Federal learning incentive method, device and system based on contribution degree

By calculating and signing rewards, qualified participants are screened and aggregation gradient ciphertexts are encrypted, which solves the problem of lack of fair incentives in federated learning and improves the enthusiasm of participants and the quality of the model.

CN119047601AInactive Publication Date: 2024-11-29CETC BIGDATA RES INST CO LTD +1

Patent Information

Application Number
CN202411545606.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-11-01
Publication Date
2024-11-29
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing federated learning lacks a fair and reasonable incentive mechanism, and it is difficult to measure the contribution of participating nodes with different data volumes, different data quality, and different data distribution, resulting in low enthusiasm among high-quality data participants, affecting the federated learning effect and system stability.

Method used

By receiving the gradient cipher text of the participants, decryption verification, calculate the contribution of each participant, generate a reward and sign and send it to the reward center, determine the qualified participants, generate the next round of participants, and transmit the aggregated gradient cipher through the encryption algorithm. Article, ensure that only qualified participants can decrypt and participate in the next round of training.

Benefits of technology

It has enhanced the enthusiasm of participants, improved the efficiency of federal learning and model quality, ensured the quality of gradient data, and achieved fair and reasonable reward distribution.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119047601A_ABST
    Figure CN119047601A_ABST
Patent Text Reader

Abstract

The invention provides a contribution degree-based federal learning excitation method, device and system. The method comprises the following steps of: receiving a gradient ciphertext sent by a participant; after gradient ciphertexts sent by the participants in a preset number or within a preset time are received, performing decryption verification on the gradient ciphertexts to obtain gradient data; calculating the rewards of the participants, signing the rewards, and sending the signed rewards to a reward center, so that the reward center sends the rewards to the accounts of the corresponding participants; determining qualified participants, and generating a list of participants participating in federated learning in the next round; generating an aggregation gradient according to the gradient data of the qualified participants, and encrypting the aggregation gradient to obtain an aggregation gradient ciphertext; and sending the aggregation gradient ciphertext to participants in the participant list, so that the participants carry out the next round of iterative training. By using the scheme of the invention, each participant can obtain corresponding excitation, the enthusiasm of the participant is improved, and the federal learning efficiency is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of federated learning, and in particular to a contribution-based federated learning incentive method, device and system. Background Art

[0002] Federated learning (FL), as a distributed machine learning method, has attracted the attention of many researchers in recent years and has been widely used in various fields, such as smart healthcare, smart cities, finance, and smart industry.

[0003] In federated learning, data is retained in the local physical domain, and participating nodes use their own data to train local models separately and jointly build a joint model. The data owner's right to hold data resources and the right to use data processing are not infringed, and the transfer and entrustment of the right to operate data products can generate income, which buffers the contradiction between data sharing and privacy security, provides a new paradigm for breaking through "data islands", and also brings new solutions to data-driven industries.

[0004] Federated learning systems usually assume that all participants will honestly and actively participate in federated learning. However, since different participants often have different amounts of data and different qualities of data, and different participants contribute differently to the final global model, in a fair incentive environment, participants with high amounts of data and high data quality have low participation enthusiasm, and the assumption is difficult to hold. In particular, in non-independent and identically distributed data scenarios, the contribution of participants is more difficult to measure. In the process of local model training, there is consumption of various resources such as computing power and bandwidth. If there is a lack of a fair and reasonable incentive mechanism, it will be difficult to guarantee the enthusiasm of participants, especially those with high-quality data, to participate in federated learning tasks, which will seriously affect the effect of federated learning and the stability of the system. Summary of the invention

[0005] The present invention provides a contribution-based federated learning incentive method, device and system to solve the problem that the existing federated learning lacks a fair and reasonable incentive mechanism and is difficult to measure the federated learning contribution of participating nodes with different data amounts, different data qualities and different data distributions.

[0006] To this end, the present invention provides the following technical solutions: In one aspect, the present invention provides a contribution-based federated learning incentive method, the method comprising: Receive the gradient ciphertext sent by the participating party; After receiving the gradient ciphertext sent by the participants within a preset number or time, decrypt and verify the gradient ciphertext to obtain gradient data; Calculate the rewards for each participant, sign the rewards and send them to the reward center, so that the reward center can send the rewards to the corresponding participant accounts; Identify qualified participants and generate a list of participants for the next round of federated learning; Generate an aggregate gradient according to the gradient data of the qualified participant, and encrypt the aggregate gradient to obtain an aggregate gradient ciphertext; The aggregated gradient ciphertext is sent to the participants in the participant list so that the participants can perform the next round of iterative training.

