A privacy-preserving incentive mechanism training method based on reverse auction in federated learning

By adopting the privacy protection incentive mechanism training method of reverse auction in federated learning, participants use generators to fit the output of the extractor and filter through reverse auctions, solving the problems of incentive mechanism and privacy protection, realizing data privacy Effective protection and motivation of participants.

CN114493808BActive Publication Date: 2024-11-26SUN YAT SEN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210109376.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-28
Publication Date
2024-11-26
Estimated Expiration
2042-01-28

AI Technical Summary

Technical Problem

The lack of effective incentives in federated learning leads to lack of motivation among participants, and existing privacy protection technologies have problems with the risk of privacy leakage and performance degradation.

Method used

Using the training method of privacy protection incentive mechanism based on reverse auction, the federated learning participants do not send the extractor directly, but use the generator to fit the extractor's output, and filter and motivate participants through reverse auction. The central server cannot Reverse the original data to satisfy personal rationality, authenticity and budget feasibility.

Benefits of technology

It effectively protects data privacy, improves the incentive mechanism for participants, ensures personal rationality and budget feasibility of the training process, and avoids the risk of privacy leakage.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114493808B_ABST
    Figure CN114493808B_ABST
Patent Text Reader

Abstract

In view of the limitations of the prior art, the present invention proposes a privacy-preserving incentive mechanism training method based on reverse auction in federated learning. This method does not send the extractor to the central server, but only uses the generator to fit the output of the extractor, so that the model uploaded by the federated learning participants to the central server is not directly trained with the original data. Therefore, the central server cannot infer the original data, thereby protecting data privacy. In addition, the present invention incentivizes the federated learning participants in the form of reverse auction, which can meet the personal rationality, authenticity and budget feasibility in the training process.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence; specifically, to a federated learning technology; more specifically, to a reverse auction-based privacy protection incentive mechanism training method in federated learning. Background Art

[0002] The concept of federated learning was first proposed by Google and used in the intelligent software GBoard, with the purpose of predicting the next word in the virtual keyboard of Android smartphones. The basic idea is to deploy a recursive neural network language model in each user's smartphone. In each round of model training, the user sends the model gradient based on the local data set to the server, thereby avoiding uploading the user's sensitive data. The two biggest problems faced by federated learning in real application are:

[0003] Incentive mechanism. Most of the work on federated learning today is focused on designing higher-performance algorithms. The lack of an incentive mechanism has seriously hindered the widespread application of federated learning. If there is no fair and reasonable reward distribution and incentive mechanism, and only the task scheduling mechanism of the central server is relied upon to force the division of labor among federated learning participants, then participants who are centered on self-interest will lack the motivation to actively participate in federated learning and are likely to upload some low-quality or meaningless model gradients. The current incentive mechanism includes: ① Building a reputation model from the historical interaction records of federated learning participants, combining reputation with contract theory. ② Using an impact factor function to measure sample importance to allocate rewards.

[0004] Privacy protection. The Chinese invention application with a publication date of 2021.05.25: A context-aware recommendation system and method based on federated learning, which applies federated learning technology to the recommendation system and combines contextual information in the recommendation system, in order to achieve a high recommendation accuracy while protecting user privacy. Although the federated learning framework provides certain protection for user privacy, the server may still indirectly infer the original data from the model structure and model gradient exposed by the participants, resulting in the leakage of private information. Existing federated learning methods mainly use three technical means to solve this problem: ① Secure multi-party computing (SMC), but it is vulnerable to collusion inference attacks; ② Differential privacy (DP), but it reduces model performance; ③ Homomorphic encryption (HE), but the encryption and decryption process is very time-consuming and computing power-consuming. Therefore, the existing technology still has certain limitations. Summary of the invention

[0005] In view of the limitations of the existing technology, the present invention proposes a privacy protection incentive mechanism training method and system based on reverse auction in federated learning. The technical solution adopted by the present invention is:

[0006] A reverse auction-based privacy-preserving incentive mechanism training method in federated learning. For a central server including a global classifier and a number of federated learning participants including extractors, classifiers, discriminators, and generators, in each round of federated learning:

[0007] The federated learning participants do not send the extractor to the central server, but use their own generators to fit the output of the extractor for the central server to train;

[0008] The central server screens the classifiers of each federated learning participant according to the bid price and generator provided by each federated learning participant in combination with a reverse auction method; and updates the global classifier by fusing the screening results.

