of a local model through a learnable layer according to the distributed global features.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and in particular relates to a point cloud data classification method based on privacy-preserving federated few-sample learning. Background Art

[0002] In vehicle autonomous driving, it is necessary to collect and identify 3D point cloud data near the vehicle to obtain environmental information such as objects around the vehicle. On the one hand, due to the limitations of vehicle carriers, classification and recognition models are often installed on local servers of various automobile companies. Data transmission across multiple local servers or local servers across companies is prone to privacy data leakage. On the other hand, data heterogeneity is not handled well, and the effect of improving model performance is limited; lack of adaptability to heterogeneous data, the model is prone to overfitting. The existing problems and shortcomings include:

[0003] 1. Insufficient processing of data heterogeneity, limited effect of improving model performance: Existing technologies often use simple aggregation methods when processing highly heterogeneous data, which cannot effectively capture the complex relationships and features in the data. This results in limited performance improvement when the model processes different types of data.

[0004] 2. Lack of adaptability to heterogeneous data, the model is prone to overfitting: Existing technologies often ignore the heterogeneity of data when training models, resulting in the model being unable to adapt to different types of data in practical applications. This makes the model prone to overfitting, that is, it performs well on training data, but performance degrades in practical applications. Existing technologies usually require a large amount of labeled data to train models. In the real world, especially in the field of autonomous driving, it can be very difficult to obtain labeled data, especially for new or rare categories. Summary of the invention

[0005] In order to solve the shortcomings of the prior art, achieve data privacy protection, and improve the performance and generalization ability of the point cloud classification prediction model, the present invention adopts the following technical solutions:

[0006] A point cloud data classification method based on privacy-preserving federated few-shot learning is applied to local nodes to build a local model for point cloud data classification, and classification training is performed based on a set of acquired point cloud data. The obtained local model parameters are transmitted to a global server so that the global server aggregates the parameters of other nodes to obtain global model parameters, and the local model parameters are obtained and updated based on the global model parameters.

[0007] The local model extracts features from the point cloud data support set and query set through an embedding layer, and adaptively enhances the data information of the local prototype features and local query features obtained respectively through an attention mechanism to obtain attention parameters. The local features are updated based on the attention parameters so that the global server aggregates and distributes the local features updated by each local node, obtains the distributed global features, and performs classification prediction of the local model through a learnable layer.

[0008] By incorporating an attention mechanism to highlight relevant information, which is critical for handling the sparsity and irregularity of point clouds, federated learning enables the model to benefit from a wider range of data variance and real-world scenarios by distributing learning across multiple nodes, which is particularly beneficial for addressing the data scarcity problem common in federated few-shot learning.

[0009] Furthermore, a local loss of federated learning is constructed, including an embedding loss and a learnable loss. The embedding loss updates the embedding layer parameters based on the difference between the predicted category and the true category, and the learnable loss updates the learnable layer parameters based on the similarity between the local prototype feature and the local query feature. The parameters updated by the local loss are sent to the global server, so that the global server constructs the global loss of federated learning, takes the collaborative minimization on each node as the global goal, aggregates the parameters updated by the local loss of each node, updates the global model parameters, and obtains the global model parameters to update the parameters of the local loss of the local node.

[0010] By minimizing the local loss in a federated manner, the local model optimizes the embedding layer and the learnable layer in the distributed network, thereby improving the performance in FSL for 3D point cloud classification. This federated adaptability of the loss function ensures that the learning is effectively distributed across nodes, and each node contributes to the refinement of the global model while preserving data privacy and locality. The global goal emphasizes collaborative minimization of the loss across all nodes. Each node calculates its local loss and updates its parameters, which are then aggregated to update the global model parameters. This process ensures that the global model can benefit from diverse local insights while maintaining data privacy and learning efficiency. Federated learning is particularly effective in scenarios with distributed data sources, enabling the model to learn robust general features and prototypes based on the 3D point cloud classification task of FSL.

[0011] Furthermore, the embedding loss adopts cross entropy loss to calculate the difference between the predicted category score Y of the point cloud data obtained by the local node and the true category label L. The embedding loss formula is as follows:

[0012]

[0013] Among them, N (node) Indicates the number of point cloud data obtained by the local node, C indicates the total number of categories, and They represent the true category label and predicted category score of the i-th point cloud data and the c-th category of the local node respectively.