[0007] Optionally, the calculating of the reward of each participant includes: Calculate the contribution of each participant; The rewards of each participant are calculated based on the contribution of each participant.

[0008] Optionally, calculating the contribution of each participant includes: calculating the contribution of each participant through marginal loss.

[0009] Optionally, the calculation of the reward of each participant according to the contribution of each participant includes: The weight of each participant in this round is calculated based on the contribution of each participant; The rewards for each participant are calculated based on their weight and budget for this round.

[0010] Optionally, determining the qualified participants includes: Determine the consistency with the previous round of aggregation model based on the gradient data of each participant; Qualified parties are determined based on the consistency.

[0011] Optionally, determining the consistency with the last round of aggregation model according to the gradient data of each participant includes: calculating the error between the gradient data of each participant and the last round of aggregation model.

[0012] Optionally, the qualified participants include: high-quality participants; Determining a high-quality participant according to the consistency includes: if the error is less than a first set value, determining the participant as a high-quality participant.

[0013] Optionally, the qualified participants include: participants other than the inferior participants; The inferior participants determined based on the consistency include: Pre-set the label corresponding to each participant and set the initial value of the label to 0; If the error corresponding to the participant is greater than or equal to a second set value, the tag value of the participant is increased by 1; If the tag value of the participant is greater than a set threshold, the participant is regarded as a low-quality participant.

[0014] Optionally, the method further comprises: The participant performs iterative training based on the received global model or aggregated gradient ciphertext in combination with local data to obtain gradient data; Encrypting the gradient data using a set encryption algorithm to generate a gradient ciphertext; Upload the gradient ciphertext.

[0015] Optionally, the encryption algorithm includes any one of the following: a symmetric encryption algorithm and an asymmetric encryption algorithm.

[0016] On the other hand, the present invention also provides a contribution-based federated learning incentive device, the device comprising: a receiving module, a data processing module, a reward calculation module, a screening module, and a sending module; The receiving module is used to receive the gradient ciphertext sent by the participant; The data processing module is used to decrypt and verify the gradient ciphertext after the receiving module receives the gradient ciphertext sent by the participants within a preset number or time to obtain gradient data; The reward processing module is used to calculate the reward of each participant after the data processing module verifies the gradient ciphertext, and send the reward to the reward center after signing the reward, so that the reward center sends the reward to the corresponding participant account; The screening module is used to determine qualified participants and generate a list of participants participating in the next round of federated learning; The data processing module is further used to generate an aggregate gradient according to the gradient data of the qualified participant, and encrypt the aggregate gradient to obtain an aggregate gradient ciphertext; The sending module is used to send the aggregated gradient ciphertext to the participants in the participant list, so that the participants can perform the next round of iterative training.

[0017] On the other hand, the present invention also provides a contribution-based federated learning incentive system, the system comprising: a cloud server, a reward center, and at least two participant nodes; the cloud server comprises the federated learning incentive device; The participant node is used to perform iterative training based on the global model or aggregated gradient ciphertext of the cloud server in combination with local data to obtain gradient data; encrypt the gradient data using a set encryption algorithm to generate gradient ciphertext; and send the gradient ciphertext to the cloud server; The cloud server is used to decrypt and verify the gradient ciphertext after receiving the gradient ciphertext sent by the participants within a preset number or time to obtain gradient data; after the verification is passed, the reward of each participant is calculated, and the reward is signed and sent to the reward center; qualified participants are determined, a list of participants participating in the next round of federated learning is generated, and an aggregate gradient is generated according to the gradient data of the qualified participants obtained by decryption, and the aggregate gradient is encrypted to obtain an aggregate gradient ciphertext; the aggregate gradient ciphertext is sent to the participants in the participant list so that the participants can perform the next round of iterative training; The reward center is used to send the reward to the corresponding participant account.

[0018] Optionally, the system further includes: a key generation center, configured to generate and distribute keys for the cloud server and nodes of each participant.

