Federal learning generalization method and system for network traffic

Abstract

The invention discloses a federated learning generalization method and system for network traffic, and belongs to the technical field of communication networks, and the method comprises the steps: firstly collecting data summary of different network nodes in a federated learning environment, and dynamically selecting the network nodes as participating clients of federated learning through an intelligent node selection method; and determining a network node set participating in each round of model training. And automatically grouping a network node set participating in each round of model training into a plurality of groups with similar data distribution by utilizing a self-adaptive clustering algorithm. A model branch is independently trained for a group through a group exclusive model training method based on adaptive clustering, and learning results of all model branches are integrated into a main model through a multi-model branch intelligent fusion method, so that the main model can integrate knowledge of each model branch. According to the method, the generalization ability of the model on unseen data distribution can be remarkably improved, and the data flow service identification precision of network communication is further improved.

Description

Technical Field

[0001] The present invention relates to the field of communication network technology, and in particular to a generalized learning algorithm for class imbalance and non-independent identically distributed (Non-IID) data problems in a federated learning environment in a communication network and an implementation method thereof. Background Art

[0002] Federated learning, as a distributed machine learning framework, allows multiple nodes to jointly train a shared model while keeping their own data locally, thereby protecting data privacy. However, in communication networks, since node data often exhibits obvious diversity and imbalance, such as class imbalance and non-independent and identically distributed (Non-IID) characteristics of data distribution, it is difficult for the trained model to achieve good generalization performance on unseen data. Especially in practical applications, this data heterogeneity poses a huge challenge to the training efficiency and generalization ability of the model. Summary of the invention

[0003] Purpose of the invention: In a communication network, within a federated learning framework, the data held by each node (router, base station, etc.) often presents significant diversity and imbalance, which poses a challenge to the learning efficiency and generalization ability of the model. In order to improve the generalization ability of the model under class imbalance and non-independent identically distributed (Non-IID) data conditions, the present invention provides a federated learning generalization method for network traffic, which optimizes the use of data and the training process of the model by introducing an intelligent node selection mechanism and an adaptive clustering algorithm, thereby improving the generalization ability of the model.

[0004] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

[0005] A generalized method for federated learning of network traffic, comprising the following steps:

[0006] Step 1: Collect data summaries of different network nodes in the federated learning environment through the data summary collection method.

[0007] Step 2: Dynamically select network nodes as participating clients of federated learning through an intelligent node selection method to determine the set of network nodes participating in each round of model training.

[0008] Step 3: Use an adaptive clustering algorithm to automatically group the set of network nodes participating in each round of model training into several groups with similar data distributions based on the data characteristics of the nodes.

[0009] Step 4: Train a model branch independently for the group obtained in step 3 using a group-specific model training method based on adaptive clustering. Each model branch focuses on learning the data distribution characteristics within the group.

[0010] Step 5: Integrate the learning results of all model branches into the main model through the multi-model branch intelligent fusion method, so that the main model can integrate the knowledge of each model branch.

[0011] Preferred: The model parameter aggregation formula in the multi-model branch intelligent fusion method is as follows: in, Indicates group Exclusive model of Indicates group Local model update, are the weights of the two aggregation methods, represents the number of nodes in the rth group, For Node The amount of data, is the data volume of node R, For Node The similarity and diversity comprehensive score, For Node The comprehensive score of similarity and diversity.

[0012] Preferably: the method of dynamically selecting network nodes by the intelligent node selection method in step 2: obtaining a comprehensive score based on the data summary information of different network nodes, and dynamically selecting a set of network nodes participating in each round of model training based on the comprehensive score.

[0013] Preferred: The comprehensive score is calculated by the following formula: in, Indicates the comprehensive rating. are the weights of similarity and diversity scores, , , is the similarity weighting coefficient, , is the diversity weighting coefficient, L is the number of labels, K is the number of features, is the mean of all users on feature k, represents the feature mean of node i on feature k, is the standard deviation of all users on feature k, represents the standard deviation of node i on feature k, is the feature distribution of label l for all users, represents the feature distribution of node i in label l, is the entropy of the global model, Represents global data distribution, Indicates the data added to node i The uncertainty of the post-global model, Indicates the amount of data, represents the probability distribution of node i on feature x, represents the probability distribution of global data on feature x, Indicates characteristics.

[0014] Preferably: the intelligent node selection method in step 2 comprises the following steps:

[0015] Step 21), calculate the comprehensive score for all nodes .

