Federal learning competitor cooperation and punishment method oriented to data heterogeneity
By constructing three types of relationship graphs and a two-layer optimization framework, the problems of multi-relationship representation and conflict of interest in federated learning are solved, achieving efficient and stable collaboration in heterogeneous data scenarios and improving the fairness and security of collaboration.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-01-05
- Publication Date
- 2026-04-14
AI Technical Summary
Existing federated learning technologies lack accurate representation of multi-party relationships in multi-party collaborative scenarios, have imperfect conflict of interest avoidance mechanisms, unbalanced optimization goals, insufficient algorithm adaptability and efficiency, and lack of penalty mechanisms. They are difficult to achieve efficient and stable collaboration without conflicts of interest, especially in scenarios with complex data heterogeneity and competition intensity.
Three types of relationship graphs (undirected competition graph, directed benefit graph, and directed data usage graph) are constructed. Combining the Nash equilibrium principle and topological sorting algorithm, the collaborative relationships among participants are optimized through conflict of interest avoidance constraints and penalty mechanisms to ensure fairness and security. A two-layer optimization framework is adopted to balance individual and ecological contributions.
It enables structured characterization and quantitative analysis of relationships among participants, strictly avoids conflicts of interest, enhances collaboration enthusiasm, adapts to large-scale scenarios, ensures stable collaboration order, and improves the practicality and security of federated learning.
Smart Images
Figure CN121860089A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field, and in particular to a method for competitor collaboration and punishment in federated learning for data heterogeneity. Background Technology
[0002] Federated learning, as a distributed machine learning paradigm, effectively balances the value of data collaboration with the need for privacy protection by enabling collaborative iteration of model parameters without disclosing the original data of each participant. It has demonstrated broad application prospects in various cross-institutional scenarios such as healthcare, intelligent transportation, and financial services. Especially in multi-participant collaboration scenarios, federated learning can fully utilize the local data resources of each participant to improve the generalization performance of the global model, thus becoming one of the core technologies for cross-domain data collaboration.
[0003] With the development of federated learning technology, researchers have gradually focused on the imperfect collaboration among participants, with competition being a common and critical issue in cross-institutional collaboration. In existing research, some scholars have attempted to achieve collaboration among competing participants through alliance structure optimization. For example, Chaudhury's team treated federated learning participants as a large alliance, exploring the construction of a core stable alliance structure; other scholars have analyzed the collaborative equilibrium problem in federated learning through linear regression tasks. Furthermore, some studies have applied meta-learning and multi-task learning techniques to personalized model construction, attempting to mitigate the conflicts of interest arising from competition.
[0004] However, existing federated learning techniques for competing participants still have many shortcomings that urgently need to be addressed, specifically in the following aspects: A lack of precise multi-faceted relationship representation systems; existing research has failed to effectively integrate competitive relationships, data benefit relationships, and actual collaborative relationships among participants, failing to provide comprehensive data support for collaborative strategy formulation. Although some studies mention competitive or benefit relationships, they have not constructed structured relationship models (such as competition graphs, benefit graphs, and data usage graphs) using graph theory, as in this application, resulting in vague relationship descriptions and difficulty in quantitative analysis; Inadequate conflict of interest avoidance mechanisms; existing technologies mostly adopt simple alliance division or parameter isolation methods, failing to establish strict conflict avoidance principles and effectively preventing direct or indirect transfer of benefits between competing participants, such as failing to guarantee that "the friend of my enemy is my enemy." The implementation of this practical logic undermines the fairness and security of collaboration; the optimization objectives are unbalanced, with existing research either focusing on maximizing the performance of the global model or only on improving the utility of a single participant, failing to achieve a balance between "maximizing the utility of participants and maximizing the contribution to the ecosystem," which easily leads to problems such as "free-riding" or "excessive sacrifice of individual interests," reducing the participants' enthusiasm for collaboration; the algorithm's adaptability and efficiency are insufficient, with existing solutions struggling to adapt to scenarios with large-scale participants and lacking efficient processing mechanisms for special problems such as dominance graph ring structures, resulting in poor practicality in real-world complex scenarios (such as multi-hospital intensive care data collaboration); fifth, the lack of a punishment mechanism, with existing technologies failing to establish clear quantitative standards for punishing behaviors that violate the conflict avoidance principle during collaboration, making it impossible to guarantee the stable progress of the collaboration order.
