Multi-party authorization system oriented to federated learning

Abstract

The invention relates to the technical field of data sharing, in particular to a federated learning-oriented multi-party authorization system, which comprises a model aggregation module for processing a local model of a participant to output a federated model; the verification module is connected with the model aggregation module and is used for verifying the use permission of the user according to the authorization certificate of the participant; the control module is respectively connected with the model aggregation module and the verification module, and is used for executing the verification module when judging that the operation safety of the multi-party authorization system does not meet the requirement according to the byte quantity proportion of the authorization certificate of the user; and adjusting the transmission rate of the authorization certificate or determining the running memory proportion of the secret share generator according to the average generation number of the secret shares, and adjusting the training period of the federal model for the first time according to the byte difference quantity before and after decoding the authorization certificate. According to the invention, the operation security of the multi-party authorization system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data sharing, and in particular to a multi-party authorization system for federated learning. Background Art

[0002] The concept of federated learning was originally proposed to build a machine learning model based on data sets that are distributed among multiple users and do not share data with each other. The main idea is to aggregate local models learned on multiple devices without sending private data to semi-honest third-party servers or other devices; considering the privacy issues of secret data distributed across multiple institutions, Yang Qiang et al. extended the application of federated learning to a wide range of use cases and divided the scenarios into three categories: horizontal federated learning, vertical federated learning, and federated transfer learning. Yang Qiang et al. further proposed the concept of trusted federated learning to emphasize the importance of model intellectual property rights when building a secure federated learning system.

[0003] In the context of privacy issues plaguing major platforms, federated learning has become a hot research direction. In the distributed learning process of federated learning, the various roles involved in learning do not exchange their own data, but only complete the overall model training by exchanging model parameters or intermediate results, thereby achieving a series of business model creation and optimization needs through sample or feature expansion while meeting data security such as financial supervision, data not leaving the domain and privacy protection.

[0004] After years of development, federated learning has also come into many categories. According to network topology, there are centralized federated learning and decentralized federated learning. According to data partitioning, there are horizontal federated learning, vertical federated learning, and federated transfer learning. The successful application of deep neural networks (DNNs) in computer vision, natural language processing, and data mining tasks comes at the cost of expensive training processes. On the one hand, training requires a lot of effort and cost in terms of expertise, dedicated hardware, and designing and training DNN models. On the other hand, a large amount of training data is required to improve model performance, and the amount of training data tends to increase monotonically with the increase in the amount of training data. Therefore, it is urgent to protect valuable training data and trained DNN models from being illegally copied, redistributed, or abused. So far, both white-box watermarking and black-box watermarking methods have been widely used in the verification of DNN model ownership. In terms of white-box watermarking, a specified binary string is added to the model as a watermark encoding, and the verifier verifies its copyright by extracting the corresponding binary string. However, most deep learning models deployed on cloud servers, including pirated models, are black-box scenarios. A black-box watermarking method uses a backdoor as a watermark key image. Specific backdoor images will trigger the backdoor and verify copyright ownership.

[0005] China Patent Publication No.: CN117763616A discloses a method and system for assessing the contribution of multiple subjects in a federated learning sharing process, the method includes: a data sharing contribution assessment model based on blockchain, combined with smart contract calculation, to obtain a contribution assessment result; the interactive process of the data sharing contribution assessment model based on blockchain includes: each entity first uses the organization authentication authorization to register to ensure the identity and trust of the user; the entity includes participants, coordinators and blockchains; the participant passes the contribution assessment statistic to the contribution assessment smart contract, the contribution assessment smart contract receives and parses the input data, and processes it according to the preset contribution assessment rules and algorithms to obtain the contribution assessment result, and stores the contribution assessment result; wherein the assessment statistic data mainly includes four dimensions: basic information, data value, local model evaluation, and model contribution weight. It can be seen that the method and system for assessing the contribution of multiple subjects in the federated learning sharing process have vulnerabilities. Since the existing federated learning protocol has been proven to have vulnerabilities, attackers can use these vulnerabilities to destroy data privacy, thereby causing errors in the data transmission process, resulting in data loss of the authorization credential, and then causing the system to fail to identify the credential, thereby causing the operation security of the multi-party authorization system to decrease. Summary of the invention