[0016] Step 22), follow Sort the top P nodes from high to low and select them to participate in the training.

[0017] Step 23), the node participates in the feedback module.

[0018] Step 24), after each round of training, collect the node participation and training results, and update the node selection strategy. The weights are updated using the following formula: in, represents the similarity score weight, represents the diversity score weight, Indicates Similarity score weights for round training, Indicates The diversity score weights for round training, represents the differential, represents the training loss, is the learning rate.

[0019] Preferred: In step 3, an adaptive clustering algorithm is used to automatically group the set of network nodes participating in each round of model training into several groups with similar data distribution according to the data characteristics of the nodes: a feature vector is extracted from the data set of each node, and the similarity between the nodes is calculated using a similarity measurement algorithm based on the feature vector, and the nodes are grouped into several clusters based on the similarity. The adaptive clustering algorithm can adaptively adjust the clustering strategy according to real-time data to dynamically form the optimal node group.

[0020] Preferably, the feature vector is a statistic extracted from the data set of each node or a low-dimensional representation obtained by a dimensionality reduction method.

[0021] Preferably: the data summary includes data volume, data distribution type, and data label distribution.

[0022] Preferably: the network node is a router or a base station.

[0023] A network traffic federated learning generalization system, used to implement the above-mentioned network traffic federated learning generalization method, includes a data summary collection unit, an intelligent node selection unit, an adaptive clustering unit, a dedicated model training unit, and a multi-model branch intelligent fusion unit, wherein:

[0024] The data summary collection unit is used to collect data summaries of different network nodes in a federated learning environment.

[0025] The intelligent node selection unit is used to dynamically select network nodes as participating clients of federated learning and determine the set of network nodes participating in each round of model training.

[0026] The adaptive clustering unit is used to automatically group the set of network nodes participating in each round of model training into a number of groups with similar data distribution according to the data characteristics of the nodes.

[0027] The exclusive model training unit independently trains a model branch for a group through a group-exclusive model training method based on adaptive clustering, and each model branch focuses on learning the data distribution characteristics within the group.

[0028] The multi-model branch intelligent fusion unit is used to integrate the learning results of all model branches into the main model, so that the main model can integrate the knowledge of each model branch.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention optimizes the use of data and the training process of the model by introducing an intelligent node selection mechanism and an adaptive clustering algorithm. The intelligent node selection mechanism dynamically selects nodes participating in model training based on the similarity and diversity of data distribution to improve training efficiency. The adaptive clustering algorithm automatically forms node groups with similar data distributions according to the characteristics of the node data, trains a dedicated model branch for each group, and integrates the learning results of all model branches into the main model through a multi-model branch intelligent fusion method, so that the main model can integrate the knowledge of each model branch, thereby significantly improving the generalization ability of the model on unseen data distributions. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is an intelligent node selection method.

[0032] Figure 2 It is an adaptive clustering algorithm. DETAILED DESCRIPTION

[0033] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0034] Example 1

[0035] Since the services of data traffic in network nodes (routers, base stations, etc.) include games, videos, social networking, and WEB browsing, the data held by network nodes often presents significant diversity and imbalance, which poses a challenge to the learning efficiency and generalization ability of the federated learning framework model, resulting in insufficient recognition progress. This embodiment provides a generalization method for federated learning of network traffic, which is specially designed to deal with non-independent identically distributed (Non-IID) and class imbalance problems, thereby improving the recognition accuracy of service traffic in the network, and treating different network nodes (routers, base stations, etc.) as participating clients of federated learning, including the following steps:

[0036] Step 1: collect data summaries of different network nodes in the federated learning environment through a data summary collection method for use in subsequent intelligent node selection and adaptive clustering algorithms. The data summary includes data volume, data distribution type, and data label distribution. The network node is a router.

[0037] The data summary collection method requires that the necessary data summaries can be effectively collected from each node, while ensuring that the data collection process does not infringe on the data privacy of the node. Responsible for collecting data summary information from each node, such as data volume, data distribution type, data label distribution, etc. This information is used for subsequent similarity and diversity score calculations. Data collection frequency: The data summary information collection cycle of each node is T, that is, the data summary information is collected every T time units. The specific T value can be adjusted according to the specific application scenario.

[0038] The data summary collection method includes the following steps:

[0039] Step 11), data volume

[0040] The total amount of data for each node is expressed as: in, Indicates the total amount of data for each node. is the number of data samples held by node i, is the size of the jth data sample of node i at time t.

