Federal learning incentive scheme based on Nash negotiation theory

Abstract

The invention provides a federated learning incentive scheme based on a Nash negotiation theory, and aims to optimize cooperation between a data provider and a requester. By introducing an SC-NBTI protocol and combining an IPFS distributed file system, efficient task management and execution are realized. In order to improve communication efficiency and protect data privacy, a probability sparse matrix compression technology is adopted to reduce data volume and communication cost of gradient transmission. Aiming at a complex Nash negotiation solution, the invention provides a greedy client selection strategy, and in combination with a bonus payment strategy, a resolving process is simplified, and income balance of all parties participating in federal learning is ensured. The method not only improves the resource allocation efficiency, but also ensures that all participants obtain fair returns. The SC-NBTI is deployed on the block chain, and automatic management is realized by using the characteristics of transparency and non-tampering of the SC-NBTI, so that the traceability and verifiability of tasks are enhanced, and disputes and default risks are reduced. The design ensures efficient proceeding of federal learning tasks.

Description

Technical Field

[0001] The present invention relates to the field of data sharing, and in particular to federated learning, and specifically to a federated learning incentive scheme based on Nash negotiation theory. Background Art

[0002] In order to protect data privacy and solve the problem of data silos, researchers have proposed the concept of federated learning. Federated learning is an emerging machine learning paradigm that allows multiple parties to jointly train models without sharing original data, thereby avoiding the risk of data leakage. This method is implemented by training the model locally and sending the updated model parameters to a central server, thereby achieving data sharing and collaboration while protecting data privacy.

[0003] However, in real data sharing scenarios, data requesters of federated learning often have to consider how to motivate data providers to participate in federated learning. Data providers may have valuable data resources, but due to issues such as data privacy and security, they may be reluctant to share data directly. Therefore, data requesters need to develop appropriate incentive mechanisms to encourage data providers to participate in federated learning. This may include providing data exchange, sharing of model training results, or economic rewards to better promote data sharing and cooperation, thereby improving the effectiveness and efficiency of federated learning.

[0004] Kang

[57] proposed an incentive scheme based on a reputation system, which uses a triple weighted subjective logic model (TWSL) to accurately manage the reputation of vehicles. Vehicles can select the best data provider based on the reputation value calculated from their interaction history with other vehicles, thereby ensuring the acquisition of high-quality shared data. Zeng

[34] designed a multi-dimensional auction incentive scheme for federated learning in a mobile edge computing environment. By considering multiple dimensions such as data quality, computing resources, and network conditions, the scheme effectively stimulated the enthusiasm of participants, while optimizing resource allocation and improving system efficiency. Deng

[58] focused on the quality perception problem in federated learning and ensured the improvement of data quality by introducing precise user incentives and model aggregation mechanisms. At the same time, by rewarding high-quality data contributors, more users are encouraged to participate, thereby enhancing the overall performance and fairness of the model. Zhu

[28] used Stackelberg game to study the incentive mechanism of federated learning. This method achieves the purpose of improving model accuracy and resource utilization efficiency by real-time adjustment of the energy consumption of available clients in federated learning. Zhan[36,59] designed an efficient incentive mechanism for edge learning through deep reinforcement learning methods. This mechanism can dynamically adjust the incentive strategy to adapt to the ever-changing network environment and user behavior, thereby optimizing resource allocation and improving learning efficiency. Lim

[35] used alliance game theory to reward federated learning participants based on marginal contributions, and simulated the dynamic decision-making process through a merge and split algorithm, achieving the goal of optimizing benefits and forming a stable alliance structure under changing network conditions. Ding

[60] proposed a contract theory incentive mechanism that considers multidimensional information under incomplete information. However, according to analytical research and experimental results, when the local data set is independent and identically distributed, the contract theory incentive mechanism can work well under both weak incomplete information and complete information. In a non-independent and identically distributed (Non-IID) environment, the contract theory incentive mechanism performs poorly.

[0005] By analyzing the above studies, we can find that the existing incentive mechanisms have not fully considered the problems of communication efficiency and Non-IID data scenarios. At the same time, they are often incentive mechanisms designed in a non-cooperative game manner. Non-cooperative games mean that they can often only represent the structured interactions of the parties involved in federated learning, and are not friendly to unstructured interaction settings. To a certain extent, the entry threshold of federated learning has been raised, resulting in the loss of high-quality data providers. In addition, the utility functions in existing studies generally include quadratic or logarithmic costs, which usually cannot guarantee the uniqueness of Nash equilibrium. Summary of the invention

[0006] The purpose of the present invention is to propose a federated learning incentive scheme based on Nash negotiation theory to solve the problem of insufficient incentives in federated learning scenarios. Through the scheme proposed by the present invention, more users can be encouraged to participate in federated learning tasks, thereby improving the accuracy of the model.