[0006] To this end, the present invention provides a multi-party authorization system for federated learning, which is used to overcome the problem in the prior art that, since the existing federated learning protocol has been proven to have loopholes, attackers can exploit these loopholes to destroy data privacy, thereby causing errors in the data transmission process, resulting in data loss of authorization credentials, and further causing the system to fail to recognize the credentials, thereby reducing the operational security of the multi-party authorization system.

[0007] To achieve the above-mentioned purpose, the present invention provides a multi-party authorization system for federated learning, comprising: a model aggregation module, which is used to process the local models of the participants to output a federated model, including a model training unit for training the local models of several participants to output a model to be aggregated and a model aggregation unit connected to the model training unit for aggregating the corresponding number of models to be aggregated to output a federated model; a verification module, which is connected to the model aggregation module and is used to verify the user's use authority according to the authorization certificate of the participant, including a secret share generator for generating a corresponding secret share according to the backdoor information code of the participant, and a secret share generator connected to the secret share generator for generating a secret share according to the backdoor information code of the participant. an authorization credential distributor for distributing secret shares to participants and a credential restorer for verifying the user's usage rights; a control module, which is respectively connected to the model aggregation module and the verification module, and is used to adjust the transmission rate of the authorization credential or determine the operating memory ratio of the secret share generator according to the average number of secret shares generated when it is determined that the operating security of the multi-party authorization system does not meet the requirements according to the byte ratio of the user's authorization credential, and to initially adjust the training cycle of the federal model according to the byte difference before and after the decoding of the authorization credential, and to secondary adjust the training cycle of the federal model according to the average training time of the federal model for the same amount of data after the federal model runs a single cycle with the training cycle after the initial adjustment.

[0008] Furthermore, the model aggregation module also includes a trusted server for adjusting the federated model process.

[0009] Furthermore, the verification module also includes a backdoor information encoder for generating the backdoor information code.

[0010] Furthermore, the process of verifying the user's usage rights is that the user provides the secret share to the trusted server, the trusted server first decrypts the secret share, and then performs secret sharing recovery on the secret share to obtain the backdoor information code, which is used as the input of the credential decoder. The credential decoder generates a set of backdoor instances for verification based on the input backdoor information code and returns it. The trusted server inputs the backdoor instance into the model and returns the confidence level, and judges the user's usage rights based on the confidence level.

[0011] Further, the control module is connected to the authorization verification unit, and is used to calculate the byte volume ratio of the user's authorization credential when verifying according to the byte volume of the user's authorization credential when verifying and the original byte volume of the user's authorization credential when verifying, and to determine that the operation security of the multi-party authorization system does not meet the requirements when the byte volume ratio of the user's authorization credential when verifying satisfies the first ratio condition or the second ratio condition;

[0012] The control module is connected to the secret share generator, and is used to preliminarily determine that the multi-party authorization stability does not meet the requirements when the byte ratio of the user's authorization credential during the verification only meets the first ratio condition, and to perform a secondary determination on whether the multi-party authorization stability meets the requirements according to the average number of secret shares generated;

[0013] The control module is connected to the credential transmission unit, and is used to reduce the transmission rate of the authorization credential when the byte ratio of the user's authorization credential during the verification only meets the second ratio condition;

[0014] Among them, the first proportion condition is that the byte ratio of the user's authorization credentials when performing verification is greater than the preset first ratio and less than or equal to the preset second ratio; the second proportion condition is that the byte ratio of the user's authorization credentials when performing verification is greater than the preset second ratio.

[0015] Furthermore, the calculation formula for the byte ratio of the user's authorization certificate during verification is:

[0016]

[0017] Among them, S is the byte ratio of the user's authorization certificate when verifying, T1 is the byte size of the user's authorization certificate when verifying, and T2 is the original byte size of the user's authorization certificate.

