A fair opportunity federated learning incentive method and system based on game theory

Abstract

The present invention discloses a game-theory-based opportunity-fair federated learning incentive method and system. The method includes: the model owner determines the reward value for the current round; based on the reward value for the current round, each data node determines its contribution rate for participating in the current round of training and performs local model training corresponding to each data node; the model owner aggregates the local models of the data nodes based on the reputation values ​​of each data node to obtain a global model; a first utility function is constructed based on the cost of each data point in the current round, and a second utility function is constructed based on the model owner's revenue; the utility of the model owner and each data node is calculated based on the first and second utility functions, and the contribution rate and reward value of the next round of training are adjusted based on the utility results. This method achieves an optimal balance between the retraining contribution rate of power data nodes and the training cost, ensuring the training effect and duration of federated learning.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and in particular relates to a game theory-based opportunity fair federated learning incentive method and system. Background Art

[0002] In power system operations, accurate data is crucial for ensuring safe and stable system operation. However, due to the complexity of power systems and the increasing volume of data, detecting data anomalies has become particularly important. Power data anomalies can be caused by a variety of reasons, including equipment failure, severe weather, or human intervention. If these anomalies are not promptly detected and addressed, they can lead to system performance degradation or even equipment damage. Therefore, the development of power data anomaly detection technology has become a major focus in the current power industry. By adopting federated learning methods, different power data nodes can collaborate to train anomaly detection models, effectively identifying and locating anomalies in power data, improving the reliability and efficiency of system operations, and thereby ensuring the stability and security of power supply.

[0003] In recent years, game theory has provided rigorous mathematical models for analyzing the outcomes of various decisions and strategies in federated learning, providing a solid theoretical foundation for addressing the complex decision-making processes of multiple agents within federated learning. Federated learning incentive mechanisms based on game theory have been widely used. However, in non-IID scenarios, sufficient opportunity fairness is not achieved for data nodes. The current focus on global model accuracy in federated learning results intensifies the bias towards training results for active data nodes with large data volumes and better network performance, while neglecting the data models of passive data nodes with small data volumes and poorer performance. This means that trained anomaly detection models are more effective for the former, while less effective and generally poorer for the latter. However, in actual power projects, the latter constitute the majority of the total equipment, so biased model training can compromise detection effectiveness. Summary of the Invention

[0004] The present invention provides a game theory-based opportunity fair federated learning incentive method and system for solving the technical problem of poor generalization ability of global models.

[0005] In a first aspect, the present invention provides a method for motivating federated learning with fair opportunity based on game theory, comprising:

[0006] Obtain the capability value and reputation value of each data node in different rounds, and determine the target data node participating in each round of training based on the selection probability;

[0007] The model owner determines the reward value for the current round , according to the reward value of the current round Each data node determines its contribution rate for the current round of training and performs local model training corresponding to each data node. The contribution rate is the ratio of the local model accuracy increment of the data node to the data node's capability value.

[0008] The model owner aggregates the local models of the data nodes according to the reputation values ​​of each data node to obtain the global model;

[0009] Constructing a first utility function based on the cost of each data point in the current round, and constructing a second utility function based on the benefits of the model owner;

[0010] The utility of the model owner and each data node is calculated according to the first utility function and the second utility function, and the contribution rate of the next round of training and the reward value of the next round are adjusted according to the utility results until the global model converges.

[0011] In a second aspect, the present invention provides a fair opportunity federated learning incentive system based on game theory, comprising:

[0012] A determination module is configured to obtain the capability value of each data node and its reputation value in different rounds, and determine the target data node participating in the training in each round based on the selection probability;

[0013] The training module is configured to determine the reward value of the current round for the model owner , according to the reward value of the current round Each data node determines its contribution rate for the current round of training and performs local model training corresponding to each data node. The contribution rate is the ratio of the local model accuracy increment of the data node to the data node's capability value.

[0014] Aggregation module, configured so that the model owner aggregates the local models of data nodes according to the reputation values ​​of each data node to obtain a global model;

[0015] A construction module configured to construct a first utility function according to the expenses of each data point in the current round, and to construct a second utility function according to the benefits of the model owner;

[0016] An iteration module is configured to calculate the utility of the model owner and each data node according to the first utility function and the second utility function, and adjust the contribution rate of the next round of training and the reward value of the next round according to the utility results until the global model converges.

