Worker training reward determination method for multi-task competition

Abstract

This invention belongs to the technical field of worker training reward determination strategies in federated learning systems, specifically relating to a method for determining worker training rewards in the face of multi-task competition. This invention is based on a scenario in a federated learning system where there are multiple task publishers and multiple worker nodes, in which there is competition among the task publishers. For task publishers, in order to gain a competitive advantage and attract more high-quality worker nodes, they need to increase the rewards they pay. However, from their own perspective, task publishers also want to obtain higher returns by lowering rewards. For worker nodes, the presence of multiple task publishers gives them the opportunity to choose the optimal contract to ensure their own returns. This invention can simultaneously satisfy individual rationality constraints and incentive compatibility constraints, balancing the relationship between task publishers and worker nodes, enabling both roles in the system to obtain expected returns, forming a virtuous cycle, and achieving the goal of incentivizing worker nodes to join the federated learning system.

Description

Technical Field

[0001] This invention belongs to the technical field of worker training compensation determination strategy in federated learning systems, specifically relating to a method for determining worker training compensation in the context of multi-tasking competition. Background Technology

[0002] The training efficiency of a federated learning system is closely related to the quality of the local model updated by the worker nodes in each iteration. However, worker nodes consume various resources, including computing and communication resources, during federated learning tasks, resulting in significant costs. Furthermore, malicious server nodes can infer important data information using intermediate gradients, jeopardizing the security of worker nodes' private data. All these factors greatly diminish the incentive for worker nodes to participate in federated learning tasks. Therefore, the task publisher in the system needs to design appropriate incentive mechanisms to determine worker compensation and attract worker nodes to participate in federated learning tasks.

[0003] Furthermore, federated learning systems suffer from information asymmetry between task publishers and worker nodes. Key information such as the data quality and computing power of each worker node is invisible to the task publisher, who can only rely on experience to judge the quality distribution of worker nodes and thus infer the probability of each worker node belonging to a certain type. To address this issue, most studies utilize contract theory to map the quality types of worker nodes to appropriate rewards, thereby attracting more high-quality worker nodes to join the federated learning system.

[0004] The current research supports scenarios with one task publisher and multiple worker nodes, meaning that a task publisher designs contracts for multiple nodes participating in the task. This assumption is unreasonable in practical applications. Summary of the Invention

[0005] The purpose of this invention is to provide a method for determining worker training rewards in a multi-task competition scenario. This method is based on contract theory and aims to attract more high-quality worker nodes to join federated learning.

[0006] A method for determining worker training compensation in a multi-job competition setting includes the following steps:

[0007] 1) In a federated learning system, multiple task publishers publish training tasks and submit pre-designed contracts to the blockchain for storage;

[0008] 2) Worker nodes in the system view different tasks and contracts submitted by multiple task publishers through blocks, and choose to sign the best contract for themselves; after signing, worker nodes use local datasets to train models.

[0009] 3) Worker nodes upload the trained model to the blockchain for storage;

[0010] 4) The task publisher retrieves the updated model from the blockchain and aggregates all the models locally;

[0011] 5) After the federated learning task is completed, the task issuer pays the corresponding remuneration to the worker nodes according to the contract signed with the worker nodes, and saves the payment record on the blockchain.

[0012] Further, step 2) specifically includes:

[0013] 2.1) The cost of worker nodes participating in tasks in the optimal contract includes training cost and communication cost;

[0014] Each worker node has its own local dataset, and training the local model once in each iteration consumes computation time. and CPU resources

[0015]

[0016]

[0017] Among them, c n The number of CPU cycles required to train a worker node with one unit of data; s n f is the size of the local dataset for the worker node; n The CPU cycle frequency of the worker node;

[0018] The process of worker nodes delivering the updated training model to the task issuer is achieved through wireless communication. Each iteration of worker node n requires communication time. and communication resources for:

[0019]

[0020]

[0021] Where σ is a fixed constant representing the size of the model parameters passed by all worker nodes after each iteration; ρ n For transmission power; h n N is the channel gain; N0 is the background noise.

[0022] The total time consumed by worker node n in one global iteration process for:

[0023]

[0024] Total energy consumed by worker node n in one global iteration process for:

[0025]

[0026] 2.2) In the federated learning system, the stakeholders who benefit are divided into two roles: task issuers and worker nodes.