[0014] Furthermore, the learnable loss is calculated by predicting the local query feature Q of the local node (node) With the local prototype feature P (node) The dissimilarity between them is obtained, and the learning loss formula is as follows:

[0015]

[0016] Among them, N (node) represents the number of point cloud data obtained by the local node, C represents the total number of categories, and d(,) represents the distance function.

[0017] Furthermore, the local loss formula is as follows:

[0018]

[0019] in, represents the embedding loss, represents the learnable loss, represents the computational component loss of the local node, λ1, λ2, and λ3 represent the hyperparameters that control the relative weights of each loss term;

[0020] The global loss formula is as follows:

[0021]

[0022] Among them, θ represents the global parameters including embedding and learnable parameters, K represents the total number of local nodes, and w k represents the contribution weight assigned to the kth local node, represents the embedding loss of the k-th local node, represents the learnable loss of the k-th local node, represents the computational component loss of the k-th local node.

[0023] The overall goal is to optimize the loss parameters of the global model while minimizing the combined loss in a distributed network, where each local node contributes to the learning process without directly sharing data. The overall goal becomes to optimize the global model through local computation and aggregate updates.

[0024] Furthermore, the embedding layer includes an abstract SA layer and a feature propagation FP layer. The abstract SA layer learns local features of a set of divided local point cloud data sets through a shared multi-layer perceptron MLP to obtain a set of converted point cloud data sets and their corresponding features; the feature propagation FP layer refines the features of the conversion points by obtaining parameter information aggregated across different scales from a global server to obtain a refined feature representation F'.

[0025] The key innovation of federated adaptability is to aggregate learned features across nodes after local processing. After each training cycle, the model parameters or learned feature representations of each node are sent to a federated server to perform aggregation operations such as federated averaging to update the global model. The updated global model is redistributed to each node for further training. This federated embedding approach allows the embedding layer PointNet++ to benefit from a diverse dataset without compromising data privacy. Each node contributes to the learning process with its local data to ensure that the global model can fully understand the various point cloud patterns. This approach is particularly beneficial in situations where data cannot be processed centrally due to privacy or logistics constraints.

[0026] Furthermore, the local node supports the subset of the category c Compute local prototype features:

[0027]

[0028] Among them, x ci represents the i-th point cloud data of category c in the support set, Represents x in the local node ci The corresponding category is the local prototype feature of c, Indicates the number of point cloud data corresponding to the support set subset of category c in the local node;

[0029] Local prototype features are used to aggregate to form a global prototype feature p of type c c , this aggregation can be done using techniques such as federated averaging to ensure that the global prototype feature p c represents the average embedding of support examples from all node categories c, based on the local query feature q and the global prototype feature p of the local node c The distance between them is used to obtain the predicted category label of the query set on the local node:

[0030]

[0031] Among them, d(,) represents the distance function.

[0032] Furthermore, the attention mechanism is as follows:

[0033] Attention weight calculation: Each local node independently calculates its local local query feature Q (node) and local prototype feature P (node) The attention weight A Q (node) and A P (node) ; This computation takes into account the specific characteristics and distribution of the data on the node, allowing the mechanism to adaptively emphasize the most informative parts of the local dataset:

[0034] A P (node) =AttentionWeights(Q (node) ,P (node) )

[0035] A Q (node) =AttentionWeights(P (node) ,Q (node) )

[0036] Attention weight normalization: Use the SoftMax function to normalize the local attention weight A Q (node) and A P (node) Normalize and multiply the normalized weights by their corresponding local features element-wise to obtain the attention-weighted features:

[0037] Q′ (node) =Q (node) ⊙SoftMax(A Q (node) )

[0038] P′ (node) =P (node) ⊙SoftMax(A P (node) )

[0039] Update local features: Combine the local attention-weighted features with the original local features to obtain an enhanced local feature set; the key point of this step is that it combines the selective emphasis of the attention mechanism with the original feature set to obtain a more informative feature representation:

[0040] Q″ (node) =Q′ (node) +A Q (node)

[0041] P″ (node) =P′ (node) +A P(node) .

