Federal learning incentive mechanism method based on block chain

Abstract

The invention discloses a block chain-based federal learning incentive mechanism method, which comprises the following steps of: setting a data quality and a credibility threshold of a participant through evaluation, protecting the benefit of the existing participant, and when the data quality and the credibility of the participant are simultaneously greater than the data quality threshold and the credibility threshold, allowing the participant to join in federal learning; a punishment mechanism is added to effectively guarantee the benefits of existing participants; through the block chain technology, the incentive process is open and transparent, and incentive fairness and enthusiasm of participants are ensured. And stage factors of model training are added into the excitation mechanism based on the model effect, so that the influence of effect improvement difficulty in different stages of model training is reduced. And the participants can obtain corresponding return due to long-time waiting, so that the enthusiasm of the participants is ensured. According to the invention, benefits of existing participants can be effectively ensured; the incentive fairness and the enthusiasm of participants are ensured; and the influence of effect improvement difficulty in different stages of model training is reduced.

Description

technical field [0001] The present invention relates to the field of federated learning for privacy computing, more specifically, a method for an incentive mechanism for federated learning based on blockchain. Background technique [0002] With the implementation of the "Data Security Law", the domestic data circulation and trading industry is expected to be further standardized, which will give birth to the development of new technology industries such as "privacy computing". Common technical paths to realize private computing include trusted computing, federated learning, secure multi-party computing, etc. This patent mainly explores the incentive mechanism method of federated learning in privacy computing. [0003] Federated learning is not only a technical standard, but also a business model. When people realize the impact of big data, their first thought is to aggregate the data together, calculate the model by remote processor, and then download the result for furthe...

AI Technical Summary

Problems solved by technology

[0009] Since there is no entry threshold for participants, customers with insufficient data resources will participate in federated learning, resulting in reduced model performance; malicious customers participating in federated learning will bring some security risks

Benefits of existing parties cannot be guaranteed

[0010] 2) Lack of punishment mechanism

[0011] Customers may cause model performance degradation or other losses due to subjective or objective reasons during the federated learning process. There is currently no corresponding punishment mechanism for this situation, and the interests of existing participants cannot be effectively guaranteed

[0012] 3) The incentive process is not transparent to all participants

[0013] The incentive process is opaque to all participants, which may lead to unfair incentives and cause participants to lack trust in the alliance

[0014] 4) The incentive mechanism based on the model effect does not consider the difficulty of improving the model effect at different stages of model training;

In the later stage of model training, the loss of the objective function is generally small, and the gradient of the model parameters is relatively stable, which makes it relatively difficult to improve the model effect.

As a result, relatively high incentives can be obtained even if the data quality and quantity are average in the early stage; in the later stage, the data quality is high and the amount of data is large, but normal incentives cannot be obtained

[0016] 5) Did not consider the customer's waiting cost

Customers did not get corresponding rewards due to the long waiting cost, which affected the enthusiasm of customers to participate

Images