[0005] In real-world cross-institutional federated learning scenarios (such as eICU data collaboration in the medical field and intelligent service collaboration among multiple enterprises), there are numerous participants with complex relationships, significant differences in data heterogeneity and competitive intensity. The aforementioned shortcomings of existing technologies make it difficult to achieve efficient and stable collaboration without conflicts of interest, which severely limits the application of federated learning technology in complex competitive scenarios.
[0006] To address this, we propose a collaborative and penalized method for federated learning competitors, oriented towards data heterogeneity. Summary of the Invention
[0007] This application aims to at least partially solve one of the technical problems in the aforementioned technologies.
[0008] To achieve the above objectives, the first aspect of this application proposes a method for competitor collaboration and punishment in federated learning oriented towards data heterogeneity, comprising the following steps:
[0009] 1) Construct three types of relationship graphs among federated learning participants: an undirected competition graph representing the competitive relationship between any two participants, a directed benefit graph representing the data benefit relationship between any two participants, and a directed data use graph representing the actual collaborative relationship between any two participants. The directed data use graph is a subgraph of the directed benefit graph.
[0010] (2) Based on the extended rule of "the enemy of my friend is my enemy", set conflict of interest avoidance constraints to ensure that federated learning participants with competitive relationships are mutually unreachable in the directed data use graph.
[0011] (3) Combining the Nash equilibrium principle to construct a two-level optimization problem, so as to maximize the contribution of each federated learning participant to the federated learning ecosystem without sacrificing their own utility;
[0012] (4) Construct the dominance graph and verify its acyclic property. Then, use the topological sorting algorithm to solve the bi-level optimization problem, output the adjacency matrix of the directed data use graph, and finally determine the collaborative relationship between federated learning participants.
[0013] In addition, the federated learning competitor collaboration and penalty method for data heterogeneity proposed above in this application may also have the following additional technical features:
[0014] As a further description of the above technical solution:
[0015] In step (1), the undirected competition graph contains a set of nodes and a set of edges. The set of nodes consists of all federated learning participants, and the edges in the set of edges are used to represent the competitive relationship between the participants represented by the corresponding two nodes.
[0016] The adjacency matrix of an undirected competitive graph is an n×n matrix, with all elements on its main diagonal being 0. If two different participants are in competition, the corresponding element in the adjacency matrix is 1; otherwise, it is 0.
[0017] The adjacency matrix is obtained by each federated learning participant reporting their competitor information to the central server.
[0018] As a further description of the above technical solution:
[0019] In step (1), the directed benefit graph contains a set of nodes and a set of edges. The set of nodes consists of all federated learning participants. When a directed edge points from federated learning participant A to federated learning participant B, it indicates that the data representing participant A can bring benefits to participant B.
[0020] The adjacency matrix of a directed benefit graph is an n×n matrix, and the value of the matrix elements is positively correlated with the degree of benefit.
[0021] The importance of each federated learning participant to the ecosystem is measured by a potential level indicator, which is the sum of the values of the benefits that participant brings to all other participants.
[0022] As a further description of the above technical solution:
[0023] In step (1), the directed data uses a graph containing a set of nodes and a set of edges. The set of nodes consists of all federated learning participants. When a directed edge points from federated learning participant C to federated learning participant D, it indicates that participant C provides model update contributions to participant D during the federated learning training process.
[0024] For directed data, the adjacency matrix of the graph is an n×n decision variable matrix. The matrix elements take values of 0 or 1. A value of 1 indicates that there is a model update contribution relationship between the two corresponding participants, while a value of 0 indicates that there is no such relationship.
[0025] When a participant's data cannot benefit another participant, the corresponding element in the decision variable matrix must be 0.
[0026] As a further description of the above technical solution:
[0027] In step (2), the conflict of interest avoidance constraint is as follows: for any two federated learning participants who are in competition, the sum of the elements of the adjacency matrix of the directed data usage graph on each path from one participant to the other participant does not exceed the length of the path minus one.
[0028] As a further description of the above technical solution:
[0029] In step (3), the inner optimization objective is to maximize the utility of each federated learning participant, which is calculated by summing the numerical values of the benefits that participant receives from all other participants who provide model update contributions.