[0019] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the contribution-based federated learning incentive method are executed.

[0020] The contribution-based federated learning incentive method, device and system provided by the present invention encrypt the uploaded gradient parameters by the participants through an encryption algorithm; the cloud server calculates the rewards of each participant and sends the rewards to the corresponding participants through the reward center, thereby enhancing the enthusiasm of the participants and improving the efficiency of federated learning. At the same time, a list of participants participating in the next round of federated learning is generated. Only the selected participants can decrypt the aggregated gradient ciphertext generated by the cloud server and participate in the next round of federated learning process, thereby effectively ensuring the quality of the gradient data, and then ensuring the quality of the model finally obtained through federated learning.

[0021] Furthermore, the contribution of each participant is evaluated through marginal loss, and the reward of each participant is determined by the aggregate weight method according to the contribution of each participant, which fully ensures the fairness and rationality of the reward. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 It is a flow chart of the contribution-based federated learning incentive method provided by the present invention; Figure 2It is a structural schematic diagram of a contribution-based federated learning incentive device provided by the present invention; Figure 3 It is a structural diagram of the contribution-based federated learning incentive system provided by the present invention. DETAILED DESCRIPTION

[0024] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0025] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0026] In order to enhance the enthusiasm of participants in federated learning, thereby improving the efficiency of model training and ensuring the quality of the model, the present invention provides a contribution-based federated learning incentive method, device and system. Participants encrypt the uploaded gradient parameters through an encryption algorithm; the cloud server calculates the rewards of each participant and sends the rewards to the corresponding participants through the reward center. At the same time, a list of participants participating in the next round of federated learning is generated. Only the screened participants can decrypt the aggregated gradient ciphertext generated by the cloud server and participate in the next round of federated learning process to ensure the quality of the gradient data and thus the quality of the model finally obtained through federated learning.

[0027] In an embodiment of the present invention, in order to ensure the security of data, the gradient ciphertext and the aggregated gradient may be transmitted in an encrypted manner, for example, but not limited to: symmetric encryption algorithms such as AES (Advanced Encryption Standard), asymmetric encryption algorithms such as RSA, SM2, ECDHE (Elliptic Curve Diffie-Hellman Ephemeral), etc.

[0028] Taking the AES symmetric encryption algorithm as an example, the encryption calculation is shown in the following formula (1): C=E(K,P)(1) Among them, C is the encrypted ciphertext, K is the key, P is the plaintext, and E is the encryption function.

[0029] The decryption calculation is shown in the following formula (2): P=D(K,P)(2) Where D is the decryption function, and the decryption process is the inverse operation of the encryption process.

[0030] In some embodiments, the key generation center may generate keys for various entities in the system, such as cloud servers, various participants, and reward centers, and distribute them to the corresponding entities. For different encryption algorithms, the method and process of generating keys may be different, which is not limited in the embodiments of the present invention.

[0031] like Figure 1 FIG. 1 is a flow chart of a contribution-based federated learning incentive method provided by the present invention, comprising the following steps: Step 101, receiving the gradient ciphertext sent by the participant.

[0032] The gradient ciphertext is sent by the participants to the cloud server. The participants perform iterative training based on the received global model or aggregated gradient ciphertext in combination with local data until the training model converges to obtain gradient data. The gradient data is then encrypted and generated using a set encryption algorithm.

[0033] After each round of federated learning, the participants upload their generated gradient ciphertexts to the cloud server.

[0034] Step 102, after receiving the gradient ciphertext sent by the participants within a preset number or time, decrypt and verify the gradient ciphertext to obtain gradient data.

[0035] Step 103, calculate the reward of each participant, sign the reward and send it to the reward center, so that the reward center sends the reward to the corresponding participant account.

[0036] In order to improve the enthusiasm of the participants, certain rewards can be given to each participant. The rewards can be, but are not limited to, any one or more of the following forms: currency, physical objects, points, customized model distribution or other forms of returns.

[0037] Taking into account the different situations of each participant, different reward rules can be set: For example, in some embodiments, the reward for each participant may be determined based on the cumulative number of participations, and the more times a participant participates, the higher the reward.