[0009] Compared with the prior art, the present invention does not send the extractor to the central server, but only uses the generator to fit the output of the extractor, so that the model uploaded by the federated learning participants to the central server is not directly trained by the original data. Therefore, the central server cannot infer the original data, thereby protecting data privacy. In addition, the present invention incentivizes the federated learning participants in the form of a reverse auction, which can meet the personal rationality, authenticity and budget feasibility of the training process.

[0010] As a preferred solution, in each round of federated learning, the federated learning participants receive global classifier updates from the central server to replace local classifiers, and then train the extractor and classifier using training data x and class label y through the following steps:

[0011] S11, input the training data x into the extractor to obtain the real feature f real ;

[0012] S12, the real feature f real Input into the classifier to obtain the predicted class label y pred ;

[0013] S13, according to the predicted class label y pred With the class label y, we get the classification task loss value L task ;

[0014] S14, to minimize the task loss value L task As the goal, the back-propagation algorithm is used to update the parameters of the extractor and the classifier.

[0015] Furthermore, in each round of federated learning, after completing the training of the extractor and the classifier, the federated learning participants train the discriminator and the generator through the following steps:

[0016] S15, input the class label y and the preset random noise z into the generator to obtain the pseudo feature f fake ;

[0017] S16, the real feature f real and the pseudo-feature f fake Input into the discriminator to obtain the predicted domain label d pred ;

[0018] S17, according to the predicted domain label d pred , get the domain task loss value L dom ;

[0019] S18, using the back propagation algorithm to minimize the domain task loss value L dom Update the parameters of the discriminator for the goal to maximize the domain task loss value L dom Update the parameters of the generator for the target.

[0020] Furthermore, in each round of federated learning, after completing the training of the discriminator and the generator, the federated learning participant will send the generator trained in this round of federated learning and the preset bid price to the central server to participate in the bidding: if the bid is successful, the classifier trained in this round of federated learning will be sent to the central server for fusion, ending this round of federated learning; otherwise, this round of federated learning will be ended directly.

[0021] As a preferred solution, in each round of federated learning, the central server is trained in the following manner:

[0022] S21, obtain the generator G provided by each federated learning participant i i and the bid price i ;

[0023] S22, traverse the class label y, and input the class label y and random noise z into each generator G respectively i In the generator G i The feature mean f corresponding to the class label y i ;

[0024] S23, using a clustering algorithm to obtain the feature mean f i The distance d to the center of the class i ,by As the generator G i the value of

[0025] S24, according to The value of is used to sort the federated learning participants i from small to large, and the first k federated learning participants are selected as successful bidders;

[0026] S25, obtaining classifiers from the successful bidders, fusing them, and updating the global classifier.

[0027] Furthermore, the step S23 uses the unsupervised clustering algorithm K-means to obtain the feature mean f i The distance d to the center of the class i .

[0028] Furthermore, the k value in step S24 is obtained according to the following formula:

[0029]

[0030] Among them, B is the budget for each round of federated learning.

[0031] Furthermore, the step S25 uses a federated average aggregation algorithm FedAvg to fuse the classifiers.

[0032] The present invention also provides the following contents:

[0033] A computer device comprises a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor. When the computer program is executed by the processor, the execution content of the federated learning participants in the aforementioned reverse auction-based privacy protection incentive mechanism training method in federated learning is implemented.

[0034] A server device includes a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor. When the computer program is executed by the processor, the execution content of the central server in the aforementioned reverse auction-based privacy protection incentive mechanism training method in federated learning is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic diagram of the principle of the reverse auction-based privacy protection incentive mechanism training method in federated learning provided by the present invention;

[0036] Figure 2 A training flow chart of an extractor and a classifier in a training method for a privacy-preserving incentive mechanism based on reverse auctions in federated learning provided by the present invention;

[0037] Figure 3 A training flow chart of a discriminator and a generator in a training method for a privacy-preserving incentive mechanism based on reverse auctions in federated learning provided by the present invention;

[0038] Figure 4 A training flow chart of a central server in a training method for a privacy-preserving incentive mechanism based on reverse auction in federated learning provided by the present invention;

[0039] Figure 5 A schematic diagram of the training principle of the central server in the reverse auction-based privacy protection incentive mechanism training method in federated learning provided by the present invention. DETAILED DESCRIPTION

[0040] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0041] It should be clear that the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present application.

[0042] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the embodiments of the present application. The singular forms of "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0043] When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the attached claims. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and do not have to be used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0044] In addition, in the description of this application, unless otherwise specified, "multiple" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0045] In order to solve the limitations of the prior art, this embodiment provides a technical solution, and the technical solution of the present invention is further described below in conjunction with the drawings and embodiments.