[0017] According to a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the game theory-based fair opportunity federated learning incentive method of any embodiment of the present invention.

[0018] In a fourth aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon. When the program instructions are executed by a processor, the processor executes the steps of the game theory-based opportunity fair federated learning incentive method of any embodiment of the present invention.

[0019] The game theory-based opportunity fair federated learning incentive method and system of this application effectively solves the problem of unfair opportunities between nodes caused by the strength of data nodes by evaluating the computing power of data nodes, and also improves the accuracy of the model in detecting data from different data nodes. By modeling the decision-making process of data nodes, the optimal solution for the training strategy of each data node is obtained, and the optimal balance between the retraining contribution rate and training cost of power data nodes is achieved, thereby ensuring the training effect and duration of federated learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0021] Figure 1 A flowchart of a game theory-based opportunity-fair federated learning incentive method provided in one embodiment of the present invention;

[0022] Figure 2 A schematic diagram of a fair federated learning solution model for power data anomaly detection provided by one embodiment of the present invention;

[0023] Figure 3 This is a structural block diagram of a game theory-based opportunity-fair federated learning incentive system provided by one embodiment of the present invention;

[0024] Figure 4 It is a structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0026] See also Figure 1 , which shows a flowchart of a game theory-based opportunity fair federated learning incentive method of this application.

[0027] like Figure 1 As shown in Figure 2, the opportunity-fair federated learning incentive method based on game theory specifically includes the following steps:

[0028] Step S101: Obtain the capability value and reputation value of each data node in different rounds, and determine the target data node participating in training in each round according to the selection probability.

[0029] In this step, the data node sends its trained local model to the power blockchain and calculates the accuracy of the local model using the test set. The expression for calculating the data node's capability value is:

[0030] ,

[0031] ,

[0032] ,

[0033] Where, is the capability value of the i-th data node, is the computing capability value of the data node, is the communication capability value of the data node, is the computing energy consumption of the data node, is the number of CPU cycles required to execute one data sample for the i-th data node, is the number of data samples of the i-th data node, is the effective capacitance coefficient of the computing chip at the i-th data node, is the clock frequency of the i-th data node, The duration for the i-th data node to communicate with the model owner.

[0034] Furthermore, each data node has an initial reputation value, and after each round of parameter training, the reputation value of the data node is updated based on the contribution value of the data node. In order to fully consider the recent performance of the data node, the forgetting function method will be used to update the reputation value. The expression for calculating the reputation value of different rounds is:

[0035] ,

[0036] ,

[0037] Where, is the reputation value of the i-th data node in the t-th round, is the reputation value of the i-th data node in the t-1th round, is a system parameter, representing the impact of recent data node behavior on reputation, determined by the model owner. is the contribution rate of the i-th data node in the t-th round, is the local model accuracy of the i-th data node in the t-th round, is the local model accuracy of the i-th data node in the t-1th round.

[0038] It should be noted that since the size of the reputation value is directly linked to the fairness of opportunity, the model owner will randomly select different data nodes based on the reputation value. The expression for calculating the selection probability is:

[0039] ,

[0040] Where, is the probability that the i-th data node is selected in the t-th round, is the number of selected data nodes, is the reputation value of the i-th data node in the t-th round, The total number of data nodes.

[0041] Step S102: The model owner determines the reward value for the current round. , according to the reward value of the current round Each data node determines its contribution rate for participating in the current round of training and performs local model training corresponding to each data node, where the contribution rate is the ratio of the local model accuracy increment of the data node to the data node's capability value.

[0042] In step S103 , the model owner aggregates the local models of the data nodes according to the reputation values ​​of the data nodes to obtain a global model.

[0043] In this step, the expression of the global model is:

[0044] ,

[0045] Where, is the global model of the tth round, is the local model of the i-th data node in the t-th round, is the reputation value of the i-th data node in the t-th round, is the reputation value of the j-th data node in the t-th round, The node set where the selected data node is located.

[0046] Step S104 : constructing a first utility function based on the cost of each data point in the current round, and constructing a second utility function based on the benefit of the model owner.

[0047] In this step, the expression of the first utility function is:

[0048] ,

[0049] Where, The utility of data nodes participating in federated learning, is the contribution rate of the i-th data node in the t-th round, is the contribution rate of the jth data node in the tth round, is the node set where the selected data node is located, is the capability value of the i-th data node, The reward value for a single round of federated learning, determined by the model owner;

[0050] The expression of the second utility function is:

[0051] ,

[0052] Where, The utility of model owners participating in federated learning, Represents the benefit of the sum of the contribution rates of the data nodes to the model owner.