[0042] Furthermore, the aggregation and allocation process of the features is as follows:

[0043] Aggregation of features: The local node sends its locally enhanced local features to the aggregator of the global server for sharing and averaging to update the global model and obtain aggregated global features:

[0044]

[0045] Among them, F (i) represents the local feature at the local node i, N represents the total number of local nodes, and F global Represents the global feature, which is the average value obtained by calculating the local feature representation of all nodes;

[0046] Feature allocation: local nodes obtain global features F global , to ensure that each node in the federated network can update its local model with the newly aggregated global insights. After receiving the global feature representation, the local features of the local node i are updated:

[0047]

[0048] The importance of this redistribution is that it allows each node to benefit from the learning and discoveries of other nodes, thereby improving the overall learning and prediction performance of the federated network. By continuously iterating through these steps of local enhancement, global aggregation, and redistribution, the FL model gradually improves and becomes more robust and accurate over time. This enhanced federated approach in the feature transfer phase ensures a synergistic combination of local learning and global model enhancement. The localized attention mechanism allows for detailed feature extraction tailored to each node's data, while the global aggregation phase ensures that these individuals have insights that contribute to the collective wisdom of the model. As a result, the model not only exhibits improved performance in the task, but also shows a high degree of adaptability and generalization, which is critical for practical applications with data diversity and privacy protection.

[0049] A point cloud data classification method based on privacy-preserving federated few-shot learning is applied to the global server to obtain the local model parameters obtained by the local node through its local model classification training on a set of point cloud data, aggregate the model parameters of all nodes, and distribute the obtained global model parameters to each local node to update the parameters of its local model, and perform classification prediction of the local model through its class learning layer;

[0050] The local model parameters are obtained by extracting features from the point cloud data support set and query set through the embedding layer of the local node, and then adaptively enhancing the data information of the extracted features through the attention mechanism, and then updating the local features according to the attention parameters corresponding to the information enhancement;

[0051] By incorporating an attention mechanism to highlight relevant information, which is critical for handling the sparsity and irregularity of point clouds, federated learning enables the model to benefit from a wider range of data variance and real-world scenarios by distributing learning across multiple nodes, which is particularly beneficial for addressing the data scarcity problem common in federated few-shot learning.

[0052] The advantages and beneficial effects of the present invention are:

[0053] 1. Enhanced data privacy: Through the federated learning framework, the present invention can use the data of multiple vehicle local model nodes for model training without sharing the original data, thereby effectively protecting data privacy.

[0054] 2. Improve model performance: By aggregating data from local model nodes of different vehicles through a central server, the present invention can obtain a more comprehensive data perspective, thereby improving the performance and generalization ability of the model.

[0055] 3. Promote collaborative learning: The federated learning framework makes collaborative learning between different entities possible, which helps to share and integrate their respective data and knowledge.

[0056] 4. Reduce the need for data centralization: In some cases, data cannot be processed centrally due to privacy or logistics restrictions. The present invention can process data while maintaining data distribution through federated learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a 3DFFL framework structure diagram for performing 3D point cloud processing on an autonomous driving vehicle in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0059] like Figure 1 As shown, the process of collaborative training in which the autonomous driving vehicle package VP (Vehicle Package) collects data, the local headquarters HQ performs preprocessing and initial model training, and the central server aggregates and redistributes the updated global model weights is demonstrated to enhance classification accuracy while maintaining data integrity and protecting data privacy.

[0060] Specifically, each vehicle package VP represents a self-contained unit of a vehicle, including an autonomous vehicle equipped with 3D data acquisition tools (such as LiDAR) and a communication system for data relay. The raw point cloud data collected by VPs are pre-processed and verified at the headquarters HQ of the respective automotive companies. These HQs are responsible for initially training the local model weights, ensuring that the data is complete and suitable for the global learning process. These locally trained weights are then securely transmitted to the central server, where they are aggregated to enhance the shared global model. This aggregation process allows collaborative learning between different entities while maintaining the privacy of each participant's data. The central server then distributes the updated global model weights back to each HQ, promoting continuous improvement of local models and protecting the confidentiality of the data. This iterative process ensures that the global model benefits from diverse, privacy-focused contributions, driving the performance and reliability of autonomous driving technology through 3DFFL. In addition, the proposed model for few samples such as Figure 1 As shown in the figure, the query features and prototype features are enhanced by adding an attention mechanism. This attention mechanism selectively emphasizes relevant information and then applies the SoftMax function to the attention features. The attention features are multiplied by the output of SoftMax and added to the attention features to obtain updated features. These updated features are then transmitted to the learnable prototype network ProtoNet (Prototypical Networks) for point cloud classification prediction. The benefits provided by the attention mechanism include selective emphasis, enhancing the model's ability to capture discriminative patterns, and improving the accuracy of point cloud classification.

