A cost-effective collaborative optimization approach for personalized federated learning in mobile edge networks

Abstract

The present invention belongs to the field of mobile edge computing communication technology, and specifically relates to a cost-effective collaborative optimization method for personalized federated learning in a mobile edge network, including the collection and processing of data sets, collecting trajectory data of a highway traffic data set and digital images of a handwritten digital data set, and adjusting the digital images to the same size, node evaluation, evaluating the quality of all nodes by a good-bad solution distance method, node selection, under the condition that all node probabilities are determined, randomly selecting mobile nodes participating in training according to probability, model training, respectively using MCLR and DNN network training processed training set data to train the model, model transmission, adjusting the trained model parameters, and uploading to an edge server. The present invention can help alleviate the problem of high data heterogeneity while improving the accuracy of model training, and greatly reduce transmission costs and communication delays.

Description

Technical Field

[0001] The present invention belongs to the field of mobile edge computing communication technology, and in particular relates to a cost-effective collaborative optimization method for personalized federated learning in a mobile edge network. Background Art

[0002] In the upcoming 5G and future 6G era, the mobile edge network (MEC) is expected to be responsible for processing a large amount of heterogeneous data generated by mobile user devices such as smart cars, smartphones and wearable devices. The rich behavioral patterns hidden in these data are of vital importance to promote the development of emerging application fields such as AI medical diagnosis and autonomous driving.

[0003] In order to explore the potential value of these huge data sets while ensuring data privacy and promote the development of edge intelligence, Federated Learning (FL) has emerged as an innovative distributed machine learning paradigm. FL allows these devices to collaboratively train a shared model without directly sharing the original data. However, although FL brings significant benefits to mobile edge networks, such as improving privacy protection and allowing computation at the data source to reduce latency, it also encounters many challenges when facing heterogeneous and potentially large-scale MEC.

[0004] First, the characteristics of the datasets generated by mobile devices (MDs) are highly diverse. Not only do the data samples vary in number, but they are often non-independent and identically distributed (non-IID). This poses difficulties for traditional FL algorithms, because these algorithms tend to develop a unified model for all MDs, which may lead to poor model performance on specific MDs. With the rapid development of MD's sensing and computing capabilities, this problem has become more prominent.

[0005] Secondly, communication and computational overhead is also an important issue. The communication and computational costs between mobile devices and edge servers are high, especially when processing large-scale data sets, which will increase training time and energy consumption;

[0006] In addition, the mobility of nodes in mobile edge networks, whether vehicles, drones or smartphones, not only affects their availability, but also has a significant impact on the effectiveness of the learning process. The network topology changes and frequent disconnections caused by the high mobility of nodes pose challenges to the performance of federated learning systems. At the same time, the current research on adaptive federated learning in mobile edge networks (MEC) for highly mobile nodes is still in its initial stage, and most existing node selection methods fail to fully consider the complexity brought by high-speed mobility and transient connections.

[0007] In order to solve the above problems, the introduction of model-agnostic meta-learning (MAML) in the field of edge computing provides a new idea. MAML allows each MD to adapt quickly with only a few data points, which to a certain extent solves the problem of large deviations in sample distribution;

[0008] However, how to combine the advantages of MAML and FL in the MEC environment to propose an adaptive personalized federated learning method to better deal with problems such as data heterogeneity, node mobility, and communication computing overhead is still a hot topic and difficulty in current research;

[0009] Therefore, the present invention proposes an adaptive personalized federated learning method combined with the superior-inferior solution distance method, providing a new solution for federated learning in mobile edge networks. Summary of the invention

[0010] The purpose of the present invention is to provide a cost-effective collaborative optimization method for personalized federated learning in mobile edge networks, which can help alleviate the problem of high data heterogeneity while improving model training accuracy and greatly reducing transmission costs and communication delays. The technical solutions adopted by the present invention are as follows:

[0011] The cost-effective collaborative optimization method for personalized federated learning in mobile edge networks includes the following steps:

[0012] A: Dataset collection and processing: collect trajectory data of the highway traffic dataset and digital images of the handwritten digital dataset, and adjust the digital images to the same size;

[0013] B: Node evaluation, using the good-bad solution distance method to evaluate the quality of all nodes;

[0014] C: Node selection: Under the condition that all node probabilities are determined, the mobile nodes participating in the training are randomly selected according to probability;

[0015] D: Model training, the model is trained using the training set data processed by MCLR and DNN network training respectively;

[0016] E: Model transmission, adjusting the trained model parameters and uploading them to the edge server;

[0017] F: Probability adjustment, setting an evaluation result threshold so that the edge server can determine whether the mobile node selected according to the probability is below the threshold;

[0018] G: Model update. The edge server sends the aggregated new parameters and probabilities back to the corresponding mobile nodes. The mobile nodes train again with their own data and the returned parameters until they reach their respective number of iterations and output the model.