[0053] Step S105: Calculate the utility of the model owner and each data node according to the first utility function and the second utility function, and adjust the contribution rate of the next round of training and the reward value of the next round according to the utility results until the global model converges.

[0054] In this step, the game between the data node and the model owner is designed as a Stackelberg game, with the model owner as the leader and the data node as the follower, and its Nash equilibrium is solved.

[0055] To solve the Nash equilibrium solution, we first solve the optimal strategy of the follower. By solving the maximum value point of the data nodes participating in federated learning, we can obtain the optimal strategy of each participating node in one round of iteration, that is, the optimal training contribution rate of the data node. The expression is:

[0056]

[0057] make have to ;

[0058] Simplifying, we can get:

[0059] ,

[0060] Where, for The optimal strategy for data nodes under Nash equilibrium, represents the partial derivative, is the capability value of the jth data node;

[0061] Next, we solve the optimal strategy of the model owner and bring in the optimal strategy of the data node to obtain:

[0062] ,

[0063]

[0064] Obviously, , this game has a unique Nash equilibrium solution.

[0065] In summary, the method of the present application, by evaluating the computing power of data nodes, effectively solves the problem of unfair opportunities between nodes caused by the strength of data nodes, and also improves the accuracy of the model in detecting data from different data nodes. By modeling the decision-making process of data nodes, the optimal solution for the training strategy of each data node is obtained, and the optimal balance between the retraining contribution rate and training cost of power data nodes is achieved, thereby ensuring the training effect and duration of federated learning.

[0066] In a specific embodiment, see Figure 2, a two-tier federated learning architecture is established, consisting of a data anomaly detection model owner and n power data nodes. Each data node holds a dataset, and all data nodes collaborate to execute the federated learning algorithm to train a global model. Model owners and data nodes interact through the power blockchain, and their interactions are supervised and fairly evaluated by the blockchain, ensuring fairness during node evaluation and reducing data node free-riding. The federated learning process proceeds in rounds, with a complete round of federated learning iteration. In each round, data nodes download the current global model and perform local training using their own datasets. The trained local model is then sent to the blockchain and aggregated by the model owner into the next global model.

[0067] In summary, this embodiment can achieve the following technical effects:

[0068] A calculation model of data node capabilities is introduced to achieve fairness among data nodes with different capabilities.

[0069] Often, the more powerful a data node is, the easier it is to improve model accuracy, which limits the performance of weaker nodes. This invention uses the quotient of model accuracy improvement and data node capability as a reward metric, providing fair incentives. Clients with different capabilities and the same contribution rate can receive the same reward, overcoming the remote discrimination between data nodes.

[0070] A dynamic reputation system is introduced to overcome the problem of reduced model generalization ability.

[0071] This paper uses dynamic reputation to take into account both the recent contribution rate and the past performance of data nodes in training. On the one hand, it sets the weight of local model aggregation based on reputation, giving nodes with long-term high contribution rates a greater weight in the model aggregation. On the other hand, reputation also determines the probability of a data node being selected for federated learning, thereby ensuring that the data of weaker nodes is not gradually neglected by federated learning due to multiple rounds of model training, thus overcoming the problem of reduced model generalization ability.

[0072] The incentive mechanism is designed by adopting the method of multi-round Stackelberg game, and the Nash equilibrium of the incentive mechanism is analyzed.

[0073] Existing incentive mechanism analysis only considers single-round game analysis, where decisions in a single round may or may not lead to a Nash equilibrium. However, with multiple rounds of game analysis, rational intelligent agents will converge to a Nash equilibrium. In this invention, the model owner acts as the leader, providing a reward rate for each round of training before the start of each round. The participating data centers act as followers, making optimal decisions based on the reward rate provided by the model owner. The data center's optimal decision influences the model owner's reward rate in the next round, ultimately leading the decisions of the model owner and the data center to converge toward a Nash equilibrium.

[0074] See also Figure 3 , which shows the structural block diagram of a fair opportunity federated learning incentive system based on game theory in this application.

[0075] like Figure 3 As shown, the fair opportunity federated learning incentive system 200 includes a determination module 210, a training module 220, an aggregation module 230, a construction module 240 and an iteration module 250.