[0046] Example 1

[0047] See also Figure 1 , a privacy-preserving incentive mechanism training method based on reverse auction in federated learning, for a central server including a global classifier and several federated learning participants including extractors, classifiers, discriminators and generators, in each round of federated learning:

[0048] The federated learning participants do not send the extractor to the central server, but use their own generators to fit the output of the extractor for the central server to train;

[0049] The central server screens the classifiers of each federated learning participant according to the bid price and generator provided by each federated learning participant in combination with a reverse auction method; and updates the global classifier by fusing the screening results.

[0050] Compared with the prior art, in the training method of the privacy-preserving incentive mechanism based on reverse auction in federated learning provided by the present invention, the extractor will not be sent to the central server, and only the generator is used to fit the output of the extractor, so that the model uploaded by the federated learning participants to the central server is not directly trained by the original data. Therefore, the central server cannot infer the original data, thereby protecting data privacy; in addition, the present invention incentivizes federated learning participants in the form of reverse auction, which can meet the personal rationality, authenticity and budget feasibility in the training process.

[0051] A reverse auction is an auction method with a single buyer and multiple sellers. The buyer proposes the goods or services he wants to obtain, and then the sellers make bids for their goods or services. The buyer decides the seller with the best price-performance ratio. Therefore, a reverse auction is very suitable for use in a single-server and multi-client scenario such as federated learning.

[0052] Specifically, individual rationality means that the remuneration obtained by participants will not be lower than their own bid price; authenticity means that participants cannot obtain more remuneration by bidding higher or lower than their own costs, that is, bidding according to one's own cost is the dominant strategy; budget feasibility means that the total remuneration ultimately paid by the publisher will not exceed its own budget.

[0053] More specifically, in order to fit the output of the extractor, as a preferred embodiment, see Figure 2 In each round of federated learning, the federated learning participants receive global classifier updates from the central server to replace local classifiers, and then train the extractor and classifier using training data x and class label y through the following steps:

[0054] S11, input the training data x into the extractor to obtain the real feature f real ;

[0055] S12, the real feature f real Input into the classifier to obtain the predicted class label y pred ;

[0056] S13, according to the predicted class label y pred With the class label y, we get the classification task loss value L task ;

[0057] S14, to minimize the task loss value L task As the goal, the back-propagation algorithm is used to update the parameters of the extractor and the classifier.

[0058] For further information, see Figure 3 In each round of federated learning, after completing the training of the extractor and the classifier, the federated learning participants train the discriminator and the generator through the following steps:

[0059] S15, input the class label y and the preset random noise z into the generator to obtain the pseudo feature f fake ;

[0060] S16, the real feature f real and the pseudo-feature f fake Input into the discriminator to obtain the predicted domain label d pred ;

[0061] S17, according to the predicted domain label d pred , get the domain task loss value L dom ;

[0062] S18, using the back propagation algorithm to minimize the domain task loss value L dom Update the parameters of the discriminator for the goal to maximize the domain task loss value L dom Update the parameters of the generator for the target.

[0063] Specifically, in the prediction domain label d pred In the example, 1 represents the true feature f real The domain label, 0 means generating feature f fake The domain label.

[0064] Furthermore, in each round of federated learning, after completing the training of the discriminator and the generator, the federated learning participant will send the generator trained in this round of federated learning and the preset bid price to the central server to participate in the bidding: if the bid is successful, the classifier trained in this round of federated learning will be sent to the central server for fusion, ending this round of federated learning; otherwise, this round of federated learning will be ended directly.

[0065] Specifically, after the end of this round of federated learning, if the next round of federated learning is still needed, before the next round of federated learning starts, the federated learning participant will re-receive the global classifier updated by the fusion of this round of federated learning from the central server to update and replace the local classifier.

[0066] As a preferred embodiment, please refer to Figure 4 as well as Figure 5 , in each round of federated learning, the central server is trained in the following way:

[0067] S21, obtain the generator G provided by each federated learning participant i i and the bid price i ;

[0068] S22, traverse the class label y, and input the class label y and random noise z into each generator G respectively i In the generator G i The feature mean f corresponding to the class label y i ;

[0069] S23, using a clustering algorithm to obtain the feature mean f i The distance d to the center of the class i ,by As the generator G i the value of

[0070] S24, according to The value of is used to sort the federated learning participants i from small to large, and the first k federated learning participants are selected as successful bidders;

[0071] S25, obtaining classifiers from the successful bidders, fusing them, and updating the global classifier.