[0019] The technical effects achieved by the present invention are:

[0020] 1. In each round of FL, nodes will be selected to participate in local training and global model aggregation. For those nodes with poor stability due to mobility, it is reasonable to naturally adjust their participation probability. Therefore, it is proposed on the server side to dynamically adjust the selection probability of each node using the output of the scoring function, which is conducive to alleviating the problem of high data heterogeneity and improving the accuracy of model training.

[0021] 2. Compared with the existing transmission model method, the present invention combines the adaptive node selection method and the personalized federated learning method based on MAML to target the high mobility and transient state of nodes. The results show that the present invention greatly reduces the transmission cost and communication delay. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of the federated learning based on the superior-inferior solution distance method of the present invention;

[0023] Figure 2 It is a framework of federated learning in mobile edge networks of the present invention;

[0024] Figure 3 It is the framework of federated learning based on the superior-inferior solution distance method of the present invention;

[0025] Figure 4 is the test accuracy of the method proposed in this invention and the baseline method on the NGSIM dataset;

[0026] Figure 5 It is the test accuracy of the method proposed in this invention and the baseline method on the MNIST dataset. DETAILED DESCRIPTION

[0027] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0028] like Figure 1-5 As shown in FIG. 1 , a cost-effective collaborative optimization method for personalized federated learning in a mobile edge network includes the following steps:

[0029] A: Dataset collection and processing: collect trajectory data of highway traffic dataset and digital images of handwritten digit dataset. The main collection is the American highway traffic dataset (NGSIM) and the handwritten digit dataset (MNIST). The highway traffic dataset contains 25 columns and more than 11.8 million rows of trajectory data for four data collection locations.

[0030] The handwritten digit dataset contains 70,000 handwritten digit images, including 60,000 training images and 10,000 test images. The images in the dataset are of different sizes, but in order to train the model, we need to resize them to the same size.

[0031] B: Node evaluation, which uses the good-bad solution distance method to evaluate the quality of all nodes and adjusts the probability of node selection based on the evaluation results;

[0032] C: Node selection: Under the condition that all node probabilities are determined, the mobile nodes participating in the training are randomly selected according to probability;

[0033] D: Model training, the model is trained using the training set data processed by MCLR and DNN network training respectively;

[0034] E: Model transmission, adjusting the trained model parameters and uploading them to the edge server;

[0035] F: Probability adjustment: Set a threshold for the evaluation result. The edge server determines whether the mobile node selected based on the probability is below the threshold. The probability of such a node being selected is reduced according to certain rules, and the reduced probability is evenly distributed to other nodes.

[0036] G: Model update: The edge server sends the aggregated new parameters and probabilities back to the corresponding mobile nodes, and the nodes use their own data and the returned parameters for the next round of training until the number of iterations is reached and the model is output.

[0037] It first uses a federated learning framework to solve the problem of malicious interception during data transmission, and then introduces a node selection method based on the superior-inferior solution distance method on the basis of federated learning to solve the problem of network topology changes caused by the high mobility of nodes in actual scenarios and the high communication cost caused by frequent disconnection. In addition, by combining personalized federated learning, the reliability and efficiency of federated learning in mobile edge networks are effectively improved;

[0038] In each round of FL, nodes will be selected to participate in local training and global model aggregation. For those nodes with poor stability due to mobility, it is reasonable to naturally adjust their participation probability. Therefore, it is proposed on the server side to use the output of the scoring function to dynamically adjust the selection probability of each node, which is conducive to alleviating the problem of high data heterogeneity and improving the accuracy of model training.

[0039] Compared with the existing transmission model method, the present invention combines the adaptive node selection method and the MAML-based personalized federated learning method for the high mobility and transient state of nodes. The results show that the present invention greatly reduces the transmission cost and communication delay.

[0040] The present invention discloses an adaptive personalized federated learning framework based on the superiority-inferiority distance method, including the collection of node (such as vehicle) mobile data, the processing and conversion of the data; initialization setting of the transmitted model parameters; inputting the pre-processed node data into the training network, and selecting different hyperparameters to train the model; in particular, before the model training, each mobile node is evaluated by the superiority-inferiority distance method;

[0041] The probability of a mobile node being selected is adjusted according to the evaluation results; the server selects the mobile nodes participating in the training by the selection probability, and after the training is completed, the probability of the mobile nodes with lower evaluation results being selected is reduced, while the probability of the remaining nodes being selected is increased, and the training is cyclical.