[0030] The outer optimization objective is to maximize the total contribution of each federated learning participant to the other participants, which is calculated by summing the benefit values that participant provides to all other participants who accept its model update contributions;
[0031] The constraints of the bilevel optimization problem include the value constraints of the adjacency matrix of the directed data use graph, the aforementioned conflict of interest avoidance constraints, and the constraint that the directed data use graph is a subgraph of the directed benefit graph.
[0032] As a further description of the above technical solution:
[0033] Step (4) The dominance graph is constructed as follows:
[0034] (A) Traverse all pairs of federated learning participants that are in competition, obtain all paths between each pair of competing participants in the directed benefit graph, and form the corresponding path set;
[0035] (B) Divide all paths ending with the same edge into a path subset;
[0036] (C) Treat each path subset as a node in the dominating graph. If the last edge of a path in the first path subset belongs to a path in the second path subset, then construct a directed edge in the dominating graph from the corresponding node of the first path subset to the corresponding node of the second path subset, thus completing the construction of the dominating graph.
[0037] As a further description of the above technical solution:
[0038] When the dominance graph has a cyclic structure, a tiered penalty is used to punish federated learning participants with competing relationships. The tiered penalty satisfies the following:
[0039] (1) The penalty level is the smallest non-negative real number that satisfies the constraints of all participants in a competitive relationship;
[0040] (2) The penalty function is the difference between the maximum possible utility and the actual utility of each participant in the competitive relationship, where the maximum possible utility refers to the maximum utility that the participant can obtain under unconstrained conditions;
[0041] (3) The upper limit of punishment is the product of the punishment level and the maximum possible utility of the corresponding participant, so as to ensure the fair distribution of punishment.
[0042] Advantages of this invention:
[0043] Based on the federated learning competitor collaboration and penalty method for data heterogeneity in this application, a competition graph, benefit graph, and data usage are constructed using graph theory. Figure 3 The meta-system quantifies the competitive relationships, data benefit value, and actual collaboration permissions among participants, enabling a structured characterization and quantifiable analysis of relationships, thus solving the problem of ambiguous relationship representation in existing technologies.
[0044] Strict conflict of interest avoidance, extending the principle that "the enemy's friend is the enemy," proposes constraints that make data usage unreachable for competing parties. By constructing mathematical judgment conditions, it strictly avoids direct and indirect transfer of benefits by competitors, ensuring fairness and security in collaboration.
[0045] By adopting a two-tier optimization framework of contribution without sacrifice, we first maximize the utility of each participant (FL-PT) and then maximize their contribution to the ecosystem, balancing individual interests with ecological value, avoiding free-riding or excessive sacrifice, and enhancing the enthusiasm for participation and collaboration.
[0046] Through a series of algorithms such as dominance graph construction, dominant path selection, and X matrix update, it can handle scenarios with large-scale participants and efficiently solve path dominance relationships and loop structure problems, adapting to scenarios with data heterogeneity and different competition intensities.
[0047] Clearly define quantitative standards for the level of punishment for innocent parties and competitors, and establish a punishment evaluation system for different scenarios through experiments to ensure the stable progress of collaborative order.
[0048] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description
[0049] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:
[0050] Figure 1 This is a schematic diagram of hospital relationships in a federated learning competitor collaboration and penalty method for data heterogeneity according to an embodiment of this application;
[0051] Figure 2 This is an embodiment of a federated learning method for competitor collaboration and penalty in response to data heterogeneity, based on one embodiment of this application. Diagram illustrating mutual competition;
[0052] Figure 3 This is a schematic diagram illustrating the relationships, problems, and optimizations of a federated learning participant-oriented collaborative and penalized method for data heterogeneity-based federated learning according to an embodiment of this application.
[0053] Figure 4 This is a path between competing nodes in a federated learning method for data heterogeneity-oriented collaboration and penalty according to an embodiment of this application. Schematic diagram. Detailed Implementation
[0054] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.
[0055] The following describes, with reference to the accompanying drawings, a method for competitor collaboration and punishment in federated learning for data heterogeneity according to an embodiment of this application.
[0056] like Figure 1-4 As shown in the embodiments of this application, the federated learning competitor collaboration and penalty method for data heterogeneity may include the following steps:
[0057] 1) Construct three types of relationship graphs among federated learning participants: an undirected competition graph to represent the competitive relationship between any two federated learning participants, a directed benefit graph to represent the data benefit relationship between any two federated learning participants, and a directed data usage graph to represent the actual collaborative relationship between any two federated learning participants. The directed data usage graph is a subgraph of the directed benefit graph.
