A federated learning service deployment method and system based on end node residual energy awareness, a device, and a medium
By constructing a federated learning service deployment method that is aware of the remaining energy of the end nodes, the CPU frequency and transmit power of the end nodes are optimized, which solves the problem of excessive energy consumption in the existing technology and realizes device energy protection and training efficiency improvement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-21
- Publication Date
- 2026-04-10
AI Technical Summary
In existing federated learning systems, the impact of the remaining energy of end nodes on device participation in decision-making is not fully considered, resulting in excessive energy consumption and device energy depletion, which affects training efficiency and stability.
By constructing a federated learning service deployment method that is aware of the remaining energy of end nodes, and by utilizing cloud game theory and edge-end matching algorithms, the CPU frequency and transmit power of end nodes are optimized to reduce energy consumption and protect device energy.
It effectively reduces the computing and communication energy consumption of end nodes, protects equipment energy, improves training efficiency and stability, and saves costs.
Smart Images
Figure CN116455907B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of federated learning, and particularly relates to a federated learning service deployment method and system based on end node residual energy perception, a device, and a medium. BACKGROUND
[0002] Federated learning can effectively solve the problem of privacy leakage by constructing a usable machine learning model through joint multi-source data without leaving the local data. In the common architecture of federated learning, the cloud-edge-end three-layer architecture has the advantages of high training efficiency and good model performance, and has been widely studied. The deployment of federated learning service in the cloud-edge-end three-layer architecture includes two stages, end node selection and edge-end association. In the end node selection stage, the end node decides whether to participate in the federated learning training independently, and the model training will consume the node computing resources and reduce the node battery capacity, so it is necessary to design an incentive mechanism to encourage the end node to participate in the model training. In the edge-end association stage, the network resources owned by the edge node are limited, and too many end nodes accessing the same edge node will cause network congestion, reducing the edge-end communication efficiency and local aggregation efficiency. In addition, when the residual energy is too low, the end node will be unable to perform model training and upload the federated learning model, thereby causing the federated learning training to stagnate, so it is necessary to introduce an end node residual energy protection mechanism in the two stages of federated learning deployment, so that the end node changes the used computing resources according to its residual energy, and realizes the energy protection of the end node.
[0003] Yunus Sarikaya and Ozgur Ercetin proposed an end node incentive method based on Stackelberg game in their paper "Motivating Workers in Federated Learning: A Stackelberg Game Perspective". In this method, the cloud is the two parties of the game, and the cloud node pays the end node a reward. The cloud node income function is designed as the federated learning training time, and the end node income function is designed as the income of participating in the training, and the end node expenditure is its energy consumption. The Lagrange multiplier method is used to calculate the optimal end node resource and end node reward, and finally determine the end nodes participating in the federated learning and the resource allocation. The disadvantage of this method is that it designs how to reduce the system energy consumption, but ignores the influence of the terminal device residual energy on the device participating in the federated learning training decision, and lacks protection of the end node energy. SUMMARY
[0004] In order to overcome the above-mentioned deficiencies of the prior art, the purpose of the present application is to propose an end node residual energy-aware federated learning service deployment method, system, device and medium, which can reduce the energy consumption of the end node by selecting a lower CPU frequency and transmission power for the end node with less residual energy, realize energy protection of the end node, and has the characteristics of cost saving and simple operation.
[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:
[0006] The end node residual energy-aware federated learning service deployment method has the following specific steps:
[0007] Step S1, according to the federated learning process, a corresponding federated learning time model and an end node energy consumption model are constructed for each end node;
[0008] Step S2, according to the federated learning training time model and the end node computing energy consumption model obtained in step S1, cloud-side game is performed;
[0009] Step S3, according to the federated learning communication time model and the end node communication energy consumption model obtained in step S1, the end nodes participating in federated training obtained in step S2 are matched with the edge nodes.