[0038] Furthermore, considering that each participant has different amounts of data, data quality, etc., their contributions to the final global model will also be different. Therefore, in order to more accurately, fairly and reasonably calculate the rewards of each participant, in other embodiments, the contribution of each participant can be calculated, and then the reward of the participant can be calculated based on the contribution of each participant.

[0039] For example, the contribution of each participant can be calculated through marginal loss.

[0040] Marginal loss helps determine the contribution of each client to the global model during model training. This setting not only ensures the security of data privacy, but also enables multi-source data to be effectively used for model training. In the solution of the present invention, the marginal loss of each participant can be calculated by comparing the changes in model performance before and after each participant joins the federated learning.

[0041] For example, in t In the round of federated learning training, use represents the global model obtained by aggregating the local models of all participants, Indicates that by aggregation The global model is obtained by combining the local models of all participants except

[0042] set up and Respectively and The loss on validation datasets such as Cafar, Emnist, Minest.

[0043] In a non-limiting embodiment, for example, the participants The marginal loss can be defined as: ; When higher quality local models are aggregated, the loss of the global model will decrease. Use δ to represent the quality threshold, and accept the local model of participant i when δi(t)≥δ (indicating that the model contributes more to the global model). The larger δ is, the greater the contribution of participant i is.

[0044] In this way, the contribution of each participant in the federal cooperation can be effectively evaluated and its marginal loss can be determined.

[0045] When calculating the reward of each participant according to their contribution, the calculation can be based on the weight method. Specifically, the weight of each participant in this round is calculated by the contribution of each participant, and then the reward of each participant is calculated according to the weight of each participant in this round and the budget of this round.

[0046] Calculating the rewards for each participant based on their contribution can ensure that each participant receives a return that matches their contribution, which can greatly promote the enthusiasm of the participants.

[0047] The cloud server can use its own private key to sign the reward and send it to the reward center.

[0048] For example, signature generation and verification can be achieved using, but not limited to, RSA, identity-based signature algorithm, certificateless signature algorithm, attribute-based signature and other algorithms.

[0049] Accordingly, after receiving the reward signatures of the corresponding participants sent by the cloud server, the reward center will verify the signatures to verify their validity. If the signatures are verified to be valid, the corresponding incentives will be sent to the corresponding accounts of the participants.

[0050] Each participant may register in advance with the reward center and provide its account to the reward center when registering; or each participant may register with the cloud server when participating in federated learning and provide its account to the cloud server when registering. When the cloud server sends the reward signature to the reward center, it also provides the participant's account to the reward center, which is not limited in the embodiments of the present application.

[0051] Step 104: determine qualified participants and generate a list of participants participating in the next round of federated learning.

[0052] In order to better ensure the quality of the model and avoid the impact of malicious data (such as data sent by malicious participants that have a negative impact on the model) or low-quality gradient data on the model quality, after each round of federated learning, the cloud server can screen the participants of this round and select qualified participants to participate in the next round of federated learning.

[0053] Specifically, the consistency with the previous round of aggregation model can be determined based on the gradient data of each participant, and the qualified participants can be determined based on the consistency; then the qualified participants can be added to the participant list.

[0054] The consistency of the gradient data of each participant with the last round of aggregation model can be calculated by calculating the error between the gradient data of each participant and the last round of aggregation model. For example, the error can be expressed by calculating the Euclidean distance. The greater the distance, the greater the error of the gradient uploaded by the participant.

[0055] In the solution of the present invention, there are many ways to determine the qualified participants, such as: In some embodiments, a high-quality participant can be selected based on the error. Specifically, if the error is less than a first set value, the corresponding participant is selected as a high-quality participant; and then the high-quality participant is added to the participant list as a qualified participant.

[0056] In other embodiments, poor quality participants may be selected based on the error and eliminated from all participants in this round, that is, participants other than the poor quality participants are added to the participant list as qualified participants.

[0057] For example, in a non-limiting embodiment, the inferior participant can be determined in the following manner: a label corresponding to each participant is set in advance, and the initial value of the label is set to 0. After each round of federated learning, it is determined whether the error corresponding to each participant is greater than or equal to a second set value; if so, the label value of the participant is increased by 1; if the label value of the participant is greater than a set threshold, the participant is regarded as an inferior participant.

[0058] Step 105, generating an aggregate gradient according to the gradient data of the qualified participants, and encrypting the aggregate gradient to obtain an aggregate gradient ciphertext.