[0007] The object of the present invention is achieved in this way:

[0008] This paper first transforms the incentive problem into the problem of finding the NBS solution of the game parties by modeling the data requester and the data provider. Then, through the client selection strategy proposed in this chapter, the calculation of the complex NBS solution is simplified to the approximation of the NBS solution, which greatly improves the calculation speed and ensures the efficiency of federated learning.

[0009] The specific method is as follows:

[0010] A federated learning incentive model based on Nash negotiation theory includes the following steps:

[0011] Step 1: The data requester creates a federated learning task and broadcasts the data request information to the relevant community. The data request information includes task ID, task budget, requested data size, etc.

[0012] Step 2: The data provider uploads the current round model to IPFS, gives the IPFS address to SC-NBTI for management, and pays the model reward for this round;

[0013] Step 3: The members who receive the message estimate the computing cost and communication cost according to their own situation, and give a minimum bid based on their total cost, and upload the bid and local data size to the sharing platform;

[0014] Step 4: The NBTI smart contract selects a set of data providers according to the greedy strategy and pushes the IPFS address of this round to the selected data providers;

[0015] Step 5: The selected data provider downloads the model parameters of this round from IPFS;

[0016] Step 6: The data provider uses local data to train the model and performs sparse processing on the training results.

[0017] Step 7. The data provider uploads the sparsely processed gradient update to IPFS and puts the IPFS address on the blockchain.

[0018] Step 8: The data requester retrieves the current round of model gradient updates uploaded by each data provider from IPFS.

[0019] Step 9: The data requester aggregates the model based on the current round of gradient updates

[0020] Step 10: The data requester calculates the current round of model revenue

[0021] Step 11: The data requester uploads the model revenue of this round to NBTI

[0022] Step 12: NBTI smart contract divides the bonus equally according to the current round of revenue

[0023] Step 13: Repeat steps (2)-(12) until the model converges or reaches the specified number of iterations. Furthermore, if data provider m decides to participate in the tth round of federated learning, the benefit for m is the reward he gets from participating in this round of federated learning. It's obvious Should be greater than or equal to 0. At the same time, he will incur corresponding costs in represents the computational overhead, Represents the communication overhead.

[0024] We define a Boolean variable Used to indicate whether data provider m participates in round t of federated learning: If so, otherwise For any client m in round t, its profit It depends on the reward he gets and the cost he spends, which can be expressed as:

[0025]

[0026] Furthermore, the benefit of the data requester in step 1 can be described as:

[0027]

[0028] Furthermore, the negotiation in step 3 can be described as the following optimization problem:

[0029]

[0030] Furthermore, the above optimization can be described as:

[0031]

[0032] Furthermore, the greedy strategy in step 4 mainly relies on the probability selection strategy, and the probability of data provider m being selected is:

[0033]

[0034] Furthermore, by integrating the above conditions, the optimization problem can be described as:

[0035]

[0036] Furthermore, by introducing the Lagrange multipliers θ and To constrain and

[0037]

[0038] Further, solving the above problem, we can get:

[0039]

[0040] Based on this closed-form solution, the optimal payment price can be calculated for the data providers selected in the client selection phase:

[0041]

[0042] The positive effects of the present invention are:

[0043] This invention introduces Nash negotiation theory into the incentive mechanism, constructs a cooperative game incentive model NBTI, lowers the entry threshold of federated learning, and expands the participation pool of federated learning. Subsequently, in order to simplify the complex calculation of Nash negotiation solutions, this invention proposes a client selection scheme based on a greedy strategy and challenges the bonus payment strategy based on this scheme, which reduces the complexity of calculation and ensures the smooth progress of federated learning. The results show that the SC-NBTI model proposed in this invention has great advantages in terms of distribution fairness and model accuracy compared with existing incentive research. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flow chart of the federated learning incentive scheme based on Nash negotiation theory.

[0045] Figure 2 This is the UML class diagram of the smart contract based on the above incentive scheme.

[0046] Figure 3 This is a graph showing the change in accuracy of each solution with the number of rounds.

[0047] Figure 4 It is a graph showing the change of loss value of each scheme with the rounds.

[0048] Figure 5 It is the distribution of bonuses for each plan. DETAILED DESCRIPTION

[0049] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0050] like Figure 1As shown in the figure, the present invention first transforms the incentive problem into the problem of finding the NBS solution of the game parties by modeling the data requester and the data provider. Then, through the client selection strategy proposed in this chapter, the calculation of the complex NBS solution is simplified to the approximation of the NBS solution, which greatly improves the calculation speed and ensures the efficiency of federated learning. Specifically, it includes the following steps:

[0051] Step 1: The data requester creates a federated learning task and broadcasts the data request information to the relevant community. The data request information includes task ID, task budget, requested data size, etc.