[0041] Step 12), data distribution type:

[0042] By counting the feature distribution of each node data, the feature mean and standard deviation are calculated: in, represents the feature mean of node i on feature k, represents the standard deviation of node i on feature k, is the value of the j-th sample of node i at time t on the k-th feature.

[0043] Step 13), data label distribution:

[0044] Statistics of label distribution of different labels for each node: in, represents the feature distribution of node i in label l, is the number of samples with label l at node i at time t.

[0045] In each collection cycle T, each node i needs to complete the calculation and upload of data summary information. After each collection cycle T, it is uploaded to the central server for subsequent intelligent node selection and adaptive clustering algorithm.

[0046] Step 2, such as Figure 1 As shown in the figure, the intelligent node selection method dynamically selects network nodes as participating clients of federated learning, and determines the set of network nodes participating in each round of model training. This mechanism calculates similarity and diversity indicators based on the data summary information of different network nodes, and dynamically selects the set of nodes participating in each round of model training based on the comprehensive scores of these indicators, so as to optimize the efficiency of data use during training and improve the generalization ability of the model.

[0047] In another embodiment, a method for dynamically selecting network nodes by an intelligent node selection method is as follows: a comprehensive score is obtained based on data summary information of different network nodes, and a set of network nodes participating in each round of model training is dynamically selected based on the comprehensive score.

[0048] Calculate the data similarity and diversity scores for each node. The similarity score measures the similarity between the node data and the global data or other node data. The diversity score measures the uniqueness of the node data to ensure that the selected node data is representative and diverse. The method to obtain the comprehensive score in step 2 is as follows:

[0049] The similarity score measures the consistency of node data with the global data distribution. Assuming there are N nodes, the data summary information of each node i includes the amount of data , the feature mean of node i on feature k and the standard deviation of node i on feature k , feature distribution of node i in label l .

[0050] (1) Feature mean similarity :

[0051] in, is the mean of all users on feature k, where K is the number of features.

[0052] (2) Feature standard deviation similarity : in, is the standard deviation of all users on feature k.

[0053] (3) Data label distribution similarity : in, is the feature distribution of label l for all users, and L is the number of labels.

[0054] Finally, the similarity score It can be obtained by taking the weighted average of the above similarities: in, , , is the similarity weighting coefficient, satisfying .

[0055] Diversity score calculation

[0056] Diversity scoring evaluates the novelty and complementarity that node data can bring to model training.

[0057] (1) Information gain calculation:

[0058] Information gain measures the improvement of node data to the global model knowledge. The formula is as follows: in, represents information gain, is the entropy of the global model, represents the uncertainty of the model’s current knowledge, is the global data distribution, Indicates the data added to node i Uncertainty in the post-global model.

[0059] (2) Data uniqueness calculation:

[0060] Data uniqueness is quantified by evaluating the distance between node data and existing global data. Assume that the data feature distribution of node i is , the characteristic distribution of global data is , calculate the distribution distance: in, represents the distribution distance, KL represents the Kullback-Leibler divergence, represents the probability distribution of node i on feature x, represents the probability distribution of global data on feature x, Indicates characteristics.

[0061] Finally, diversity scores It can be obtained by taking the weighted average of the above diversities: in, represents the diversity score, , is the diversity weighted coefficient, satisfying , represents information gain, represents the distribution distance, Represents a node.

[0062] Based on the comprehensive similarity score and diversity scores Get the final comprehensive score : in, Indicates the comprehensive rating. represents the comprehensive similarity score, represents the diversity score, are the weights of similarity and diversity scores, ranging from .

[0063] Therefore, the comprehensive score is calculated as follows: in, Indicates the comprehensive rating. are the weights of similarity and diversity scores, , , is the similarity weighting coefficient, , is the diversity weighting coefficient, L is the number of labels, K is the number of features, is the mean of all users on feature k, represents the feature mean of node i on feature k, is the standard deviation of all users on feature k, represents the standard deviation of node i on feature k, is the feature distribution of label l for all users, represents the feature distribution of node i in label l, is the entropy of the global model, Represents global data distribution, Indicates the data added to node i The uncertainty of the post-global model, Indicates the amount of data, represents the probability distribution of node i on feature x, represents the probability distribution of global data on feature x, Indicates characteristics.