[0058] The undirected competition graph consists of a node set and an edge set. The node set comprises all federated learning participants, and the edges in the edge set represent the competitive relationship between two corresponding nodes representing federated learning participants. The adjacency matrix of the undirected competition graph is also included. for matrix( (the number of participants in federated learning), satisfying the formula (The main diagonal elements are 0). ( and and (There is a competitive relationship) ( and and The adjacency matrix of the undirected competitive graph (without competitive relationships) is obtained by each federated learning participant reporting its competitor information to the central server;
[0059] The directed benefit graph consists of a node set and an edge set. The node set comprises all federated learning participants. A directed edge in the edge set pointing from the first federated learning participant to the second federated learning participant indicates that the data from the first federated learning participant can benefit the second federated learning participant. The adjacency matrix of the directed benefit graph is also included. for matrix( (Number of participants in federated learning), elements in the matrix The magnitude of the value corresponds to the degree of benefit; the larger the value, the greater the benefit. The importance of each federated learning participant to the federated learning ecosystem is measured by the potential level index, which satisfies the formula... (in Adjacency matrix The Middle Line 1 The elements of the column represent right (the degree of benefit);
[0060] The directed data usage graph includes a node set and an edge set. The node set consists of all federated learning participants. A directed edge in the edge set pointing from the third federated learning participant to the fourth federated learning participant indicates that the third federated learning participant contributes to the model update of the fourth federated learning participant during the federated learning training process. The adjacency matrix of the directed data usage graph is also included. for Decision variable matrix ( (the number of participants in federated learning), satisfying the formula ( express Towards Make a contribution (indicating no contribution), and when There will always be a time when... ( (The corresponding element in the adjacency matrix of the directed benefit graph).
[0061] 2) Based on the principle that "the enemy of my friend is my enemy," an extension of this principle is used to establish a conflict of interest avoidance constraint. This ensures that participants in a federated learning relationship who are in competition with each other are mutually unreachable in the directed data usage graph. The conflict of interest avoidance constraint satisfies the formula... ,in For participants in directed benefit graphs from federated learning arrive Path length, Use adjacent nodes in the graph for directed data and Decision variables;
[0062] The specific content of the conflict of interest avoidance constraint is as follows: for any two federated learning participants who have a competitive relationship... In the directed benefit graph from All paths (in , , (where is the path length), all satisfy the formula That is, for each path, the sum of the elements in the graph adjacency matrix corresponding to the directed data is less than the length of the path, ensuring... Unreachable in directed data usage graph .
[0063] 3) Combining the Nash equilibrium principle, a two-level optimization problem is constructed to maximize the contribution of each federated learning participant to the federated learning ecosystem without sacrificing their own utility. The inner objective function of the two-level optimization problem is: The outer objective function is ;
[0064] The inner-layer optimization objective is to maximize the utility of each federated learning participant, which is expressed by the formula... Calculate (where Decision variables for directed data using graphs (As the benefit weights of the directed benefit graph), the outer optimization objective is to maximize the total contribution of each federated learning participant to the total contribution of other federated learning participants, expressed by the formula... The constraints for calculating the bilevel optimization problem include: (Decision variable value constraints), the aforementioned conflict of interest avoidance constraints, (when (At that time, the data uses the constraint of the benefit map subplot).
[0065] 4) By constructing a dominance graph and verifying its acyclic property, a topological sorting algorithm is used to solve the bi-level optimization problem, outputting the adjacency matrix of the directed data usage graph, and determining the collaborative relationships among federated learning participants. The steps for constructing the dominance graph include:
[0066] (1) Traverse all pairs of federated learning participants with competing relationships, obtain all paths from one federated learning participant to another in the directed benefit graph, and form a path set. (in This is a set of tuples representing competing relationships. For the first (the set of paths corresponding to each competing tuple);
[0067] (2) Connect all with the same edge Divide the path at the end into a path subset. ,satisfy ;
[0068] (3) Treat each path subset as a node of the dominating graph. If the end edge of a path in the first path subset is an edge of a path in the second path subset, then construct a directed edge from the corresponding node of the first path subset to the corresponding node of the second path subset in the dominating graph to complete the construction of the dominating graph.