[0061] In one embodiment of the present invention, federated learning FL is integrated into the above method, and the adaptive loss meaning of FL is constructed to adapt to the distributed characteristics of FL. Its loss calculation includes federated embedding loss and federated learnable loss, both of which are calculated locally at each federated node (local headquarters HQ) and then aggregated to the federated server (central server) to update the global model.

[0062] Federated Embedding Loss: The embedding loss of each federated node is calculated using the cross entropy loss, which measures the difference between the predicted category score Y and the true label L of the local sample set (collected 3D point cloud data) of the federated node. The embedding loss formula of the node is as follows:

[0063]

[0064] Among them, N (node) represents the number of samples of the federated node, C represents the total number of categories, and denote the true label and predicted category score of the i-th sample and c-th category of the federated node, respectively.

[0065] Federated Learnable Loss: The learnable loss of each federated node uses the concept of local prototypes to measure the local query sample Q of the federated node. (node) With the local prototype vector P (node) The dissimilarity between nodes, the learnable loss formula of the node is as follows:

[0066]

[0067] Among them, d(,) represents the distance function.

[0068] Local overall loss: The overall loss of each federated node is combined with the local embedding loss and local learnable loss Then, the local overall loss is aggregated across all nodes to update the global model. The local overall loss formula for a node is as follows:

[0069]

[0070] Among them, λ1, λ2, and λ3 represent the hyperparameters that control the relative weights of their respective loss terms. represents the loss of the computational components used by the local network of the federated nodes. By minimizing the overall loss in a federated manner, the model optimizes the embeddings and learnable components in the distributed network, leading to improved performance in FSL for 3D point cloud classification. This federated adaptability of the loss function ensures that learning is effectively distributed across nodes, with each node contributing to the refinement of the global model while preserving data privacy and locality.

[0071] Overall goal: When integrating FL into the overall goal of feature extraction and learning methods, the goal is to optimize the parameters of the global model while minimizing the combined loss in the distributed network. This involves adapting the goal to fit the federated framework, where each node contributes to the learning process without directly sharing data. The overall goal in the federated setting becomes a collaboration to optimize the global model through local computation and aggregate updates. The overall federated goal expression is as follows:

[0072]

[0073] where θ represents the global parameters of the combined model (including embeddings and learnable components), K is the total number of nodes involved in the FL process, represents the local embedding loss calculated at the kth federated node, represents the local learnable loss calculated at the kth federated node, represents the loss of the computational component computed at the kth federated node, w k It represents the contribution weight assigned to the kth federated node, which can be based on the number of samples or the computing power of the node.

[0074] The overall goal emphasizes collaborative minimization of loss across all nodes. Each node computes its local loss and updates its parameters, which are then aggregated to update the global model parameters θ. This process ensures that the global model can benefit from diverse local insights while maintaining data privacy and learning efficiency. The federated approach is particularly effective in scenarios with distributed data sources, enabling the model to learn robust general features and prototypes based on the 3D point cloud classification task of FSL.

[0075] In one embodiment of the present invention, federated few-shot learning (FSL) is integrated into the above method, and FL is combined with PointNet++ (embedding) and ProtoNet (learnable) components to build a model that can accurately classify new point clouds into their respective categories, even though each category has limited labeled examples. The model leverages the FL framework to enable it to learn from a variety of datasets without data sharing, maintaining privacy and data diversity. The model is designed to effectively capture discriminative features from input point clouds. It combines attention mechanisms to highlight relevant information, which is critical for dealing with the sparsity and irregularity of point clouds. The FL method enables the model to benefit from a wider range of data differences and real-world scenarios by distributing learning across multiple nodes, which is particularly beneficial for solving the data scarcity problem commonly seen in federated few-shot federated FSL.