[0042] This approach reduces the complexity and performance degradation caused by high-speed node movement and transient connections, reduces the training time cost in communication, and eliminates security issues in data transmission.

[0043] In order to more clearly explain the present technical solution, the present technical solution discloses the following solution or embodiment, which demonstrates the process of the cost-effective collaborative optimization method of personalized federated learning in a mobile edge network, as follows:

[0044] The steps of B are:

[0045] S1: Based on the evaluation of the superior-inferior solution distance method, data on the moving speed, direction and connection frequency of each node are collected, and these data are preprocessed to ensure accuracy and consistency. The preprocessing steps include data cleaning, missing value processing and outlier detection. At the same time, in order to eliminate the dimensional influence between different standards, the data is converted into dimensionless form. Therefore, different standardization methods are used for the data related to the three factors;

[0046] For the moving speed, because the speed usually has an upper limit, and for the moving direction, it needs to be converted into a numerical value, such as the frequency or consistency of the direction change, so the interval index is used to eliminate the dimensional influence between the speed standard and the direction standard. For the interval index, if its optimal interval is [a, b], take:

[0047] M=max{a-min{x i},max{x i}-b} (1)

[0048] Among them, min(x) and max(x) are the minimum and maximum values ​​in the data set, respectively, and are calculated according to the following formula

[0049] To convert:

[0050]

[0051] For connection frequency, considering that the data set has large differences or outliers in connection frequency, Z score normalization is used:

[0052]

[0053] Through the above steps, the normalized data is integrated into a decision matrix Z, the weight of the node attributes is considered, and the weight vector W is introduced. By multiplying the normalized decision matrix with the weight vector, the weighted normalization matrix Y is obtained;

[0054] Suppose there are m options and n criteria, then the matrix is ​​represented as:

[0055]

[0056] The matrix Y to be scored, which has been processed by forward and normalization, contains extremely large data, from which the ideal optimal solution and the worst solution are extracted;

[0057] Therefore, we take out each index, that is, the largest number in each column, to form the ideal optimal solution vector, and then define the positive ideal solution A + and negative ideal solution A - , which represent the best and worst node attributes respectively, are:

[0058] A + =(max i y i1 ,max i y i2 ,...,max i y in ) (5)

[0059] A - =(min i y i1 ,min i y i2 ,...,min i y in ) (6)

[0060] By calculating the Euclidean distance, the distance from the nth node to the positive ideal solution and the negative ideal solution is obtained in the following equations (7) and (8):

[0061] Distance to the ideal solution:

[0062]

[0063] The comprehensive distance index B(n) is introduced to measure the comprehensive distance of each node relative to the positive ideal solution and the negative ideal solution. The smaller the index is, the closer the node is to the positive ideal solution and the higher the score is, which is:

[0064]

[0065] Step F is:

[0066] S2: Probability adjustment is performed based on the node evaluation results. In each round of federated learning (FL), some nodes are selected to participate in local training and global model aggregation. Nodes with poor stability due to mobility naturally adjust their participation probability. Therefore, it is proposed to use the output of the scoring function on the server side to dynamically adjust the selection probability of each node.

[0067] In particular, the probability of nodes identified as unstable by the above S1 and selected by the control in C will be reduced based on the parameter x in the following formula (10), while the selection probability of other nodes will be increased accordingly, specifically:

[0068]

[0069] in, and They represent the probability of a node being selected in the kth global round and the reduction in probability in the next round respectively. The min function returns the minimum value of all parameters. The parameter x represents the ratio of the cumulative number of times marked by a process to the number of times the node is selected, and the value range is x∈(0,1];

[0070] β∈[0,1] as coefficient is as follows:

[0071]

[0072] α controls a given ratio x (x+β) α The size of the probability reduction achieved. For example, a larger value of α will lead to a more significant probability reduction. This is because when α increases in a smaller range (υ,ò), the probability reduction occurs in a wider range (β,1), causing the node selection probability to drop more rapidly.

[0073] At the same time, a larger α makes node selection very sensitive to identification errors, which may result in the failure to select independent and identically distributed (iid) nodes in subsequent rounds. However, a smaller α value cannot always effectively distinguish nodes, because the change in probability may be only marginal;

[0074] In order to ensure that the probability change remains within a controllable range, it is recommended to use a value of β. In addition, the specific values ​​of parameters α and β will be further adjusted through experiments.

[0075] After getting the marked node, that is After that, the probability of these nodes is reduced accordingly according to the following formula (13):

[0076]

[0077] in, n∈N is used for the (k+1)th round.

[0078] Exclude the marked nodes from the total nodes, that is, let

[0079] In the following formula (14), the probability of reducing the marked nodes is evenly distributed to the unmarked nodes, which is:

[0080]

[0081] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.