[0069] (4) In the above steps, when the dominance graph has a ring structure, a tiered penalty is used to punish federated learning participants with competing relationships. The tiered penalty satisfies the following: penalty level
[0070] To satisfy the minimum non-negative real number of the constraints for all competing federated learning participants, the penalty function satisfies the formula... (in for The maximum possible utility for (actual utility), the penalty upper limit satisfies the formula This ensures the fair distribution of punishment.
[0071] The above methodological steps are applicable to a certain number of federated learning participants. In cross-departmental scenarios, including federated learning collaboration in healthcare, drug development, digital banking, and electric vehicle charging services, a solver is used to solve constrained optimization problems. The objective function of the constrained optimization problem is... The constraints include bi-level optimization objective function constraints, conflict of interest avoidance constraints, and penalty function constraints.
[0072] The penalty level for uncompetitive federated learning participants (innocent participants) is set to 0, i.e., satisfying the formula... (in (For innocent parties), the model performance of federated learning participants outperforms both locally trained models and models trained using traditional federated learning methods. Model performance is measured by mean squared error (MSE) or mean test accuracy (MTA), and meets the following criteria. and ,or and .
[0073] The following provides further explanation of Example 2:
[0074] Construct an undirected competition graph (e.g.) Figure 1 ),Sure and , and There are competing relationships among the participants, while the other participants are not in competition with each other. Based on this, an adjacency matrix is constructed. ,conform to A constraint where the element corresponding to a competitive relationship is 1;
[0075] Constructing a directional benefit graph ( Figure 2 Based on directed edges and benefit parameters, an adjacency matrix is constructed. ,calculate Potential level indicators It conforms to the calculation formula and parameter settings of the potential level index;
[0076] Constructing a directed data usage graph ( Figure 3 Construct a decision variable matrix ,in The value can be 1. hour ;
[0077] Verify conflict of interest avoidance constraints ( Figure 2 , Figure 3 ):against Defined competitive pairs Extracting the directed benefit graph from arrive path (Path length) )and (Path length) ), calculate the values on the path according to the constraint formula. Matrix elements and Satisfy the constraints and ensure and They are not reachable from each other in the data usage graph;
[0078] Construct and solve a bi-level optimization problem based on the Nash equilibrium principle. The goal is to maximize participant utility in the inner level and maximize ecological contribution in the outer level. For example, based on uploaded files and Matrix computation: inner layer utility outer layer contribution The constraints include The aforementioned conflict of interest constraints and hour ;
[0079] Construct the dominance graph and solve it. Figure 4 ), Traverse the competing pairs , Extract all paths from the directed benefit graph to form a path set. Divide the path into subsets based on the last edge, for example, by edge. Paths with trailing edges are subsets Construct a dominance graph if The end belong If a certain path is selected, then construct... Given directed edges, after verifying that the dominating graph is acyclic, solve the bi-level optimization problem using the topological sorting algorithm, and output the optimized result. A matrix is used to determine the final collaborative relationships: ;
[0080] Penalty mechanism adaptation and performance verification: If the dominance graph contains a cycle structure, set the penalty level based on the level penalty rule definition. (The smallest non-negative real number that satisfies all constraints, its value derived from the penalty parameter setting of the uploaded file), to compete for participants. For example, its maximum possible utility (based on Matrix calculation), penalty upper limit To ensure fair punishment for innocent individuals with no competitive relationship. Its penalty function value The MSE and mean test accuracy (MTA) of the model trained by this method both meet the requirements. .
[0081] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0082] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0083] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.
Claims
1. A method for competitor collaboration and penalty in federated learning oriented towards data heterogeneity, characterized in that, Includes the following steps: 1) Construct three types of relationship graphs among federated learning participants: an undirected competition graph representing the competitive relationship between any two participants, a directed benefit graph representing the data benefit relationship between any two participants, and a directed data use graph representing the actual collaborative relationship between any two participants. The directed data use graph is a subgraph of the directed benefit graph. (2) Based on the extended rule of "the enemy of my friend is my enemy", set conflict of interest avoidance constraints to ensure that federated learning participants with competitive relationships are mutually unreachable in the directed data use graph. (3) Combining the Nash equilibrium principle to construct a two-level optimization problem, so as to maximize the contribution of each federated learning participant to the federated learning ecosystem without sacrificing their own utility; (4) Construct the dominance graph and verify its acyclic property. Then, use the topological sorting algorithm to solve the bi-level optimization problem, output the adjacency matrix of the directed data use graph, and finally determine the collaborative relationship between federated learning participants.