[0010] The construction of the federated learning time model and the end node energy consumption model in step S1 has the following specific operations:
[0011] The federated learning training time model is constructed, and the training time of the end node i is denoted as The number of CPU cycles used by the end node i for model training is denoted as c i The CPU frequency used by the end node i is denoted as f i , wherein can be represented as:
[0012]
[0013] The federated learning communication time model is constructed, and the communication time between the end node i and the edge node j is denoted as The channel gain between the end node i and the edge node j is denoted as h ij , the model parameter size is denoted as w, the communication bandwidth is denoted as B, and the transmission power of the end node i is denoted as p i , wherein can be represented as:
[0014]
[0015] The end node computing energy consumption model is constructed, and represents the computing energy consumption of the end node i, and k represents the effective capacitance coefficient of the end node i, wherein is expressed as:
[0016]
[0017] The end node communication energy consumption model is constructed, and the represents the communication energy consumption of the end node i accessing the edge node j, p i represents the transmission power of the end node i, then is expressed as:
[0018]
[0019] The step S2 specifically operates as follows:
[0020] Step S2.1, according to the federated learning time model, a cloud node delay minimum benefit function is constructed;
[0021] Step S2.2, according to the end node energy consumption model and the end node residual energy, an end node benefit function is constructed;
[0022] Step S2.3, the cloud node and the end node benefit function are solved, and the end node participating in training and the computing resource allocation scheme are obtained.
[0023] The end node benefit function in the step S2.2 is specifically:
[0024] U i (q i ,f i ) represents the reward obtained by the end node i for model training, q i represents the CPU frequency unit price paid by the cloud node to the end node i, and cost i represents the energy consumption overhead factor of the end node i, E i,res represents the residual energy of the end node i after completing the current federated learning training cycle, E i,max represents the maximum available energy of the end node i, and α represents the node residual energy factor, which indicates the influence degree of the end node residual energy on the decision of the end node, and U i (q i ,f i ) is expressed as:
[0025]
[0026] The step S2.3 solves the cloud node and the end node benefit function, and obtains the end node participating in training and the computing resource allocation scheme, which is specifically:
[0027] Step S2.31, initialize the population individuals, each individual corresponds to the CPU frequency unit price scheme paid by the cloud node to all end nodes, and the end node calculates the CPU frequency that maximizes its benefit function under the current reward;
[0028] Step S2.32, using the binary tournament algorithm, select individuals with short training time as the next generation population;
[0029] Step S2.33, two-point crossover operation is performed on the population individuals;
[0030] Step S2.34, mutation operation is performed on the individuals, and a new end node CPU frequency scheme under the reward scheme is generated;
[0031] Step S2.35, repeat steps S2.32 to S2.34 to a maximum number of iterations, finally determine the CPU frequency of the end node and the end node reward scheme, and the CPU frequency of the end node when not participating in training is 0.
[0032] The specific operation of the step S3 is as follows:
[0033] Step S3.1, according to the federated learning communication time model and the end node communication energy consumption model, and combining the end node residual energy, construct the end node preference function, select the distance as the edge node preference function;
[0034] Step S3.2, solve the edge-end preference function, realize the edge-end association and determine the end node transmission power.
[0035] The end node preference function in the step S3.1 is specifically:
[0036] Let the federated learning communication benefit obtained by the end node i be U i , use β to represent the communication reward factor of the end node i, use θ t and θ e to represent the communication time factor and the residual energy factor respectively, use E i,res to represent the residual energy of the end node i after completing the current federated learning communication, use co i to represent the end node energy consumption overhead factor, and express U i as:
[0037]
[0038] Wherein, is defined as:
[0039]
[0040] Neglecting the aggregation time of the edge node, the transmission delay of the edge node j and the cloud node is represented as T j , and is defined as:
[0041]
[0042] A federated learning service deployment system based on end node residual energy perception comprises:
[0043] A federated learning module is configured to construct a corresponding federated learning time model and an end node energy consumption model for each end node.
[0044] A cloud game module is configured to perform cloud game on the federated learning training time model and the end node computing energy consumption model obtained in step S1.
[0045] An edge-end matching module is configured to perform edge-end matching on the federated learning communication time model and the end node communication energy consumption model obtained in step S1 and the end nodes participating in federated training obtained in step S2.