Claims

1. A cost-effective collaborative optimization method for personalized federated learning in mobile edge networks, characterized by: The following steps are involved: A: Dataset collection and processing: collect trajectory data of the highway traffic dataset and digital images of the handwritten digital dataset, and adjust the digital images to the same size; B: Node evaluation, using the good-bad solution distance method to evaluate the quality of all nodes; The steps of B are: S1: Based on the evaluation of the superior-inferior solution distance method, data on the movement speed, direction, and connection frequency of each node are collected, preprocessed to ensure accuracy and consistency, and converted into dimensionless form; In S1, an interval-type index is used to eliminate the dimensional influence between the speed standard and the direction standard. For the interval-type index, if its optimal interval is [a, b], take: M=max{a-min{x i },max{x i }-b} (1) Among them, min(x) and max(x) are the minimum and maximum values ​​in the data set, respectively, and are converted according to the following formula: C: Node selection: Under the condition that all node probabilities are determined, the mobile nodes participating in the training are randomly selected according to probability; D: Model training, the model is trained using the training set data processed by MCLR and DNN network training respectively; E: Model transmission, adjusting the trained model parameters and uploading them to the edge server; F: Probability adjustment, setting an evaluation result threshold so that the edge server can determine whether the mobile node selected according to the probability is below the threshold; The F step is: S2: Probability adjustment is performed based on the node evaluation results. In each round of federated learning, some nodes are selected to participate in local training and global model aggregation. The participation probability of nodes with poor stability due to mobility is adjusted. Therefore, it is proposed to use the output of the scoring function on the server side to dynamically adjust the selection probability of each node. G: Model update. The edge server sends the aggregated new parameters and probabilities back to the corresponding mobile nodes. The mobile nodes train again with their own data and the returned parameters until they reach their respective number of iterations and output the model.

2. The cost-effective collaborative optimization method for personalized federated learning in a mobile edge network according to claim 1, characterized in that: In step B, after evaluating the quality of all nodes, the probability of a node being selected is adjusted according to the evaluation result.

3. The cost-effective collaborative optimization method for personalized federated learning in mobile edge networks according to claim 1, characterized in that: In the step F, the probability of the mobile node being selected is reduced, and the reduced probability is evenly distributed to other nodes.

4. The cost-effective collaborative optimization method for personalized federated learning in a mobile edge network according to claim 1, characterized in that: The dataset has differences or outliers in connection frequency, and is normalized using Z-score:

5. The cost-effective collaborative optimization method for personalized federated learning in a mobile edge network according to claim 4, characterized in that: Through the above steps, the normalized data is integrated into a decision matrix Z, the weight of the node attributes is considered, and the weight vector W is introduced. By multiplying the normalized decision matrix with the weight vector, the weighted normalization matrix Y is obtained; Suppose there are m options and n criteria, then the matrix is ​​represented as: The matrix Y to be scored, which has been processed by forward and normalization, contains extremely large data, from which the ideal optimal solution and the worst solution are extracted; Therefore, we take out each index, that is, the largest number in each column, to form the ideal optimal solution vector, and then define the positive ideal solution A + and negative ideal solution A - , which represent the best and worst node attributes respectively, are: TO + (max) i and i1 ,max i and i2 ,...,max i and in ) (5) HAS - =(min i y i1 ,min i y i2 ,...,min i y in ) (6) By calculating the Euclidean distance, the distance from the nth node to the positive ideal solution and the negative ideal solution is obtained in the following equations (7) and (8): The comprehensive distance index B(n) is introduced to measure the comprehensive distance of each node relative to the positive ideal solution and the negative ideal solution, which is:

6. The cost-effective collaborative optimization method for personalized federated learning in a mobile edge network according to claim 1, characterized in that: In S2, the probability of the node identified as unstable by S1 and selected by C is reduced based on the parameter x in the following formula (10), while the selection probability of other nodes is increased accordingly, specifically: in, and They represent the probability of a node being selected in the kth global round and the reduction in probability in the next round respectively. The min function returns the minimum value of all parameters. The parameter x represents the ratio of the cumulative number of times marked by a process to the number of times the node is selected, and the value range is x∈(0,1]; β∈[0,1] is the coefficient as follows: α controls a given ratio x (x+β) α The size of the probability reduction achieved is when the marked node is obtained, that is, After that, the probability of these nodes is reduced accordingly according to the following formula (13): in, Used for round k+1.

7. The cost-effective collaborative optimization method for personalized federated learning in a mobile edge network according to claim 6, characterized in that: Exclude the marked nodes from the total nodes, that is, let In the following formula (14), the probability of reducing the marked nodes is evenly distributed to the unmarked nodes, which is:

Images