[0064] These calculation results are used to dynamically select the set of nodes participating in model training to optimize the data usage efficiency and model generalization ability during training.

[0065] The intelligent node selection method in step 2 includes the following steps:

[0066] Step 21), calculate the comprehensive score for all nodes .

[0067] Step 22), follow Sort the top P nodes from high to low and select them to participate in the training.

[0068] Step 23), the node participates in the feedback module.

[0069] Step 24), after each round of training, collect the node participation and training results, and update the node selection strategy. The weights are updated using the following formula: in, represents the similarity score weight, represents the diversity score weight, Indicates Similarity score weight for round (current round) training, Indicates The diversity score weight of the round (current round) training, represents the differential, represents the training loss, is the learning rate.

[0070] Specific steps:

[0071] (1) Initialization: Setting the initial similarity and diversity score weights and .

[0072] (2) Data summary information collection: Collect data summary information from all nodes .

[0073] (3) Score calculation Calculate the similarity score of each node and diversity scores , and then calculate the comprehensive score .

[0074] (4) Node selection: based on comprehensive score Sort and select the first P nodes to participate in training.

[0075] (5) Training and feedback: Perform model training, collect feedback, and update weights and .

[0076] (6) Loop: Repeat steps 2 to 5 until the training stop condition is met.

[0077] Through this intelligent node selection mechanism, the federated learning system can dynamically select the optimal nodes to participate in training, improving training efficiency and the generalization ability of the model.

[0078] Step 3, such as Figure 2 As shown, an adaptive clustering algorithm is used to automatically group the set of network nodes participating in each round of model training into several groups with similar data distributions according to the data characteristics of the nodes.

[0079] In step 3, the adaptive clustering algorithm is used to automatically group the set of network nodes participating in each round of model training into several groups with similar data distribution according to the data characteristics of the nodes: extract feature vectors from the data set of each node, mainly including the statistical characteristics of the data (such as mean, variance), data distribution, feature vectors, etc. The total data volume of node i is expressed as , whose eigenvector is expressed as . Eigenvector This can be a statistic extracted from the dataset at each node or a low-dimensional representation obtained through dimensionality reduction methods such as PCA.

[0080] The similarity between nodes is calculated using a similarity measurement algorithm based on the feature vector, and the nodes are grouped into several clusters based on the similarity. Assume there are N nodes, and the feature vector of each node i is .

[0081] Similarity It can be calculated as follows:

[0082] Cosine Similarity: Euclidean distance: Similarity can be calculated using cosine similarity or Euclidean distance. Indicates similarity, is the feature vector of node i, is the feature vector of node j.

[0083] You can use K-means clustering algorithm or other clustering algorithms (such as DBSCAN) to cluster nodes into several clusters. Suppose we use K-means clustering:

[0084] (1) Initialize Q cluster centers , , ..., , ..., .

[0085] (2) Assign each node i to the nearest cluster center: (3) Update cluster center: in, represents the set of nodes in cluster q, is the feature vector of node h.

[0086] (4) Repeat steps 2 and 3 until the cluster center no longer changes or the upper limit of the number of iterations is reached.

[0087] The adaptive clustering algorithm can adaptively adjust the clustering strategy according to real-time data and dynamically form the optimal node group. The algorithm automatically groups nodes into several groups with similar data distributions based on the data characteristics of the nodes. The node data characteristics within each group are similar, which helps to reduce the impact of non-independent and identically distributed (Non-IID) data on model training.

[0088] After each round of training, clustering and model training are adaptively adjusted according to model performance and node data changes.

[0089] (1) Adjust clustering: regularly recalculate the feature vector of the node and similarity, and re-cluster.

[0090] (2) Optimize the model: adjust the model parameters and training strategies based on the clustering results and model performance.

[0091] Through the above-mentioned adaptive clustering algorithm, node groups with similar data distribution can be effectively formed dynamically according to the characteristics of node data, and a dedicated model branch can be trained for each group, thereby improving the overall performance and generalization ability of the federated learning model.

[0092] After obtaining the feature vector of the node data, the similarity between nodes is calculated using a similarity measurement algorithm, and the nodes are automatically grouped into several clusters based on these similarities. The similarity measurement algorithm is required to accurately reflect the similarity of node data distribution, and the clustering algorithm needs to be able to adaptively adjust the clustering strategy according to real-time data to dynamically form the optimal node group.