[0046] A federated learning service deployment device based on end node residual energy perception comprises:
[0047] A memory is configured to store required data.
[0048] A processor is configured to implement the federated learning service deployment method based on end node residual energy perception described in steps S1 to S3 when executing the computer program.
[0049] A computer readable storage medium stores a computer program, and the computer program can implement the federated learning service deployment method based on end node residual energy perception when executed by a processor.
[0050] Compared with the prior art, the present application has the following advantages:
[0051] 1. The present application fully considers the influence of the residual energy of the end node, designs a cloud game algorithm in the end node selection stage, determines the end nodes participating in distributed training and the allocation amount of computing resources thereof, and introduces the residual energy of the end node into a game revenue function, so that the end node with less residual energy selects a lower CPU frequency, thereby reducing the computing energy consumption of the end node and realizing energy protection of the end node.
[0052] 2. In the edge-end matching stage, the present application designs an edge-end matching algorithm, determines the edge nodes accessed by the end nodes and the transmission power allocation of the end nodes, and introduces the residual energy of the end nodes into an end node preference function, so that the end node with less residual energy selects a lower transmission power, thereby reducing the communication energy consumption of the end node and realizing energy protection of the end node.
[0053] 3. Step S2.3 of the present application uses a genetic algorithm to solve the cloud node revenue function, forms a candidate solution of the end node computing reward, and the end node solves the revenue function according to the candidate computing reward, thereby successfully solving the complex cloud game problem constructed by the present application.
[0054] In summary, the present application has the characteristics of cost saving and simple operation. Attached Figure Description
[0055] Figure 1 This is a flowchart illustrating the implementation of the present invention.
[0056] Figure 2 This is a flowchart illustrating the cloud-based game playing implementation of the present invention.
[0057] Figure 3 This is a flowchart illustrating the delayed reception algorithm implementation in edge matching of the present invention.
[0058] Figure 4 This is a comparison chart of node energy standard deviation and training time between the present invention and a federated learning service deployment method with no residual energy awareness; wherein... Figure 4 (a) is a comparison chart of the standard deviation of the residual energy at the end nodes. Figure 4 (b) is a comparison chart of federal learning and training time.
[0059] Figure 5 This is a comparison chart of the node remaining energy of the present invention and the federated learning service deployment method without remaining energy awareness. Detailed Implementation
[0060] The present invention will now be described in detail with reference to the accompanying drawings.
[0061] Reference Figure 1 The deployment method of federated learning services based on end-node remaining energy awareness includes the following steps:
[0062] Step S1: Based on the federated learning process, construct a corresponding federated learning time model and end-node energy consumption model for each end node. The specific operations are as follows:
[0063] The federated learning training time model is constructed as follows:
[0064]
[0065] in, c represents the local learning time of end node i. i f represents the number of CPU cycles used by end node i for local model training. i This represents the available computing resources of end node i.
[0066] The federated learning communication time model is constructed as follows:
[0067]
[0068] in, This represents the communication time from when endpoint i uploads its local model to edge server j, where w represents the model parameter size, B is the bandwidth between the endpoint and the edge, and p... i h represents the transmit power of end node i.ij denotes the channel gain between the end node i and the edge server j, and n0denotes the noise power.
[0069] The end node computing energy consumption model is constructed as:
[0070]
[0071] wherein, denotes the computing energy consumption of the end node i, and k denotes the effective capacitance coefficient of the end node i.
[0072] The end node communication energy consumption model is constructed as:
[0073]
[0074] wherein, denotes the communication energy consumption of the end node i, and p i denotes the transmission power of the end node i, denotes the communication delay between the end node i and the edge server j.
[0075] Step S2, according to the federated learning training time model and the end node computing energy consumption model obtained in step S1, cloud-side game is performed, and the specific operation is as follows:
[0076] Step S2.1, according to the federated learning time model, a cloud node delay minimum benefit function is constructed;
[0077] Step S2.2, according to the end node energy consumption model and the end node residual energy, an end node benefit function is constructed, and specifically:
[0078] The cloud node benefit function is constructed, and is expressed as:
[0079]
[0080] wherein, denotes the average value of the federated learning training time, denotes the end node training rate, denotes the set of end nodes that perform the federated learning training last among all K end nodes.