[0018] Furthermore, the reduced transmission rate of the authorization credential is determined by the difference between the byte ratio of the user's authorization credential during verification and a preset second ratio.

[0019] Further, the control module is connected to the secret share generator, and is used to make a second determination that the stability of multi-party authorization does not meet the requirement when the average number of secret shares generated meets the first number condition or the second number condition, and to make a preliminary determination that the validity of the authorization verification does not meet the requirement when the average number of secret shares generated only meets the second number condition, and to make a second determination on whether the validity of the authorization verification meets the requirement according to the byte difference before and after the decoding of the authorization credential;

[0020] The control module is connected to the secret share generator, and is used to increase the operating memory ratio of the secret share generator when the average generated quantity of the secret share only meets the first quantity condition;

[0021] The increased operating memory ratio of the secret share generator is determined by the difference between the average number of secret shares generated and the preset first number;

[0022] Among them, the first quantity condition is that the average generated number of secret shares is greater than the preset first number and less than or equal to the preset second number; the first quantity condition is that the average generated number of secret shares is greater than the preset second number.

[0023] Furthermore, the control module is connected to the secret share generator and the model training unit respectively, so as to determine that the validity of the authorization verification does not meet the requirements when the difference between the bytes before and after decoding the authorization credential is greater than a preset difference, and to increase the training cycle of the federation model;

[0024] The training cycle of the increased federated model is determined by the difference between the byte difference before and after the decoding of the authorization credential and the preset difference.

[0025] Furthermore, the control module is connected to the secret share generator and the model training unit respectively, so as to determine that the processing efficiency of the federation model does not meet the requirements when the average training time of the federation model for the same amount of data is greater than the preset training time, and reduce the training cycle of the increased federation model;

[0026] The reduced training cycle of the federated model is determined by the difference between the average training time of the federated model for the same amount of data and the preset training time.

[0027] Compared with the prior art, the beneficial effect of the present invention lies in that the system of the present invention adjusts the transmission rate of the authorization credential according to the byte ratio of the user's authorization credential during verification by setting a model aggregation module, a verification module and a control module. Due to network fluctuations at the user end, errors occur in the data transmission process, resulting in data loss of the authorization credential, which in turn causes the system to fail to identify the credential, thereby failing authorization. By reducing the transmission rate of the authorization credential, the probability of data transmission errors caused by network fluctuations is reduced. By adjusting the running memory ratio of the secret share generator according to the average number of secret shares generated, since the secret share generator may need to rely on other components or services, incompatible changes may be introduced when the component is updated, resulting in incompatibility between the secret share generator and the component, which in turn causes the generator to be unable to generate the correct number of shares. By increasing the running memory ratio of the secret share generator, the resource utilization of the generator is improved. The usage rate is increased, and the impact of problems caused by incompatibility is reduced. The training cycle of the federal model is adjusted according to the byte difference before and after the decoding of the authorization credential. Since the character encoding of the credential data may be different, the credential decoder fails to process it correctly, resulting in parsing errors, improper handling of the boundary conditions of the credential data, and then causes problems with the credential decoder, resulting in unsuccessful verification. By increasing the training cycle of the federal model, the model further optimizes the credential data, reducing the impact of verification failures caused by different character encodings of the credential data. The training cycle of the federal model is adjusted secondary according to the average training time of the federal model for the same amount of data. Since an excessively large training cycle will lead to an extension of the training time, users need to wait longer to transmit the local model to the central server, which leads to an extension of the overall training time. By reducing the training cycle of the federal model, the processing efficiency of the federal model is improved, and the operational security of the multi-party authorization system is improved.

[0028] Furthermore, in the system described in the present invention, the authorization of the model is no longer owned by a single central server, but is implemented through multi-party authorization. No party can independently determine the right to use the model, which ensures the fairness and reliability of the authorization. Only after collecting a sufficient number of authorization credentials can the user use the model, thus avoiding unauthorized use. By implementing a multi-party authorization mechanism, access to the model becomes more difficult and restricted. Users need to spend a certain amount of time and effort to collect authorization credentials, which increases the scarcity and value of the model. At the same time, the multi-party authorization mechanism can also dynamically adjust the number of authorizations and the acquisition threshold according to demand, further enhancing the value of the model. Through the multi-party authorization mechanism, the system effectively prevents unauthorized illegal use and dissemination. Only those users with legal authorization can use the model, which can prevent piracy, infringement and unauthorized use.