[0072] Specifically, for each successful bidder, the central server will pay

[0073] Furthermore, the step S23 uses the unsupervised clustering algorithm K-means to obtain the feature mean f i The distance d to the center of the class i .

[0074] Furthermore, the k value in step S24 is obtained according to the following formula:

[0075]

[0076] Among them, B is the budget for each round of federated learning.

[0077] Furthermore, the step S25 uses a federated average aggregation algorithm FedAvg to fuse the classifiers.

[0078] Example 2

[0079] A computer device comprises a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor. When the computer program is executed by the processor, the execution content of the federated learning participants in the reverse auction-based privacy protection incentive mechanism training method in federated learning as described in Example 1 is implemented.

[0080] Specifically, the computer device referred to in this embodiment may be, but is not limited to, a personal computer, a laptop computer, a mobile phone, a tablet computer, and other devices.

[0081] Example 3

[0082] A server device comprises a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor. When the computer program is executed by the processor, the execution content of the central server in the reverse auction-based privacy protection incentive mechanism training method in federated learning as described in Example 1 is implemented.

[0083] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A privacy-preserving incentive mechanism training method based on reverse auction in federated learning, for a central server including a global classifier and several federated learning participants including extractors, classifiers, discriminators and generators, It is characterized in that In each round of federated learning: The federated learning participants do not send the extractor to the central server, but use their own generators to fit the output of the extractor for the central server to train; The central server screens the classifiers of each federated learning participant according to the bid price and generator provided by each federated learning participant in combination with a reverse auction method; and updates the global classifier by fusing the screening results; In each round of federated learning, the federated learning participants receive global classifier updates from the central server to replace local classifiers with training data. x and class labels y , the extractor and classifier are trained by the following steps: S11, the training data x Input into the extractor to obtain the real features ; S12, the real feature Input into the classifier to obtain the predicted class label ; S13, based on the predicted class label With class label y , get the classification task loss value ; S14, to minimize the task loss value As a target, the back propagation algorithm is used to update the parameters of the extractor and the classifier; In each round of federated learning, after completing the training of the extractor and the classifier, the federated learning participants train the discriminator and the generator through the following steps: S15, label the class y And preset random noise z Input into the generator to obtain pseudo features ; S16, the real feature And pseudo features Input into the discriminator to obtain the predicted domain label ; S17, based on the predicted domain label , get the domain task loss value ; S18, using the back propagation algorithm to minimize the domain task loss value Update the parameters of the discriminator for the goal to maximize the domain task loss value updating parameters of the generator for the target; In each round of federated learning, the central server is trained in the following way: S21, obtain each federated learning participant Generators provided and the bid price ; S22, traversing the class labels y , the class label y and random noise z Input to each generator separately In the generator Corresponding to the class label y The characteristic mean ; S23, using a clustering algorithm to obtain the feature mean Distance to the center of the class ,by As a generator the value of S24, according to The values ​​of the federated learning participants are as follows: Sort by k Federated learning participants are screened as successful bidders; S25, obtaining classifiers from the successful bidders, fusing them, and updating the global classifier.

2. According to the reverse auction-based privacy protection incentive mechanism training method in federated learning according to claim 1, It is characterized in that In each round of federated learning, after completing the training of the discriminator and the generator, the federated learning participant will send the generator trained in this round of federated learning and the preset bid price to the central server to participate in the bidding: if the bid is successful, the classifier trained in this round of federated learning will be sent to the central server for fusion, ending this round of federated learning; otherwise, this round of federated learning will be ended directly.

3. According to the reverse auction-based privacy protection incentive mechanism training method in federated learning according to claim 1, It is characterized in that The step S23 uses the unsupervised clustering algorithm K-means to obtain the feature mean Distance to the center of the class .

4. According to the reverse auction-based privacy protection incentive mechanism training method in federated learning according to claim 3, It is characterized in that In step S24 k The value is obtained according to the following formula: ; in, B Budget for each round of federal learning.

5. According to the privacy protection incentive mechanism training method based on reverse auction in federated learning according to claim 4, It is characterized in that The step S25 uses the federated average aggregation algorithm FedAvg to fuse the classifiers.

6. A computer device, Features: It includes a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor. When the computer program is executed by the processor, it implements the execution content of the federated learning participants in the reverse auction-based privacy protection incentive mechanism training method in federated learning as described in any one of claims 1 to 5.

7. A server device, Features: It includes a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor. When the computer program is executed by the processor, it implements the execution content of the central server in the privacy protection incentive mechanism training method based on reverse auction in federated learning as described in any one of claims 1 to 5.