[0093] Step 4: Train a model branch independently for the group obtained in step 3 through a group-specific model training method based on adaptive clustering. Each model branch focuses on learning the data distribution characteristics within the group, so as to achieve in-depth learning of different data distributions. The model branch focuses on learning the common characteristics of the node data within the group to improve the model's fit and generalization ability to the data distribution. It is required to be able to select or adjust the most suitable model structure and parameters for different groups to maximize the learning effect of each group model branch.

[0094] During the model training process, clustering strategies and model training strategies are dynamically adjusted based on training feedback and changes in data distribution within the group. It is required to be able to monitor the effect of model training in real time, and adaptively optimize clustering grouping and model parameters based on the monitoring results to ensure that the model can continue to adapt to changes in data distribution, maintain an efficient training process and excellent generalization ability.

[0095] Train a dedicated model branch for each group (cluster). Assume that the rth group contains nodes, the group-specific model is expressed as .

[0096] (1) Initialize the model .

[0097] (2) Each node Train the model on local data And update the model parameters, .

[0098] (3) Aggregate the model of each node to update the group model .

[0099] Step 5: Integrate the learning results of all model branches into the main model through the multi-model branch intelligent fusion method, so that the main model can integrate the knowledge of each model branch, thereby significantly improving the generalization ability of the model on unseen data distributions.

[0100] The model parameter aggregation formula in the multi-model branch intelligent fusion method is as follows: in, Indicates group Exclusive model of Indicates group Local model update, are the weights of the two aggregation methods, represents the number of nodes in the rth group, For Node The amount of data, is the data volume of node R, For Node The similarity and diversity comprehensive score, For Node The comprehensive score of similarity and diversity.

[0101] Example 2

[0102] This embodiment provides a network traffic federated learning generalization system, which is used to implement the above-mentioned network traffic federated learning generalization method, including a data summary collection unit, an intelligent node selection unit, an adaptive clustering unit, a dedicated model training unit, and a multi-model branch intelligent fusion unit, wherein:

[0103] The data summary collection unit is used to collect data summaries of different network nodes in a federated learning environment.

[0104] The data summary collection unit first collects data summary information from all nodes in the federated learning environment. Data summary information includes but is not limited to data volume, data distribution type, data label distribution and other information. It is required to be able to effectively collect necessary data summaries from each node while ensuring that the data collection process does not infringe on the data privacy of the node. In each collection cycle T, each node i needs to complete the calculation and upload of data summary information, which will be uploaded to the central server after each collection cycle T for subsequent intelligent node selection and adaptive clustering algorithms.

[0105] The intelligent node selection unit is used to dynamically select network nodes as participating clients of federated learning and determine the set of network nodes participating in each round of model training.

[0106] The intelligent node selection unit includes a similarity and diversity score calculation module, a node dynamic selection module and a node participation feedback module, wherein:

[0107] The similarity and diversity score calculation module is based on the data summary information collected by the data summary information collection module. The similarity and diversity score calculation module is responsible for calculating the data similarity and diversity indicators of each node. The similarity indicator is used to evaluate the consistency of the node data with the global data distribution. The diversity indicator evaluates the novelty and complementarity that the node data can bring to the model training. This module requires the algorithm to accurately and efficiently score the similarity and diversity of the data of each node.

[0108] The node dynamic selection module is responsible for formulating and implementing the node selection strategy based on the score results calculated by the similarity and diversity score calculation module. This strategy needs to dynamically select a part of all nodes as participants in model training, and the selection criteria are based on a comprehensive consideration of similarity and diversity scores. This module requires that the selection strategy can be flexibly adjusted to adapt to changes in data distribution in the federated learning environment, and ensure that the selected node set can maximize the learning efficiency and generalization ability of the model.

[0109] After a node is selected to participate in model training, the node participation feedback module is responsible for collecting feedback information about the node's participation in training, including but not limited to training efficiency, model performance improvement, and possible changes in data distribution. The feedback mechanism needs to evaluate the effect of node participation and provide a reference for subsequent node selection. This module requires the system to efficiently collect feedback information and adjust the node selection strategy based on this information to continuously optimize the model training process.

[0110] The adaptive clustering unit is used to automatically group the set of network nodes participating in each round of model training into a number of groups with similar data distribution according to the data characteristics of the nodes.