[0029] Furthermore, unlike the traditional centralized authorization system, this system adopts a distributed authorization credential generation and management mechanism for multiple participants. This distributed approach protects the authorization information while reducing the risk of single point failure, while improving the scalability and robustness of the system. In the system, the generation and management of authorization credentials are distributed among multiple participants. Even if a node is attacked or fails, other nodes can still continue to operate, maintaining the availability and stability of the system. The authorization credentials and data in this system are scattered across multiple nodes, and attackers need to attack multiple nodes at the same time to obtain complete information, which increases the difficulty of the attack.

[0030] Furthermore, the system of the present invention adjusts the transmission rate of the authorization credential by setting a preset first ratio and a preset second ratio. Due to fluctuations in the network on the user side, errors occur in the data transmission process, resulting in data loss in the authorization credential, which in turn causes the system to fail to recognize the credential, resulting in authorization failure. By reducing the transmission rate of the authorization credential, the probability of data transmission errors caused by network fluctuations is reduced, thereby further achieving an improvement in the operational security of the multi-party authorization system.

[0031] Furthermore, the system described in the present invention adjusts the running memory ratio of the secret share generator by setting a preset first number and a preset second number. Since the secret share generator may need to rely on other components or services, incompatible changes may be introduced when the components are updated, resulting in incompatibility between the secret share generator and the components, which in turn causes the generator to be unable to generate the correct number of shares. By increasing the running memory ratio of the secret share generator, the resource utilization of the generator is improved, the impact of problems caused by incompatibility is reduced, and the operational security of the multi-party authorization system is further improved.

[0032] Furthermore, the system of the present invention adjusts the training cycle of the federal model by setting a preset difference amount. Since the character encoding of the credential data may be different, the credential decoder may not process it correctly, resulting in parsing errors, improper handling of boundary conditions of the credential data, and further problems with the credential decoder, resulting in unsuccessful verification. By increasing the training cycle of the federal model, the model further optimizes the credential data, reducing the impact of verification failures caused by different character encodings of the credential data, and further achieving improved operational security for the multi-party authorization system.

[0033] Furthermore, the system of the present invention performs secondary adjustments to the training cycle of the federal model by setting a preset training duration. Since an excessively long training cycle will extend the training duration, users will need to wait longer to transmit the local model to the central server, thereby extending the overall training duration. By reducing the training cycle of the federal model, the processing efficiency of the federal model is improved, further achieving an improvement in the operational security of the multi-party authorization system. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is an overall structural block diagram of a multi-party authorization system for federated learning according to an embodiment of the present invention;

[0035] Figure 2 A specific flow chart of the federated model training aggregation process of a multi-party authorization system for federated learning according to an embodiment of the present invention;

[0036] Figure 3 A specific flow chart of a process of processing a user's local model to output a federated model in a multi-party authorization system for federated learning according to an embodiment of the present invention;

[0037] Figure 4 The present invention is a logical flow chart of a multi-party authorization system for federated learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0039] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0040] In addition, it should be noted that in the description of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0041] See also Figure 1 , Figure 2 , Figure 3 as well as Figure 4As shown, they are respectively an overall structural block diagram of a multi-party authorization system for federated learning according to an embodiment of the present invention, a specific flow chart of a federated model training aggregation process, a specific flow chart of a process of processing a user's local model to output a federated model, and a logic flow chart. A multi-party authorization system for federated learning according to the present invention comprises:

[0042] A model aggregation module, used for processing the local models of the participants to output a federated model, including a model training unit used for training the local models of several participants to output a model to be aggregated and a model aggregation unit connected to the model training unit to perform an aggregation operation on the corresponding number of models to be aggregated to output a federated model;