[0081] The end node benefit function is constructed, and is expressed as:
[0082]
[0083] wherein, U i (q i ,f i ) denotes the reward obtained by the end node i for model training, q i denotes the CPU frequency unit price paid by the cloud node to the end node i, and cost idenotes the energy consumption overhead factor of the end node i i,res denotes the remaining energy of the end node i after completing the current federated learning training cycle i,max denotes the maximum available energy of the end node i, denotes the node remaining energy factor, the size of which indicates the degree of influence of the end node remaining energy on the decision of the end node; the first part denotes the computing reward obtained by the end node i, the greater the CPU frequency used, the more the reward obtained, and the second part denotes the energy consumption overhead of the end node, and the smaller the remaining energy, the greater the energy consumption overhead.
[0084] The cloud node computing constraint is constructed and is represented as:
[0085]
[0086] wherein budgetComp denotes the total budget of the cloud node federated learning training time.
[0087] The end node CPU frequency constraint is constructed and is represented as:
[0088] f min ≤f i ≤f max
[0089] wherein f min and f max respectively denote the minimum available CPU frequency and the maximum available CPU frequency of the end node.
[0090] Referring to Figure 2 , step S2.3 solves the cloud node and end node revenue functions to obtain the end node participating in training and the computing resource allocation scheme, specifically:
[0091] Step S2.31, initialize the population individuals, each individual corresponds to a CPU frequency unit price scheme of the cloud node to all end nodes, and the end node calculates the CPU frequency of the end node under the current reward to maximize the revenue function of the end node;
[0092] Step S2.32, use the binary tournament algorithm to select individuals with short training time as the next generation population;
[0093] Step S2.33, perform two-point crossover operation on the population individuals;
[0094] Step S2.34, perform mutation operation on the individuals and generate the end node CPU frequency scheme under the new reward scheme;
[0095] Step S2.35, repeat steps S2.32 to S2.34 to a maximum number of iterations, and finally determine the CPU frequency of the end node and the end node reward scheme, and the CPU frequency of the end node is 0 when the end node does not participate in training.
[0096] Step S3, according to the federated learning communication time model and the end node communication energy consumption model obtained in step S1, the end nodes participating in federated training obtained in step S2 are edge-end matched, and the specific operation is as follows:
[0097] Step S3.1, according to the federated learning communication time model and the end node communication energy consumption model, and combining the residual energy of the end node, an end node preference function is constructed, and the distance is selected as the edge node preference function, specifically:
[0098]
[0099] Wherein, U i represents the federated learning communication income of the end node i, β is the communication reward factor of the end node i, θ t represents the communication time factor, θ e represents the residual energy factor, E i,res represents the residual energy of the end node i after completing the current federated learning communication, co i represents the end node energy consumption overhead factor; the first part represents the communication reward obtained by the end node i, and the smaller the communication delay, the more the reward obtained, and the second part represents the energy consumption overhead, and the smaller the residual energy, the greater the energy consumption overhead. In the first part of the formula, the is defined as:
[0100]
[0101] Neglecting the aggregation time of the edge node, the transmission delay between the edge node j and the cloud node is represented as T j , and is defined as:
[0102]
[0103] The communication budget constraint and the end node transmission power constraint are constructed, and are represented as:
[0104]
[0105] p min ≤p i ≤p max
[0106] Wherein, budgetComm represents the total budget of federated learning communication, p min and p max represent the minimum available transmission power and the maximum available transmission power of the end node, respectively.
[0107] Step S3.2, the edge-end preference function is solved, the edge-end association is realized, and the end node transmission power is determined.
[0108] The joint greedy algorithm and the delay acceptance algorithm determine the communication reward factor, and the edge-end association scheme and the end node transmission power. The specific steps are described as follows: determining the value range of the communication reward factor β according to the communication budget; solving the corresponding edge-end matching scheme and the end node transmission power under different values of β, and selecting the optimal scheme as the final β value, edge-end association scheme and end node transmission power result.