[0111] The adaptive clustering unit automatically divides nodes into several groups with similar data distribution based on node data features, each group consists of nodes with similar data distribution, further improving the adaptability and generalization ability of the model to different data distributions. The adaptive clustering unit includes a data feature extraction module, a similarity measurement and cluster formation module, and an adaptive adjustment and optimization module.

[0112] The data feature extraction module is responsible for extracting feature vectors from the data set of each node. These feature vectors should accurately reflect the distribution characteristics of the node data. It is required to use an efficient feature extraction algorithm in the extraction process to ensure that representative feature vectors can be extracted quickly and accurately from a large amount of data while ensuring that data privacy is not violated. The total amount of data at node i is expressed as , whose eigenvector is expressed as . Eigenvector This can be a statistic extracted from the dataset at each node or a low-dimensional representation obtained through dimensionality reduction methods such as PCA.

[0113] After obtaining the feature vector of the node data, the similarity measurement and cluster formation module uses the similarity measurement algorithm to calculate the similarity between nodes, and automatically groups the nodes into several clusters based on these similarities. This module requires that the similarity measurement algorithm can accurately reflect the similarity of node data distribution, and the clustering algorithm needs to be able to adaptively adjust the clustering strategy according to real-time data to dynamically form the optimal node group.

[0114] During the model training process, the module dynamically adjusts the clustering strategy and model training strategy based on the training feedback and changes in the data distribution within the group. The module is required to be able to monitor the effect of model training in real time, and adaptively optimize the clustering grouping and model parameters based on the monitoring results to ensure that the model can continue to adapt to changes in data distribution, maintain an efficient training process and excellent generalization ability.

[0115] The exclusive model training unit independently trains a model branch for the group through a group-exclusive model training method based on adaptive clustering, and each model branch focuses on learning the data distribution characteristics within the group. For each group formed by adaptive clustering, it is responsible for independently training a model branch. The model branch focuses on learning the common characteristics of the node data within the group to improve the model's fit and generalization ability for the data distribution. The module is required to select or adjust the most suitable model structure and parameters for different groups to maximize the learning effect of each group model branch.

[0116] The multi-model branch intelligent fusion unit is used to integrate the learning results of all model branches into the main model, so that the main model can integrate the knowledge of each model branch.

[0117] The multi-model branch intelligent fusion unit is responsible for integrating the learning results of each group-specific model branch formed by the adaptive clustering algorithm, so that the main model can integrate the knowledge of these model branches, thereby significantly improving the generalization ability of the model on diversified data distribution. It includes a model branch results collection module, a fusion strategy design module, a fusion effect evaluation and optimization module, and a knowledge transfer and sharing module.

[0118] The fusion strategy design module is responsible for collecting the output results and parameters of all group-specific model branches trained by the adaptive clustering algorithm. The collection process is required to be efficient and safe, ensuring that the output of each model branch accurately reflects the data learning results of each group and protecting the intellectual property rights of the model branch.

[0119] After collecting the model branch results, the model branch results collection module is responsible for designing fusion strategies, including but not limited to weighted averaging, model distillation, meta-learning and other methods, to integrate the knowledge of each model branch into the main model. This module requires that the strategy design should consider both the performance of each model branch and its contribution to the global generalization ability to ensure that the integrated main model has the best generalization performance.

[0120] The fusion effect evaluation and optimization module is responsible for evaluating the generalization ability and performance of the fused main model, including but not limited to evaluation on an independent test set. Based on the evaluation results, the module will optimize the fusion strategy, adjust the weights of the model branches, or change the fusion method to further improve the performance of the main model. This module is required to have fast feedback and adjustment capabilities to continuously optimize the generalization ability of the main model.

[0121] During the main model fusion process, the knowledge transfer and sharing module is responsible for realizing knowledge transfer and sharing between model branches to strengthen the main model's understanding and learning of the data characteristics of a specific group. This module is required to be able to identify the knowledge complementarity that may exist between the branches of each model, and through an effective knowledge transfer strategy, promote the main model to comprehensively learn the specific knowledge of each group, thereby improving the overall generalization ability, and then improve the accuracy of data flow business recognition in network communications.