[0076] In addition, the learnable component for classification, represented by ProtoNet, is adapted to work within a federated framework, which ensures that the model can not only learn common features across distributed datasets, but also retains robustness on unknown point clouds. The ultimate goal is to achieve accurate and robust classification performance on new point clouds, effectively overcoming the inherent problems of FSL in the 3D field, while also benefiting from the cooperation and privacy protection capabilities of FL.

[0077] Adaptability of PointNet++ (embedding) in federated learning:

[0078] The PointNet++ architecture is an extension of the original PointNet, designed specifically for point cloud analysis, and further enhanced by incorporating federated learning FL technology, which is critical to addressing the challenge of data scarcity while ensuring data privacy and taking full advantage of the diversity of data sources. PointNet++ runs on multiple nodes (such as local devices or servers), each with its own subset of data, and the layered architecture includes an abstraction (SA) layer and a feature propagation (FP) layer, which is suitable for distributed operation.

[0079] Federated SA layer: Each node receives a subset of the input point set X and applies the SA layer locally, which consists of partitioning the local point set into smaller regions S = {S1, S2, .., SM}, where M represents the number of regions formed within each node. The local features of these regions are learned through a shared multi-layer perceptron (MLP), and the output of each node is a set of transformed points X' and their corresponding features F.

[0080] Federated FP layer: Similar to the SA layer, the FP layer is performed locally on each node, which refines the features of the transformation points by aggregating information across different scales to obtain the refined feature representation F'. After local processing, the key innovation of federated adaptability is to aggregate the learned features across nodes. After each training cycle, the model parameters or learned feature representations of each node are sent to a federated server to perform aggregation operations such as federated averaging to update the global model. The updated global model is redistributed to each node for further training. This federated embedding approach allows PointNet++ to benefit from a diverse dataset without compromising data privacy. Each node contributes to the learning process with its local data to ensure that the global model can fully understand various point cloud patterns. This method is beneficial in cases where data cannot be processed centrally due to privacy or logistics constraints.

[0081] The adaptability of FL ensures the collaborative nature of the learning process while preserving privacy. The global model becomes more robust and generalizable by incorporating diverse data insights, making it well suited for real-world applications where data is distributed and diverse. FL is combined with ProtoNet for FSL to learn embeddings in a privacy-preserving manner, where the process of learning and using prototypes for classification is distributed across multiple nodes in the federated network.

[0082] Federated Prototype Computation: In FL, each node in the network holds a subset of support examples (3D point clouds) for each category c The local prototype feature of each node is calculated as follows:

[0083]

[0084] Among them, x ci represents the i-th 3D point cloud data of category c, represents the local prototype feature of category c at a specific node, represents the number of supporting examples of category c at the local node. Then, the local prototype features are aggregated to form the global prototype features. This aggregation can be done using techniques such as federated averaging to ensure that the global prototype feature p c represents the average embedding from all support examples of node class c.

[0085] Federated classification: In a federated context, query examples may be distributed on different nodes. Each node calculates the distance between its local query example and the global prototype to obtain the predicted category label of the query example on the local node:

[0086]

[0087] Among them, q represents the query example feature, p c represents the global prototype feature of category c, and d(,) represents the distance function.

[0088] Integrating FL into the embedding learning process of ProtoNet enables the model to effectively utilize diverse data sources. Each node contributes to forming a more robust and representative global prototype by computing a local prototype based on its available data. This approach not only protects data privacy, but also enhances the generalization ability of the model on diverse datasets. Therefore, the federated version of ProtoNet becomes more adaptable and efficient in scenarios where data cannot be centralized or privacy issues are critical.

[0089] Enhanced FL in Feature Transfer:

[0090] Integrating FL into the feature transfer process involves a careful adaptation of the attention mechanism to a collaborative learning setting. This adaptation aims to optimize the feature extraction and classification process in a distributed network of nodes, each of which possesses a unique subset of data, thereby improving the overall performance and robustness of the model.