2. The method for competitor collaboration and penalty in federated learning oriented towards data heterogeneity according to claim 1, characterized in that, In step (1), the undirected competition graph contains a set of nodes and a set of edges. The set of nodes consists of all federated learning participants, and the edges in the set of edges are used to represent the competitive relationship between the participants represented by the corresponding two nodes. The adjacency matrix of an undirected competitive graph is an n×n matrix, with all elements on its main diagonal being 0. If two different participants are in competition, the corresponding element in the adjacency matrix is 1; otherwise, it is 0. The adjacency matrix is obtained by each federated learning participant reporting their competitor information to the central server.
3. The method for competitor collaboration and penalty in federated learning oriented towards data heterogeneity according to claim 1, characterized in that, In step (1), the directed benefit graph contains a set of nodes and a set of edges. The set of nodes consists of all federated learning participants. When a directed edge points from federated learning participant A to federated learning participant B, it indicates that the data representing participant A can bring benefits to participant B. The adjacency matrix of a directed benefit graph is an n×n matrix, and the value of the matrix elements is positively correlated with the degree of benefit. The importance of each federated learning participant to the ecosystem is measured by a potential level indicator, which is the sum of the values of the benefits that participant brings to all other participants.
4. The method for competitor collaboration and penalty in federated learning oriented towards data heterogeneity according to claim 1, characterized in that, In step (1), the directed data uses a graph containing a set of nodes and a set of edges. The set of nodes consists of all federated learning participants. When a directed edge points from federated learning participant C to federated learning participant D, it indicates that participant C provides model update contributions to participant D during the federated learning training process. For directed data, the adjacency matrix of the graph is an n×n decision variable matrix. The matrix elements take values of 0 or 1. A value of 1 indicates that there is a model update contribution relationship between the two corresponding participants, while a value of 0 indicates that there is no such relationship. When a participant's data cannot benefit another participant, the corresponding element in the decision variable matrix must be 0.
5. The method for competitor collaboration and penalty in federated learning oriented towards data heterogeneity according to claim 1, characterized in that, In step (2), the conflict of interest avoidance constraint is as follows: for any two federated learning participants who are in competition, the sum of the elements of the adjacency matrix of the directed data usage graph on each path from one participant to the other participant does not exceed the length of the path minus one.
6. The method for competitor collaboration and penalty in federated learning oriented towards data heterogeneity according to claim 1, characterized in that, In step (3), the inner optimization objective is to maximize the utility of each federated learning participant, which is calculated by summing the numerical values of the benefits that participant receives from all other participants who provide model update contributions. The outer optimization objective is to maximize the total contribution of each federated learning participant to the other participants, which is calculated by summing the benefit values that participant provides to all other participants who accept its model update contributions; The constraints of the bilevel optimization problem include the value constraints of the adjacency matrix of the directed data use graph, the aforementioned conflict of interest avoidance constraints, and the constraint that the directed data use graph is a subgraph of the directed benefit graph.
7. The method for competitor collaboration and penalty in federated learning oriented towards data heterogeneity according to claim 1, characterized in that, Step (4) The dominance graph is constructed as follows: (A) Traverse all pairs of federated learning participants that are in competition, obtain all paths between each pair of competing participants in the directed benefit graph, and form the corresponding path set; (B) Divide all paths ending with the same edge into a path subset; (C) Treat each path subset as a node in the dominating graph. If the last edge of a path in the first path subset belongs to a path in the second path subset, then construct a directed edge in the dominating graph from the corresponding node of the first path subset to the corresponding node of the second path subset, thus completing the construction of the dominating graph.
8. The method for competitor collaboration and penalty in federated learning oriented towards data heterogeneity according to claim 7, characterized in that, When the dominance graph has a cyclic structure, a tiered penalty is used to punish federated learning participants with competing relationships. The tiered penalty satisfies the following: (1) The penalty level is the smallest non-negative real number that satisfies the constraints of all participants in a competitive relationship; (2) The penalty function is the difference between the maximum possible utility and the actual utility of each participant in the competitive relationship, where the maximum possible utility refers to the maximum utility that the participant can obtain under unconstrained conditions; (3) The upper limit of punishment is the product of the punishment level and the maximum possible utility of the corresponding participant, so as to ensure the fair distribution of punishment.