[0027] For the task issuer, they can infer the probability of each worker node belonging to a certain type based on past experience; assuming the quality of worker nodes is divided into M classes, and arranged in ascending order, represented as θ1<…<θ m <…<θ M ,m∈{1,…,M};Each worker node belongs to θ M The probability of type p m Then there is

[0028] Due to the information asymmetry between the task publisher and the worker nodes, the task publisher i designs specific contract abstractions for worker nodes of different quality types. Wherein, the CPU cycle frequency of worker node n is f n The revenue corresponding to this type of worker node is

[0029] For a given contract, the revenue function U that the task issuer obtains from worker node n is... p for:

[0030]

[0031] Because there is competition among different task publishers in a federated learning system, and each task publisher strives to maximize its own interests, the probability pc of a worker node selecting a task publisher is introduced. i Then the overall revenue function of the task publisher for:

[0032]

[0033] For worker node n that has signed the contract submitted by task publisher i, its corresponding revenue function U a for:

[0034]

[0035] Then the overall revenue function of the worker node for:

[0036]

[0037] 2.3) When information asymmetry exists, the contract designed by the task issuer needs to simultaneously satisfy individual rationality constraints and incentive compatibility;

[0038] Definition 1: Individual rationality constraint: The condition for each worker node to agree to participate in a federated learning task is that the reward after completing the task is non-negative, as expressed below:

[0039]

[0040] Definition 2: Incentive Compatibility Constraint: The condition for each worker node to obtain the maximum benefit is to select a contract designed according to its own quality type, as shown below:

[0041]

[0042] The calculation of the optimal contract first considers the relaxation problem without monotonicity constraints, and then checks whether the obtained solution satisfies the monotonicity condition; simultaneously, worker nodes consider maximizing their own revenue and use the optimal p. * c i Finally, by simplifying the constraints based on the local downward excitation constraints, the simplest optimal contract is obtained, calculated as follows:

[0043]

[0044] st

[0045]

[0046]

[0047]

[0048]

[0049] 2.4) The specific iterative process is as follows:

[0050] First, using the contract of task publisher A as a constant, the optimal contract of task publisher B is derived using this constant condition. Then, using the optimal contract of task publisher B as a constant, the optimal contract of task publisher A is derived again using this condition. The above process is repeated until the optimal contracts of task publishers A and B converge. At this point, the optimal contract corresponding to each type of worker node can be calculated. This equilibrium condition simultaneously satisfies the individual rationality constraint and incentive compatibility constraint in each contract. That is, while ensuring non-negative income under competitive conditions, multiple task publishers also ensure that each worker node can choose the optimal contract that suits its own conditions and obtain the best income.

[0051] The beneficial effects of this invention are as follows:

[0052] This invention is based on a scenario in a federated learning system with multiple task publishers and multiple worker nodes, where there is competition among the task publishers. To gain a competitive advantage and attract more high-quality worker nodes, task publishers need to increase their compensation. However, from a self-interested perspective, they may also lower their compensation to obtain higher returns on tasks. For worker nodes, the presence of multiple task publishers gives them the opportunity to choose the optimal contract to ensure their own returns. This invention is not only applicable to scenarios with multiple task publishers that better meet practical needs, but also satisfies individual rationality constraints and incentive compatibility constraints. It can balance the relationship between task publishers and worker nodes, ensuring that both roles in the system can obtain expected returns, leading to virtuous cycle and ultimately incentivizing worker nodes to join the federated learning system. Attached Figure Description

[0053] Figure 1 This is a diagram of the architecture of the present invention. Detailed Implementation

[0054] The present invention will now be further described with reference to the accompanying drawings.

[0055] This invention relates to a strategy for determining worker training compensation, specifically a method for determining worker training compensation based on contract theory and oriented towards multi-task competition in a federated learning system with multiple task publishers.

[0056] This invention is based on a scenario in a federated learning system where there are multiple task publishers and multiple worker nodes, and in this scenario, there is competition among the task publishers. For task publishers, in order to gain a competitive advantage and attract more high-quality worker nodes, they need to increase the rewards they offer. However, from their own perspective, task publishers also want to obtain higher returns on tasks by lowering their rewards. For worker nodes, the presence of multiple task publishers gives them the opportunity to choose the optimal contract to ensure their own returns.

[0057] The federated learning system consists of task publishers, workers, and a blockchain. There are multiple task publishers and multiple workers, and the worker nodes are heterogeneous. The blockchain resides in the cloud and is used for storage and computation. Furthermore, information asymmetry exists between the two roles. Therefore, when designing contracts for worker nodes of different quality types, it is necessary to calculate the cost and benefit functions of worker nodes participating in federated learning tasks, and thus design the optimal contract.