[0091] Attention Mechanism: In a federated setting, each participating node computes its local query feature Q (node) and prototype feature P (node) The attention mechanism is customized to run at the node level and includes the following steps:

[0092] 1) Attention weight calculation: Each node independently calculates the attention weight of its local features. This calculation takes into account the specific characteristics and distribution of the data on the node, allowing the mechanism to adaptively emphasize the most informative parts of the local dataset;

[0093] A P (node) =AttentionWeights(Q (node) ,P (node) )

[0094] A Q (node) =AttentionWeights(P (node) ,Q (node) )

[0095] 2) SoftMax application: Then use the SoftMax function to adjust the local attention weight (A Q (node) , A P (node) ) are normalized. By performing element-wise multiplication of these normalized weights with the local features, the attention-weighted features are obtained:

[0096] Q′ (node) =Q (node) ⊙SoftMax(A Q (node) )

[0097] P′ (node) =P (node) ⊙SoftMax(A P (node) )

[0098] 3) Update local features: Combine the local attention-weighted features with the original features to obtain an enhanced feature set. The key point of this step is that it combines the selective emphasis of the attention mechanism with the original feature set to obtain a more informative feature representation.

[0099] Q″ (node) =Q′ (node) +A Q (node)

[0100] P″ (node) =P′ (node) +A P (node)

[0101] 4) Aggregation of federated features: After local feature enhancement, the model enters the global feature aggregation stage. This process is crucial for FL because it integrates diverse insights from various nodes;

[0102] Specifically, feature sharing and aggregation: Each node in the federated network calculates its local feature representation and then periodically shares these features with the central aggregator, which performs federated averaging on all nodes to update the global model. The aggregated global feature table is as follows:

[0103]

[0104] Among them, F (i) is the feature representation at node i, N is the total number of nodes in the federated network, and F global is the global feature representation, which is the average value obtained by calculating the local feature representations of all nodes.

[0105] This equation represents how the global model aggregates local insights by averaging the feature representations of each node.

[0106] 5) Redistribution of global model:

[0107] Once the global feature representation F is calculated global , it is redistributed back to all nodes in the network to ensure that each node in the federated network can update its local model with the newly aggregated global insight. After receiving the global feature representation, the local model update at node i is represented as:

[0108]

[0109] in, represents the updated local feature representation at node i after incorporating insights from the global model. The importance of this redistribution is that it allows each node to benefit from the learning and discoveries of other nodes, thereby improving the overall learning and prediction performance of the federated network. By continuously iterating through these steps of local enhancement, global aggregation, and redistribution, the FL model gradually improves and becomes more robust and accurate over time. This enhanced federated approach in the feature transfer phase ensures a synergistic combination of local learning and global model enhancement. The localized attention mechanism allows for detailed feature extraction tailored to the data of each node, while the global aggregation phase ensures that these individuals have insights that contribute to the collective wisdom of the model. As a result, the model not only exhibits improved performance in the task, but also shows a high degree of adaptability and generalization, which is critical for practical applications with data diversity and privacy protection.

[0110] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A point cloud data classification method based on privacy-preserving federated few-shot learning is applied to local nodes and is characterized by: Construct a local model for point cloud data classification, and perform classification training based on a set of acquired point cloud data, transmit the obtained local model parameters to the global server so that the global server aggregates the parameters of other nodes to obtain the global model parameters, obtain and update the local model parameters based on the global model parameters; The local model extracts features from the point cloud data support set and query set through an embedding layer, and adaptively enhances the data information of the local prototype features and local query features obtained respectively through an attention mechanism to obtain attention parameters. The local features are updated based on the attention parameters so that the global server aggregates and distributes the local features updated by each local node, obtains the distributed global features, and performs classification prediction of the local model through a learnable layer.

2. The point cloud data classification method based on privacy-preserving federated few-sample learning according to claim 1 is characterized in that: Constructing a local loss for federated learning, including an embedding loss and a learnable loss, wherein the embedding loss updates the embedding layer parameters based on the difference between the predicted category and the true category, and the learnable loss updates the learnable layer parameters based on the similarity between the local prototype feature and the local query feature; sending the updated parameters of the local loss to the global server, so that the global server constructs the global loss for federated learning, takes the collaborative minimization on each node as the global goal, aggregates the updated parameters of the local loss of each node, updates the global model parameters, and obtains the global model parameters to update the parameters of the local loss of the local node.

3. The point cloud data classification method based on privacy-preserving federated few-sample learning according to claim 2, characterized in that: The embedding loss adopts cross entropy loss to calculate the difference between the predicted category score Y of the point cloud data obtained by the local node and the true category label L. The embedding loss formula is as follows: Among them, N (node) Indicates the number of point cloud data obtained by the local node, C indicates the total number of categories, and They represent the true category label and predicted category score of the i-th point cloud data and the c-th category of the local node respectively.