[0052] Step 2: The data provider uploads the current round model to IPFS, gives the IPFS address to SC-NBTI for management, and pays the model reward for this round;

[0053] Step 3: The members who receive the message estimate the computing cost and communication cost according to their own situation, and give a minimum bid based on their total cost, and upload the bid and local data size to the sharing platform;

[0054] Step 4: The NBTI smart contract selects a set of data providers according to the greedy strategy and pushes the IPFS address of this round to the selected data providers;

[0055] Step 5: The selected data provider downloads the model parameters of this round from IPFS;

[0056] Step 6: The data provider uses local data to train the model and performs sparse processing on the training results.

[0057] Step 7. The data provider uploads the sparsely processed gradient update to IPFS and puts the IPFS address on the blockchain.

[0058] Step 8: The data requester retrieves the current round of model gradient updates uploaded by each data provider from IPFS.

[0059] Step 9: The data requester aggregates the model based on the current round of gradient updates

[0060] Step 10: The data requester calculates the current round of model revenue

[0061] Step 11: The data requester uploads the model revenue of this round to NBTI

[0062] Step 12: NBTI smart contract divides the bonus equally according to the current round of revenue

[0063] Step 13: Repeat steps (2)-(12) until the model converges or reaches the specified number of iterations. Furthermore, if data provider m decides to participate in the tth round of federated learning, the benefit for m is the reward he gets from participating in this round of federated learning. It's obvious Should be greater than or equal to 0. At the same time, he will incur corresponding costs in represents the computational overhead, Represents the communication overhead.

[0064] We define a Boolean variable Used to indicate whether data provider m participates in round t of federated learning: If so, otherwise For any client m in round t, its profit It depends on the reward he gets and the cost he spends, which can be expressed as:

[0065]

[0066] Furthermore, the benefit of the data requester in step 1 can be described as:

[0067]

[0068] Furthermore, the negotiation in step 3 can be described as the following optimization problem:

[0069]

[0070] Furthermore, the above optimization can be described as:

[0071]

[0072] Furthermore, the greedy strategy in step 4 mainly relies on the probability selection strategy, and the probability of data provider m being selected is:

[0073]

[0074] Furthermore, by integrating the above conditions, the optimization problem can be described as:

[0075]

[0076]

[0077] Furthermore, by introducing the Lagrange multipliers θ and To constrain and

[0078]

[0079] Further, solving the above problem, we can get:

[0080]

[0081] Based on this closed-form solution, the optimal payment price can be calculated for the data providers selected in the client selection phase:

[0082]

Claims

1. A federated learning incentive scheme based on Nash negotiation theory, characterized in that: The method comprises the following steps: Step 1: The data requester creates a federated learning task and broadcasts the data request information to the relevant community. The data request information includes task ID, task budget, requested data size, etc. Step 2: The data provider uploads the current round model to IPFS, gives the IPFS address to SC-NBTI for management, and pays the model reward for this round; Step 3: The members who receive the message estimate the computing cost and communication cost according to their own situation, give a minimum bid based on their total cost, and upload the bid and local data size to the sharing platform; Step 4: The NBTI smart contract selects a set of data providers according to the greedy strategy and pushes the IPFS address of this round to the selected data providers; Step 5: The selected data provider downloads the model parameters of this round from IPFS; Step 6: The data provider uses local data to train the model and performs sparse processing on the training results; Step 7: The data provider uploads the sparsely processed gradient update to IPFS and puts the IPFS address on the blockchain; Step 8: The data requester retrieves the model gradient update uploaded by each data provider from IPFS; Step 9: The data requester aggregates the model based on the current round of gradient updates; Step 10: The data requester calculates the current round of model revenue; Step 11: The data requester uploads the model revenue of this round to NBTI; Step 12: The NBTI smart contract divides the bonus equally according to the current round of revenue; Step 13, repeat steps (2)-(12) until the model converges or reaches a specified number of iterations; Furthermore, if data provider m decides to participate in round t of federated learning, the benefit for m is the reward he gets from participating in this round of federated learning. It's obvious Should be greater than or equal to 0; at the same time, he will incur corresponding costs in represents the computational overhead, represents the communication overhead; We define a Boolean variable Used to indicate whether data provider m participates in round t of federated learning: If so, otherwise For any client m in round t, its profit It depends on the reward he gets and the cost he spends, which can be expressed as: Furthermore, the benefit of the data requester in step 1 can be described as: Furthermore, the negotiation in step 3 can be described as the following optimization problem: Furthermore, the above optimization can be described as: Furthermore, the greedy strategy in step 4 mainly relies on the probability selection strategy, and the probability of data provider m being selected is: Furthermore, by integrating the above conditions, the optimization problem can be described as: Furthermore, by introducing the Lagrange multipliers θ and To constrain and Further, solving the above problem, we can get: Based on this closed-form solution, the optimal payment price can be calculated for the data providers selected in the client selection phase.

Images