[0076] The determination module 210 is configured to obtain the capability value of each data node and the reputation value in different rounds, and determine the target data node for each round of training based on the selection probability; the training module 220 is configured for the model owner to determine the reward value of the current round. , according to the reward value of the current round Each data node determines its contribution rate for participating in the current round of training and performs local model training corresponding to each data node, wherein the contribution rate is the ratio of the local model accuracy increment of the data node to the capability value of the data node; the aggregation module 230 is configured for the model owner to aggregate the local models of the data nodes according to the reputation value of each data node to obtain a global model; the construction module 240 is configured to construct a first utility function according to the overhead of each data point in the current round, and to construct a second utility function according to the benefit of the model owner; the iteration module 250 is configured to calculate the utility of the model owner and each data node according to the first utility function and the second utility function, and adjust the contribution rate of the next round of training and the reward value of the next round according to the utility results until the global model converges.

[0077] It should be understood that Figure 3 Modules and references documented in Figure 1 Therefore, the operations and features described above for the method and the corresponding technical effects also apply to Figure 3 The modules in it will not be described in detail here.

[0078] In other embodiments, embodiments of the present invention further provide a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor is caused to execute the game theory-based opportunity fair federated learning incentive method in any of the above method embodiments;

[0079] As an embodiment, the computer-readable storage medium of the present invention stores computer-executable instructions, and the computer-executable instructions are configured as follows:

[0080] Obtain the capability value and reputation value of each data node in different rounds, and determine the target data node participating in each round of training based on the selection probability;

[0081] The model owner determines the reward value for the current round , according to the reward value of the current round Each data node determines its contribution rate for the current round of training and performs local model training corresponding to each data node. The contribution rate is the ratio of the local model accuracy increment of the data node to the data node's capability value.

[0082] The model owner aggregates the local models of the data nodes according to the reputation values ​​of each data node to obtain the global model;

[0083] Constructing a first utility function based on the cost of each data point in the current round, and constructing a second utility function based on the benefits of the model owner;

[0084] The utility of the model owner and each data node is calculated according to the first utility function and the second utility function, and the contribution rate of the next round of training and the reward value of the next round are adjusted according to the utility results until the global model converges.

[0085] The computer-readable storage medium may include a program storage area and a data storage area. The program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created based on the use of the game theory-based fair opportunity federated learning incentive system. Furthermore, the computer-readable storage medium may include high-speed random access memory and may also include memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state memory device. In some embodiments, the computer-readable storage medium may optionally include memory remotely located relative to the processor, and such remote memory may be connected to the game theory-based fair opportunity federated learning incentive system via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0086] Figure 4 Schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Figure 4As shown, the device includes: a processor 310 and a memory 320. The electronic device may also include: an input device 330 and an output device 340. The processor 310, the memory 320, the input device 330 and the output device 340 may be connected via a bus or other means. Figure 4 The example of a bus connection is shown in FIG. Memory 320 is the aforementioned computer-readable storage medium. Processor 310 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions, and modules stored in memory 320, thereby implementing the above-mentioned method embodiment of the fair opportunity federated learning incentive method based on game theory. Input device 330 can receive input digital or character information, and generate key signal input related to user settings and function control of the fair opportunity federated learning incentive system based on game theory. Output device 340 may include a display device such as a display screen.

[0087] The electronic device can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided by the embodiment of the present invention.

[0088] As an embodiment, the electronic device is applied to a game theory-based opportunity fair federated learning incentive system, and is used for a client, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to:

[0089] Obtain the capability value and reputation value of each data node in different rounds, and determine the target data node participating in each round of training based on the selection probability;

[0090] The model owner determines the reward value for the current round , according to the reward value of the current round Each data node determines its contribution rate for the current round of training and performs local model training corresponding to each data node. The contribution rate is the ratio of the local model accuracy increment of the data node to the data node's capability value.

[0091] The model owner aggregates the local models of the data nodes according to the reputation values ​​of each data node to obtain the global model;

[0092] Constructing a first utility function based on the cost of each data point in the current round, and constructing a second utility function based on the benefits of the model owner;

[0093] The utility of the model owner and each data node is calculated according to the first utility function and the second utility function, and the contribution rate of the next round of training and the reward value of the next round are adjusted according to the utility results until the global model converges.