[0058] 1) The cost of worker nodes participating in tasks in the optimal contract includes training cost and communication cost.

[0059] Each worker node has its own local dataset, and training the local model once in each iteration consumes computation time. and CPU resources Assume that the number of CPU cycles required for a worker node to train one unit of data is c. n The local dataset size of the worker node is s n The CPU cycle frequency of the worker node is f n The calculation is as follows:

[0060]

[0061]

[0062] The process of worker nodes delivering the updated training model to the task issuer is achieved through wireless communication. Each iteration of worker node n requires communication time. and communication resources Assume that the magnitude of the model parameters passed by all worker nodes after each iteration is a fixed constant σ, and the transmission power is ρ. n The channel gain is h n The background noise is N0. The calculation is as follows:

[0063]

[0064]

[0065] That is, the total time consumed by worker node n in one global iteration process. The calculation is as follows:

[0066]

[0067] Total energy consumed by worker node n in one global iteration process The calculation is as follows:

[0068]

[0069] 2) In the federated learning system, the beneficiaries are divided into two roles: task issuers and worker nodes.

[0070] For task publishers, they can infer the probability of each worker node belonging to a certain type based on past experience. Assume the quality of worker nodes is divided into M classes, arranged in ascending order, represented as θ1 < ... < θ m <…<θ M ,m∈{1,…,M}. Each worker node belongs to θ M The probability of type p m Then there is This invention specifically discusses the scenario where there are two task issuers. Therefore, there are... The probability that a worker node selects the i-th task publisher is pc. i .

[0071] Due to the information asymmetry between the task publisher and the worker nodes, the task publisher i can design specific contracts for worker nodes of different quality types, which can be abstracted as follows: Wherein, the CPU cycle frequency of worker node n is f n The revenue corresponding to this type of worker node is

[0072] For a given contract, the revenue function that the task issuer receives from worker node n is U. p The calculation is as follows:

[0073]

[0074] Since there is competition among different task publishers in the federated learning system, and each task publisher strives to maximize its own interests, based on formula (7), the probability pc of a worker node selecting a task publisher is introduced. i The overall revenue function of the task publisher for that task is... The calculation is as follows:

[0075]

[0076] For worker node n that has signed the contract submitted by task publisher i, its corresponding revenue function U a The definition is as follows:

[0077]

[0078] Then the overall revenue function of the worker node The calculation is as follows:

[0079]

[0080] 3) When information asymmetry exists, the contract designed by the task issuer needs to simultaneously satisfy individual rationality constraints and incentive compatibility.

[0081] Definition 1: Individual rationality constraint: The condition for each worker node to agree to participate in a federated learning task is that the reward after completing the task is non-negative, as expressed below:

[0082]

[0083] Definition 2: Incentive Compatibility Constraint: The condition for each worker node to obtain the maximum benefit is to select a contract designed according to its own quality type, as shown below:

[0084]

[0085] The calculation of the optimal contract first considers the relaxation problem without monotonicity constraints, and then checks whether the obtained solution satisfies the monotonicity condition. Simultaneously, worker nodes consider maximizing their own revenue and use the optimal p. * c i Finally, by simplifying the constraints based on the local downward excitation constraint, the simplest optimal contract is obtained, calculated as follows:

[0086]

[0087] st

[0088]

[0089]

[0090]

[0091]

[0092] 4) The specific iterative process is as follows:

[0093] First, using the contract of task publisher A as a constant, the optimal contract of task publisher B is derived using this constant condition. Then, using the optimal contract of task publisher B as a constant, the optimal contract of task publisher A is derived again using this condition. This process is repeated until the optimal contracts of task publishers A and B converge. At this point, the optimal contract corresponding to each type of worker node can be calculated. This equilibrium condition simultaneously satisfies the individual rationality constraint and incentive compatibility constraint in each contract. That is, while ensuring non-negative returns for multiple task publishers under competitive conditions, it also ensures that each worker node can choose the optimal contract that suits its own conditions and obtain the best return.

[0094] The contract proposed in this invention is not only applicable to scenarios with multiple task publishers that better meet practical needs, but also satisfies individual rationality constraints and incentive compatibility constraints. It can balance the relationship between task publishers and worker nodes, enabling both roles in the system to obtain expected benefits, forming a virtuous cycle, and achieving the goal of incentivizing worker nodes to join the federated learning system.