[0109] Referring to Figure 3 , the delay acceptance algorithm has the following specific process:
[0110] 1. A preference list of end nodes for each edge node is established according to the edge-end distance, and each edge node can access n end nodes;
[0111] 2. The edge node initiates an association invitation to the first n end nodes in the preference list, and deletes these end nodes from the preference list;
[0112] 3. The end node calculates the maximum benefit and the corresponding transmission power according to the preference function, and decides whether to associate with the edge node; when the end node has not been associated with the edge node, the invitation of the edge node is accepted; when the end node has been associated with other edge nodes, the edge node with a large preference function value is retained;
[0113] 4. Steps 2 to 3 are repeated until all end nodes are connected to an edge node, or all edge nodes' preference lists are empty, and finally the edge-end association scheme and the transmission power used by the end node are obtained.
[0114] The federated learning service deployment system based on end node residual energy perception comprises:
[0115] A federated learning module is configured to construct a corresponding federated learning time model and an end node energy consumption model for each end node.
[0116] A cloud game module is configured to perform cloud game on the federated learning training time model and the end node computing energy consumption model obtained in step S1.
[0117] An edge-end matching module is configured to perform edge-end matching on the end nodes participating in federated training obtained in step S2, based on the federated learning communication time model and the end node communication energy consumption model obtained in step S1.
[0118] The federated learning service deployment device based on end node residual energy perception comprises:
[0119] A memory is configured to store required data.
[0120] A processor is configured to implement the federated learning service deployment method based on end node residual energy perception described in steps S1 to S3 when executing the computer program.
[0121] A computer readable storage medium stores a computer program, which, when executed by a processor, can implement a federated learning service deployment method based on end node residual energy perception.
[0122] The effect of the application is further described below in combination with simulation experiments:
[0123] 1. Simulation parameter setting:
[0124] The following network scenario is set in the application: within a range of 500*500m, 5 edge nodes and 20 end nodes are randomly set, and the edge-end channel gain is: h(dB)=-128.1-37.6log 10 (d), wherein d(km) represents the transmission distance between the edge server and the end nodes covered by the base station where the edge server is located. The related simulation parameter settings are shown in Table 1:
[0125] Table 1 Simulation parameter setting
[0126]
[0127]
[0128] 2. Simulation content and result analysis:
[0129] Under the above environment and parameters, the federated learning service deployment algorithm of the application and the existing federated learning service deployment algorithm which does not consider the residual energy of the end nodes are used to deploy the federated learning service, and the federated learning training is realized through numerical simulation, and the results are referred to Figure 4 and Figure 5 .
[0130] Figure 4 The standard deviation of the residual energy of the end nodes and the federated learning training time are compared, and under the application, the standard deviation of the residual energy of the end nodes is lower than that of the comparative method, and the training time is higher than that of the comparative method; Figure 5 The residual energy of the end nodes of the application and the comparative application is compared, and the residual energy greater than 0 is taken as a reference, and it can be seen from the figure that in the application, the end nodes with exhausted energy appear in the 80th training cycle of federated training, and in the comparative algorithm, half of the end nodes have energy of 0 in the first 20 cycles of federated training.
[0131] The simulation results show that the federated learning service deployment method based on end node residual energy perception can make the end nodes with less residual energy use lower CPU frequency and transmit power for training, reduce the energy consumption of the end nodes, and realize the energy protection of the end nodes.