[0094] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods of each embodiment or certain portions of the embodiments.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A fair opportunity federated learning incentive method based on game theory, characterized by: include: Obtain the capability value and reputation value of each data node in different rounds, and determine the target data node participating in each round of training based on the selection probability; The model owner determines the reward value for the current round , according to the reward value of the current round Each data node determines its contribution rate for the current round of training and performs local model training corresponding to each data node. The contribution rate is the ratio of the local model accuracy increment of the data node to the data node's capability value. The expression for calculating the data node's capability value is: ， ， ， Where, is the capability value of the i-th data node, is the computing capability value of the data node, is the communication capability value of the data node, is the computing energy consumption of the data node, is the number of CPU cycles required to execute one data sample for the i-th data node, is the number of data samples of the i-th data node, is the effective capacitance coefficient of the computing chip at the i-th data node, is the clock frequency of the i-th data node, The duration of one communication between the i-th data node and the model owner; The model owner aggregates the local models of the data nodes according to the reputation values ​​of each data node to obtain the global model; Constructing a first utility function based on the cost of each data point in the current round, and constructing a second utility function based on the benefits of the model owner; The utility of the model owner and each data node is calculated according to the first utility function and the second utility function, and the contribution rate of the next round of training and the reward value of the next round are adjusted according to the utility results until the global model converges.

2. The game theory-based opportunity fair federated learning incentive method according to claim 1, characterized in that: in, The expression for calculating the reputation value of different rounds is: ， ， Where, is the reputation value of the i-th data node in the t-th round, is the reputation value of the i-th data node in the t-1th round, is a system parameter, representing the impact of recent data node behavior on reputation, determined by the model owner. is the contribution rate of the i-th data node in the t-th round, is the local model accuracy of the i-th data node in the t-th round, is the local model accuracy of the i-th data node in the t-1th round.

3. The game theory-based opportunity fair federated learning incentive method according to claim 2, characterized in that, The expression for calculating the selection probability is: ， Where, is the probability that the i-th data node is selected in the t-th round, is the number of selected data nodes, is the reputation value of the i-th data node in the t-th round, The total number of data nodes.

4. The game theory-based opportunity fair federated learning incentive method according to claim 1, characterized in that: The expression of the global model is: ， Where, is the global model of the tth round, is the local model of the i-th data node in the t-th round, is the reputation value of the i-th data node in the t-th round, is the reputation value of the j-th data node in the t-th round, The node set where the selected data node is located.

5. The game theory-based opportunity fair federated learning incentive method according to claim 1, characterized in that: The expression of the first utility function is: ， Where, The utility of data nodes participating in federated learning, is the contribution rate of the i-th data node in the t-th round, is the contribution rate of the jth data node in the tth round, is the node set where the selected data node is located, is the capability value of the i-th data node, The reward value for a single round of federated learning, determined by the model owner; The expression of the second utility function is: ， Where, The utility of the model owner in federated learning, Represents the benefit of the sum of the contribution rates of the data nodes to the model owner.

6. A fair opportunity federated learning incentive system based on game theory, characterized by: include: A determination module is configured to obtain the capability value of each data node and its reputation value in different rounds, and determine the target data node participating in the training in each round based on the selection probability; The training module is configured to determine the reward value of the current round for the model owner , according to the reward value of the current round Each data node determines its contribution rate for the current round of training and performs local model training corresponding to each data node. The contribution rate is the ratio of the local model accuracy increment of the data node to the data node's capability value. The expression for calculating the data node's capability value is: ， ， ， Where, is the capability value of the i-th data node, is the computing capability value of the data node, is the communication capability value of the data node, is the computing energy consumption of the data node, is the number of CPU cycles required to execute one data sample for the i-th data node, is the number of data samples of the i-th data node, is the effective capacitance coefficient of the computing chip at the i-th data node, is the clock frequency of the i-th data node, The duration of one communication between the i-th data node and the model owner; Aggregation module, configured so that the model owner aggregates the local models of data nodes according to the reputation values ​​of each data node to obtain a global model; A construction module configured to construct a first utility function according to the expenses of each data point in the current round, and to construct a second utility function according to the benefits of the model owner; An iteration module is configured to calculate the utility of the model owner and each data node according to the first utility function and the second utility function, and adjust the contribution rate of the next round of training and the reward value of the next round according to the utility results until the global model converges.

7. An electronic device, characterized in that: include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images