[0059] The cloud server aggregates the gradient data obtained by decryption to obtain an aggregated gradient. It should be noted that the embodiment of the present application does not limit the aggregation method of the gradient data, which can be determined according to actual application needs.

[0060] Then, the cloud server encrypts the aggregated gradient using its own key to obtain an aggregated gradient ciphertext.

[0061] Step 106: Send the aggregated gradient ciphertext to the participants in the participant list so that the participants can perform the next round of iterative training.

[0062] The cloud server can encrypt the aggregate gradient through a dynamic broadcast encryption algorithm or an attribute-based encryption algorithm, and send the aggregate gradient ciphertext to the participants in the participant list, ensuring that only the participants with access rights can decrypt the ciphertext.

[0063] For example, the calculation method of the dynamic broadcast encryption algorithm is roughly as follows: The dynamic broadcast encryption method consists of four algorithms: system establishment, key generation, encryption and decryption. Among them: System establishment: input security parameter λ and the maximum number of participants n, and output system parameters and master key ek.

[0064] Key generation: Input system parameters and user serial number IDi, and output the identity label labi and decryption key dki of the corresponding federated learning user.

[0065] Encryption: Let the training user set be S=(ID1,…,IDn), and the revoked user set be RS. The broadcaster first generates (Hdr,SK) for S\RS using the initial encryption key ek, where Hdr is called the ciphertext header and SK is called the session key. Then, the message M (i.e., the aggregated gradient) is encrypted using K to obtain the broadcast ciphertext CM, and (Hdr,RS,CM) is broadcast to each user.

[0066] Decryption: Assume that user IDi is a participant in the participant list, (Hdr, RS, CM) is the ciphertext it receives, and decrypt CM to recover the message plaintext M.

[0067] For another example, the cloud server uses an attribute-based encryption algorithm for aggregated gradients. The participants first verify whether the user attributes meet the access policy. If so, the aggregated gradient ciphertext can be decrypted. Otherwise, it cannot be decrypted.

[0068] The contribution-based federated learning incentive method provided by the present invention is that the participants encrypt the uploaded gradient parameters through an encryption algorithm; the cloud server calculates the rewards of each participant and sends the rewards to the corresponding participants through the reward center, thereby enhancing the enthusiasm of the participants and improving the efficiency of federated learning. At the same time, a list of participants participating in the next round of federated learning is generated. Only the selected participants can decrypt the aggregated gradient ciphertext generated by the cloud server and participate in the next round of federated learning process, thereby effectively ensuring the quality of the gradient data, and then ensuring the quality of the model finally obtained through federated learning.

[0069] Accordingly, the embodiment of the present invention also provides a contribution-based federated learning incentive device, such as Figure 2 Shown is a structural schematic diagram of the device.

[0070] The federated learning incentive device 200 includes: a receiving module 201, a data processing module 202, a reward processing module 203, a screening module 204, and a sending module 205. Among them: The receiving module 201 is used to receive the gradient ciphertext sent by the participant; The data processing module 202 is used to decrypt and verify the gradient ciphertext after the receiving module 201 receives the gradient ciphertext sent by the participants within a preset number or time, so as to obtain the gradient data; The reward processing module 203 is used to calculate the reward of each participant after the data processing module 202 verifies the gradient ciphertext, and send the reward to the reward center after signing the reward, so that the reward center sends the reward to the corresponding participant account; The screening module 204 is used to determine qualified participants and generate a list of participants participating in the next round of federated learning; The data processing module 202 is also used to generate an aggregate gradient according to the gradient data of the qualified participants, and encrypt the aggregate gradient to obtain an aggregate gradient ciphertext; The sending module 205 is used to send the aggregated gradient ciphertext to the participants in the participant list, so that the participants can perform the next round of iterative training.

[0071] The specific implementation methods of the above modules can refer to the description of the federated learning incentive method of the present invention, which will not be repeated here.

[0072] Accordingly, the embodiment of the present invention also provides a contribution-based federated learning incentive system, such as Figure 3 The figure shows a structural diagram of the system.