Claims
1. A federated learning service deployment method based on end node residual energy awareness, characterized in that, The specific steps are as follows: Step S1, according to the federated learning process, a corresponding federated learning time model and an end node energy consumption model are constructed for each end node; the specific operation of the federated learning time model and the end node energy consumption model in step S1 is as follows: A federated learning training time model is constructed, and the end node i training time is denoted as The number of CPU cycles used by the end node i for model training is denoted as The CPU frequency used by the end node i is denoted as Wherein Can be expressed as: ； A federated learning communication time model is constructed, and the communication time between the end node i and the edge node j is denoted as The channel gain between the end node i and the edge node j is denoted as h ij The model parameter size is denoted as The communication bandwidth is denoted as The transmission power of the end node i is denoted as Wherein Can be expressed as: ， Wherein, n0 represents the noise power; A computing energy consumption model of end nodes is constructed, using denotes the computing energy consumption of end node i, using denotes the effective capacitance coefficient of end node i, wherein is denoted as: ； A communication energy consumption model of end nodes is constructed, and the energy consumption of end nodes is calculated using represents the communication energy consumption of end node i accessing edge node j, represents the transmission power of end node i, and is represented as: ； Step S2, according to the federated learning training time model and the end node computing energy consumption model obtained in step S1, cloud game is carried out; the specific operation of step S2 is as follows: Step S2.1, according to the federated learning time model, a cloud node delay minimum benefit function is constructed; Step S2.2, according to the end node energy consumption model and the end node residual energy, an end node benefit function is constructed; the end node benefit function in step S2.2 is specifically: Using denotes the reward of end node i for model training, using denotes the CPU frequency unit price paid by the cloud node to end node i, using denotes the energy consumption overhead factor of end node i, denotes the remaining energy of end node i after completing the current federated learning training cycle, denotes the maximum available energy of end node i, using denotes the node remaining energy factor, which indicates the degree of influence of the remaining energy of the end node on the decision-making of the end node, and is denoted as: ； Step S2.3, the cloud node and end node benefit functions are solved to obtain the end node participating in training and its computing resource allocation scheme; the cloud node and end node benefit functions are solved in step S2.3 to obtain the end node participating in training and its computing resource allocation scheme, which is specifically: Step S2.31, initialize the population individuals, each individual corresponds to the CPU frequency unit price scheme of the cloud node to all end nodes, and the CPU frequency of the end node that makes the maximum benefit function under the current remuneration is obtained; Step S2.32, use binary tournament algorithm to select individuals with short training time as the next generation population; Step S2.33, two-point crossover operation is performed on the population individuals; Step S2.34, mutation operation is performed on the individuals, and the CPU frequency scheme of the end node under the new remuneration scheme is generated; Step S2.35, repeat steps S2.32 to S2.34 to a maximum number of iterations to finally determine the CPU frequency of the end node and the end node remuneration scheme, and the CPU frequency of the end node when it does not participate in training is 0; Step S3, according to the federated learning communication time model and the end node communication energy consumption model obtained in step S1, the end nodes participating in federated training obtained in step S2 are matched; the specific operation of step S3 is as follows: Step S3.1, according to the federated learning communication time model and the end node communication energy consumption model, and combining the end node residual energy, an end node preference function is constructed, and distance is selected as the edge node preference function; the end node preference function in step S3.1 is specifically: Let the communication reward of end node i in federated learning be denoted as , where denotes the communication reward factor of end node i, and and denote the communication time factor and the residual energy factor, respectively, where denotes the residual energy of end node i after completing the current federated learning communication, and denotes the energy consumption overhead factor of the end node, and is expressed as: wherein is defined as: Neglecting the aggregation time at the edge node, the transmission delay from edge node j to the cloud node is denoted as , where d is the distance between edge node j and the cloud node. is defined as: ； Step S3.2, the edge-end preference function is solved to realize edge-end association and determine the end node transmission power.
2. A system for implementing the federated learning service deployment method based on end node residual energy awareness according to claim 1, characterized in that, It comprises: a federated learning module for constructing a corresponding federated learning time model and an end node energy consumption model for each end node; a cloud game module for cloud game on the federated learning training time model and the end node computing energy consumption model obtained in step S1; an edge-end matching module for matching the end nodes participating in federated training obtained in step S2 according to the federated learning communication time model and the end node communication energy consumption model obtained in step S1.
3. A federated learning service deployment device based on end node residual energy awareness, characterized by, It comprises: a memory for storing required data; a processor for executing a computer program to realize the federated learning service deployment method based on end node residual energy perception as claimed in claim 1.
4. A computer readable storage medium, storing a computer program, wherein the computer program, when executed by a processor, enables the method of claim 1 to be implemented.