[0122] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A generalized method for federated learning of network traffic, characterized in that: The following steps are involved: Step 1: Collect data summaries of different network nodes in the federated learning environment through a data summary collection method; Step 2: Dynamically select network nodes as participating clients of federated learning through an intelligent node selection method to determine the set of network nodes participating in each round of model training; Step 3: Use an adaptive clustering algorithm to automatically group the set of network nodes participating in each round of model training into several groups with similar data distributions according to the data characteristics of the nodes; Step 4: Train a model branch independently for the group obtained in step 3 using a group-specific model training method based on adaptive clustering, where each model branch focuses on learning the data distribution characteristics within the group; Step 5: Integrate the learning results of all model branches into the main model through the multi-model branch intelligent fusion method, so that the main model can integrate the knowledge of each model branch.

2. The method for generalizing federated learning of network traffic according to claim 1, characterized in that: The model parameter aggregation formula in the multi-model branch intelligent fusion method is as follows: in, Indicates group Exclusive model of Indicates group Local model update, are the weights of the two aggregation methods, represents the number of nodes in the rth group, For Node The amount of data, is the data volume of node R, For Node The similarity and diversity comprehensive score, For Node The comprehensive score of similarity and diversity.

3. The method for generalizing federated learning of network traffic according to claim 2, characterized in that: The method of dynamically selecting network nodes by the intelligent node selection method in step 2 is: obtaining a comprehensive score based on the data summary information of different network nodes, and dynamically selecting a set of network nodes participating in each round of model training based on the comprehensive score.

4. The method for generalizing federated learning of network traffic according to claim 3, characterized in that: The comprehensive score is calculated using the following formula: in, Indicates the comprehensive rating. are the weights of similarity and diversity scores, , , is the similarity weighting coefficient, , is the diversity weighting coefficient, L is the number of labels, K is the number of features, is the mean of all users on feature k, represents the feature mean of node i on feature k, is the standard deviation of all users on feature k, represents the standard deviation of node i on feature k, is the feature distribution of label l for all users, represents the feature distribution of node i in label l, is the entropy of the global model, Represents global data distribution, Indicates the data added to node i The uncertainty of the post-global model, Indicates the amount of data, represents the probability distribution of node i on feature x, represents the probability distribution of global data on feature x, Indicates characteristics.

5. The method for generalizing federated learning of network traffic according to claim 4, characterized in that: The intelligent node selection method in step 2 includes the following steps: Step 21), calculate the comprehensive score for all nodes ; Step 22), follow Sort by high to low and select the first P nodes to participate in the training; Step 23), the node participates in the feedback module; Step 24), after each round of training, collect the node participation and training effect, and update the node selection strategy; update the weight by the following formula: in, represents the similarity score weight, represents the diversity score weight, Indicates Similarity score weights for round training, Indicates The diversity score weights for round training, represents the differential, represents the training loss, is the learning rate.

6. The method for generalizing federated learning of network traffic according to claim 5, characterized in that: In step 3, an adaptive clustering algorithm is used to automatically group the set of network nodes participating in each round of model training into several groups with similar data distribution according to the data characteristics of the nodes: a feature vector is extracted from the data set of each node, the similarity between the nodes is calculated using a similarity measurement algorithm based on the feature vector, and the nodes are grouped into several clusters based on the similarity. The adaptive clustering algorithm can adaptively adjust the clustering strategy according to real-time data and dynamically form the optimal node group.

7. The method for generalizing federated learning of network traffic according to claim 6, characterized in that: The feature vector is a statistic extracted from the data set of each node or a low-dimensional representation obtained by a dimensionality reduction method.

8. The method for generalizing federated learning of network traffic according to claim 7, characterized in that: The data summary includes data volume, data distribution type, and data label distribution.

9. The method for generalizing federated learning of network traffic according to claim 8, characterized in that: The network node is a router or a base station.

10. A generalization system based on the network traffic federated learning generalization method of claim 1, characterized in that: It includes data summary collection unit, intelligent node selection unit, adaptive clustering unit, exclusive model training unit, and multi-model branch intelligent fusion unit, among which: The data summary collection unit is used to collect data summaries of different network nodes in the federated learning environment; The intelligent node selection unit is used to dynamically select network nodes as participating clients of federated learning and determine the set of network nodes participating in each round of model training; The adaptive clustering unit is used to automatically group the set of network nodes participating in each round of model training into a number of groups with similar data distribution according to the data characteristics of the nodes; The exclusive model training unit independently trains a model branch for a group through a group-exclusive model training method based on adaptive clustering, and each model branch focuses on learning the data distribution characteristics within the group; The multi-model branch intelligent fusion unit is used to integrate the learning results of all model branches into the main model, so that the main model can integrate the knowledge of each model branch.

Images