[0073] The contribution-based federated learning incentive system includes: a cloud server 20, a reward center 10, and at least two participant nodes 30. The cloud server 20 includes the above-mentioned federated learning incentive device 200, and the cloud server 20 and the reward center 10 and each participant node 30 may be connected by wire or wirelessly, which is not limited in this embodiment of the present invention. The participant node 30 is used to perform iterative training based on the global model or aggregated gradient ciphertext of the cloud server 20 and combined with local data to obtain gradient data; encrypt the gradient data using a set encryption algorithm to generate gradient ciphertext; and send the gradient ciphertext to the cloud server 20; The cloud server 20 is used to decrypt and verify the gradient ciphertext after receiving the gradient ciphertext sent by the participants within a preset number or time to obtain gradient data; after the verification is passed, calculate the reward of each participant, sign the reward and send it to the reward center; determine qualified participants, generate a list of participants participating in the next round of federated learning, generate an aggregate gradient based on the gradient data of the qualified participants obtained by decryption, encrypt the aggregate gradient to obtain an aggregate gradient ciphertext; send the aggregate gradient ciphertext to the participants in the participant list, so that the participants can perform the next round of iterative training; The reward center 10 is used to send the reward to the corresponding participant account.

[0074] Furthermore, in some non-limiting embodiments, the federated learning incentive system may also include a key generation center (not shown) for generating and distributing keys for the cloud server 20 and each participant node 30 .

[0075] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0076] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0077] In several embodiments provided by the present invention, it should be understood that the disclosed device can be implemented in other ways.

[0078] The present invention also provides a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and the computer program can be executed when it is run. Figure 1 The storage medium may include a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc. The storage medium may also include a non-volatile memory (non-volatile) or a non-transitory memory, etc.

[0079] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless means.

[0080] The embodiments of the present invention are described in detail above. The present invention is described in detail using specific implementation methods herein. The description of the above embodiments is only used to help understand the method and system of the present invention. It is only a part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should belong to the scope of protection of the present invention, and the content of this specification should not be understood as limiting the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A contribution-based federated learning incentive method, It is characterized in that The method comprises: Receive the gradient ciphertext sent by the participating party; After receiving the gradient ciphertext sent by the participants within a preset number or time, decrypt and verify the gradient ciphertext to obtain gradient data; Calculate the rewards for each participant, sign the rewards and send them to the reward center, so that the reward center can send the rewards to the corresponding participant accounts; Identify qualified participants and generate a list of participants for the next round of federated learning; Generate an aggregate gradient according to the gradient data of the qualified participant, and encrypt the aggregate gradient to obtain an aggregate gradient ciphertext; Sending the aggregated gradient ciphertext to the participants in the participant list so that the participants can perform the next round of iterative training; The calculation of the rewards for each participant includes: Calculate the contribution of each participant through marginal loss; In the t In the round of federated learning training, use represents the global model obtained by aggregating the local models of all participants, Indicates that by aggregation The global model obtained by combining the local models of all participants except set up and Respectively and The loss on the validation dataset; The participants The marginal loss is defined as: ; When higher-quality local models are aggregated, the loss of the global model will decrease. We use δ to represent the quality threshold, and accept the local model of participant i when δi(t)≥δ, which means that the model contributes more to the global model. The larger δ is, the greater the contribution of participant i is. Calculate the rewards of each participant according to their contribution; The calculation of the reward of each participant according to the contribution of each participant includes: The weight of each participant in this round is calculated based on the contribution of each participant; The rewards for each participant are calculated based on their weight and budget for this round.

2. According to the contribution-based federated learning incentive method of claim 1, It is characterized in that The determination of qualified participants includes: Determine the consistency with the previous round of aggregation model based on the gradient data of each participant; Qualified parties are determined based on the consistency.

3. According to the contribution-based federated learning incentive method of claim 2, It is characterized in that Determining the consistency with the previous round of aggregation model based on the gradient data of each participant includes: Calculate the error between the gradient data of each participant and the previous round of aggregation model.

4. The contribution-based federated learning incentive method according to claim 3, It is characterized in that The qualified participants include: high-quality participants; High-quality participants determined based on the consistency include: If the error is smaller than a first set value, the participant is regarded as a high-quality participant.