[0095] (1) In the federated learning system, multiple task publishers publish training tasks and submit pre-designed contracts to the blockchain for storage.

[0096] (2) Worker nodes in the system can view different tasks and contracts submitted by multiple task publishers through blocks and choose to sign the best contract for themselves.

[0097] (3) After signing, the worker node uses the local dataset to train the model.

[0098] (4) The worker nodes upload the trained model to the blockchain for storage.

[0099] (5) The task publisher retrieves the updated model from the blockchain and aggregates all the models locally.

[0100] (6) After the federated learning task is completed, the task issuer pays the corresponding remuneration to the worker node according to the contract signed with the worker node, and saves the payment record on the blockchain.

[0101] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for determining worker training compensation in the context of multi-job competition, characterized in that, Includes the following steps: 1) In a federated learning system, multiple task publishers publish training tasks and submit pre-designed contracts to the blockchain for storage; 2) Worker nodes in the system view different tasks and contracts submitted by multiple task publishers through blocks, and choose to sign the best contract for themselves; after signing, worker nodes use local datasets to train models. 3) Worker nodes upload the trained model to the blockchain for storage; 4) The task publisher retrieves the updated model from the blockchain and aggregates all the models locally; 5) After the federated learning task is completed, the task issuer pays the corresponding remuneration to the worker nodes according to the contract signed with the worker nodes, and saves the payment record on the blockchain.

2. The method for determining worker training compensation in a multi-tasking competition environment according to claim 1, characterized in that: Step 2) specifically refers to: 1) The cost of worker nodes participating in tasks in the optimal contract includes training cost and communication cost; Each worker node has its own local dataset, and training the local model once in each iteration consumes computation time. and CPU resources ; in, The number of CPU cycles required to train a worker node with one unit of data. The size of the local dataset for the worker node; The CPU cycle frequency of the worker node; The process of worker nodes delivering the updated trained model to the task issuer is achieved through wireless communication. Each iteration requires communication time. and communication resources for: in, The size of the model parameters passed to all worker nodes after each iteration is a fixed constant. For transmission power; Channel gain; Background noise; worker nodes Total time consumed in a single global iteration process for: worker nodes Total energy consumed in a single global iteration process for: 2) In the federated learning system, the beneficiaries are divided into two roles: task issuers and worker nodes. For task publishers, they can infer the probability of each worker node belonging to a certain type based on past experience; assuming the quality of worker nodes is divided into M categories, and arranged in ascending order, it is represented as follows: Each worker node belongs to The probability of type is Then there is ; Due to the information asymmetry between the task publisher and the worker nodes, the task publisher... Design specific contract abstractions for worker nodes of different quality types. Among them, worker nodes The CPU cycle frequency is The revenue corresponding to this type of worker node is ; For a contract, the task issuer starts from the worker node. The obtained profit function for: Because there is competition among different task publishers in a federated learning system, and each task publisher strives to maximize its own interests, the probability of a worker node selecting a task publisher is introduced. Then the overall revenue function of the task publisher for: For worker nodes Signed the task issuer The submitted contract and its corresponding payoff function for: Then the overall revenue function of the worker node for: 3) When information asymmetry exists, the contract designed by the task issuer needs to simultaneously satisfy individual rationality constraints and incentive compatibility; Definition 1: Individual rationality constraint: The condition for each worker node to agree to participate in a federated learning task is that the reward after completing the task is non-negative, as expressed below: Definition 2: Incentive Compatibility Constraint: The condition for each worker node to obtain the maximum benefit is to select a contract designed according to its own quality type, as shown below: The calculation of the optimal contract first considers the relaxation problem without monotonicity constraints, and then checks whether the obtained solution satisfies the monotonicity condition; simultaneously, worker nodes consider maximizing their own revenue and use the optimal contract. Finally, by simplifying the constraints based on the local downward excitation constraints, the simplest optimal contract is obtained, calculated as follows: 4) The specific iterative process is as follows: First, let's look at the task issuer. The optimal contract for task publisher B is derived using a constant value. Then, using the optimal contract for task publisher B as a constant value, the optimal contract for task publisher A is derived again. This process is repeated until the optimal contracts for task publishers A and B converge. At this point, the optimal contract for each type of worker node can be calculated. This equilibrium condition satisfies both the individual rationality constraint and the incentive compatibility constraint in each contract. That is, while ensuring non-negative returns for multiple task publishers under competitive conditions, it also ensures that each worker node can choose the optimal contract that suits its own conditions and obtain the best return.

Images