4. The point cloud data classification method based on privacy-preserving federated few-sample learning according to claim 2, characterized in that: The learnable loss is calculated by predicting the local query feature Q of the local node. (node) With the local prototype feature P (node) The dissimilarity between them is obtained, and the learning loss formula is as follows: Among them, N (node) represents the number of point cloud data obtained by the local node, C represents the total number of categories, and d(,) represents the distance function.

5. The point cloud data classification method based on privacy-preserving federated few-sample learning according to claim 2, characterized in that: The local loss formula is as follows: in, represents the embedding loss, represents the learnable loss, represents the computational component loss of the local node, λ1, λ2, and λ3 represent the hyperparameters that control the relative weights of each loss term; The global loss formula is as follows: Among them, θ represents the global parameters including embedding and learnable parameters, K represents the total number of local nodes, and w k represents the contribution weight assigned to the kth local node, represents the embedding loss of the k-th local node, represents the learnable loss of the k-th local node, represents the computational component loss of the k-th local node.

6. The point cloud data classification method based on privacy-preserving federated few-sample learning according to claim 1, characterized in that: The embedding layer includes an abstraction layer and a feature propagation layer. The abstraction layer learns local features of a group of divided local point cloud data sets through a shared multi-layer perceptron to obtain a group of converted point cloud data sets and their corresponding features. The feature propagation layer refines the features of the conversion points by acquiring parameters aggregated across different scales of the global server to obtain refined features.

7. The point cloud data classification method based on privacy-preserving federated few-sample learning according to claim 1, characterized in that: The local node supports the subset of the class c Compute local prototype features: Among them, x ci represents the i-th point cloud data of category c in the support set, Represents x in the local node ci The corresponding category is the local prototype feature of c, Indicates the number of point cloud data corresponding to the support set subset of category c in the local node; Local prototype features are used to aggregate to form a global prototype feature p of type c c , the local node is based on the local query feature q and the global prototype feature p c The distance between them is used to obtain the predicted category label of the query set on the local node: Among them, d(,) represents the distance function.

8. The point cloud data classification method based on privacy-preserving federated few-sample learning according to claim 1, characterized in that: The attention mechanism is as follows: Attention weight calculation: Each local node independently calculates its local local query feature Q (node) and local prototype feature P (node) The attention weight A Q (node) and A P (node) : A p (node) =AttentionWeights(Q (node) ,P (node) ) A Q (node) =AttentionWeights(P (node) ,Q (node) ) Attention weight normalization: Use the SoftMax function to normalize the local attention weight A Q (node) and A P (node) Normalize and multiply the normalized weights by their corresponding local features element-wise to obtain the attention-weighted features: Q′ (node) =Q (node) ☉SoftMax(A` Q (node) ) P′ (node) =P (node) ☉SoftMax(A P (node) ) Update local features: Combine the local attention-weighted features with the original local features to obtain enhanced local features: Q ″ (node) =Q′ (node) +A Q (node) P″ (node) =P′ (node) +A P (node) 。 9. The point cloud data classification method based on privacy-preserving federated few-sample learning according to claim 8, characterized in that: The aggregation and allocation process of the features is as follows: Aggregation of features: The local node sends its locally enhanced local features to the aggregator of the global server for sharing and averaging to update the global model and obtain aggregated global features: Among them, F (i) represents the local feature at the local node i, N represents the total number of local nodes, and F global Represents global features; Feature allocation: local nodes obtain global features F global , update the local features of local node i:

10. A point cloud data classification method based on privacy-preserving federated few-shot learning, applied to a global server, is characterized by: Obtain the local model parameters obtained by the local node through its local model to classify a set of point cloud data, aggregate the model parameters of all nodes, and distribute the obtained global model parameters to each local node to update the parameters of its local model, and perform classification prediction of the local model through its class learning layer; The local model parameters are obtained by extracting features from the point cloud data support set and query set through the embedding layer of the local node, and then adaptively enhancing the data information of the extracted features through the attention mechanism, and then updating the local features according to the attention parameters corresponding to the information enhancement.

Citation Information

Cited By

  • Longitudinal graph federal learning method based on multi-head attention mechanism

    CN120069008A

  • Sensitive data identification desensitization method and system based on federal global prototype

    CN120493080A