5. The contribution-based federated learning incentive method according to claim 4, It is characterized in that The qualified participants include: participants other than inferior participants; The inferior participants determined based on the consistency include: Pre-set the label corresponding to each participant and set the initial value of the label to 0; If the error corresponding to the participant is greater than or equal to a second set value, the tag value of the participant is increased by 1; If the tag value of the participant is greater than a set threshold, the participant is regarded as a low-quality participant.

6. The contribution-based federated learning incentive method according to any one of claims 1 to 5, It is characterized in that The method further comprises: The participant performs iterative training based on the received global model or aggregated gradient ciphertext in combination with local data to obtain gradient data; Encrypting the gradient data using a set encryption algorithm to generate a gradient ciphertext; Upload the gradient ciphertext.

7. The contribution-based federated learning incentive method according to claim 6, It is characterized in that The encryption algorithm includes any one of the following: a symmetric encryption algorithm and an asymmetric encryption algorithm.

8. A contribution-based federated learning incentive device, It is characterized in that The device comprises: a receiving module, a data processing module, a reward calculation module, a screening module, and a sending module; The receiving module is used to receive the gradient ciphertext sent by the participant; The data processing module is used to decrypt and verify the gradient ciphertext after the receiving module receives the gradient ciphertext sent by the participants within a preset number or time to obtain gradient data; The reward processing module is used to calculate the reward of each participant after the data processing module verifies the gradient ciphertext, and send the reward to the reward center after signing the reward, so that the reward center sends the reward to the corresponding participant account; wherein, calculating the reward of each participant includes: Calculate the contribution of each participant through marginal loss; In the t In the round of federated learning training, use represents the global model obtained by aggregating the local models of all participants, Indicates that by aggregation The global model obtained by combining the local models of all participants except set up and Respectively and The loss on the validation dataset; The participants The marginal loss is defined as: ; When higher-quality local models are aggregated, the loss of the global model will decrease. We use δ to represent the quality threshold, and accept the local model of participant i when δi(t)≥δ, which means that the model contributes more to the global model. The larger δ is, the greater the contribution of participant i is. The calculation of the rewards for each participant also includes: The weight of each participant in this round is calculated based on the contribution of each participant; The rewards for each participant are calculated based on their current round weight and budget; The screening module is used to determine qualified participants and generate a list of participants participating in the next round of federated learning; The data processing module is further used to generate an aggregate gradient according to the gradient data of the qualified participant, and encrypt the aggregate gradient to obtain an aggregate gradient ciphertext; The sending module is used to send the aggregated gradient ciphertext to the participants in the participant list, so that the participants can perform the next round of iterative training.

9. A contribution-based federated learning incentive system, It is characterized in that The system comprises: a cloud server, a reward center, and at least two participant nodes; the cloud server comprises the federated learning incentive device according to claim 8; The participant node is used to perform iterative training based on the global model or aggregated gradient ciphertext of the cloud server in combination with local data to obtain gradient data; encrypt the gradient data using a set encryption algorithm to generate gradient ciphertext; and send the gradient ciphertext to the cloud server; The cloud server is used to decrypt and verify the gradient ciphertext after receiving the gradient ciphertext sent by the participants within a preset number or time to obtain gradient data; after the verification is passed, the reward of each participant is calculated, and the reward is signed and sent to the reward center; qualified participants are determined, a list of participants participating in the next round of federated learning is generated, and an aggregate gradient is generated according to the gradient data of the qualified participants obtained by decryption, and the aggregate gradient is encrypted to obtain an aggregate gradient ciphertext; the aggregate gradient ciphertext is sent to the participants in the participant list so that the participants can perform the next round of iterative training; The reward center is used to send the reward to the corresponding participant account.

10. The contribution-based federated learning incentive system according to claim 9, It is characterized in that The system further comprises: The key generation center is used to generate and distribute the keys of the cloud server and the nodes of each participant.

11. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the steps of the contribution-based federated learning incentive method described in any one of claims 1 to 7 are performed.

Citation Information

Patent Citations

  • Federal learning model and data security and correctness detection method

    CN115292738A

  • Fair privacy calculation method based on federated node contribution

    CN116306910A

Cited By

  • Decentralized hierarchical federated learning method and system, and edge server

    CN119338029A

  • Federal learning incentive method, device and system based on reverse auction

    CN119514732A

  • Decentralized hierarchical federated learning method and system, and edge server

    WO2026130579A1