[0043] A verification module, which is connected to the model aggregation module and is used to verify the user's use rights according to the authorization certificate of the participant, including a secret share generator for generating corresponding secret shares according to the backdoor information code of the participant, an authorization certificate distributor connected to the secret share generator for distributing the secret shares to the participant, and a certificate restorer for verifying the user's use rights;

[0044] A control module, which is connected to the model aggregation module and the verification module respectively, and is used to adjust the transmission rate of the authorization certificate or determine the operating memory ratio of the secret share generator according to the average number of secret shares generated when it is determined that the operating security of the multi-party authorization system does not meet the requirements according to the byte ratio of the user's authorization certificate, and to initially adjust the training cycle of the federal model according to the byte difference before and after the authorization certificate is decoded, and to secondary adjust the training cycle of the federal model according to the average training time of the federal model for the same amount of data after the federal model runs a single cycle with the training cycle after the initial adjustment.

[0045] Specifically, the role of aggregation is to merge the local models updated by several users to generate a federated model.

[0046] Specifically, the byte difference before and after the authorization voucher is decoded means the difference between the byte amount of the authorization voucher data and the byte amount of the data after the authorization voucher data is decoded by the voucher decoder.

[0047] Specifically, the process of processing the user's local model to output the federated model is as follows: first, the local model of the participant is uploaded to the trusted server, and the trusted server feeds back the federated model available to each node through the model training unit and the model aggregation unit, and then the trusted server generates backdoor information. In the model training unit, the trusted server converts the generated backdoor information into a backdoor information code through the backdoor information encoder. After that, a part of the training information is selected from each node, the backdoor information code is embedded therein, a backdoor instance is generated, and the backdoor instance is used as a backdoor data set. The generation of the backdoor data set is also included in the model training unit. After constructing the backdoor data set, the trusted server also needs to construct an authorization authentication unit. The trusted server generates a certain number of secret shares based on the Shamir secret sharing scheme using the backdoor information code. After completing the above process, the trusted server encrypts the obtained certain number of secret shares, packages them with the backdoor data set, and distributes them to each node. Next, each node trains a model with a backdoor through the backdoor data set and uploads it to the trusted server. The trusted server calls the model training unit and the model aggregation unit again, aggregates the backdoor model and distributes it to each node. In this way, each node obtains the federated model with the backdoor and an encrypted secret sharing share.

[0048] Specifically, if the model is trained using a backdoor dataset, the structure of the model will be fine-tuned by the backdoor dataset and become a model with a backdoor.

[0049] Specifically, the process of determining the secret share generator is as follows: all participants negotiate to determine the minimum number of participants that can represent the "overall" authorization, that is, the threshold. Then, N secret shares are generated based on the Shamir algorithm according to the threshold n, the backdoor information code c and the number of federated participants N.

[0050] Specifically, the operation process of the authorization credential distributor is as follows: after generating the secret share and backdoor data set, the authorization credential distributor first homomorphically encrypts the key, and then distributes the backdoor data set and secret share to each federated node.

[0051] Specifically, the operation process of the credential restorer is as follows: the user provides the share to the trusted server, the server first decrypts the share, and then recovers the share through secret sharing to obtain the backdoor information code. The backdoor information code is used as the input of the credential decoder. The credential decoder generates a set of backdoor instances for verification based on the input backdoor information code and returns it. The server inputs the instance into the model and returns the confidence level, and judges the user's authority based on the confidence level.

[0052] Specifically, the steps of federated model training aggregation include:

[0053] Step S1, initialization of global model parameters;

[0054] Step S2: The trusted server transmits global model parameters to all participants and selects a number of participants for training according to a fixed ratio;

[0055] Step S3, the selected participant initializes the local model using the global model parameters, performs multiple rounds of gradient descent, and updates the parameters multiple times;

[0056] Step S4, the selected participant transmits the local model parameters to the trusted server;

[0057] Step S5: The trusted server aggregates the local model parameters into global model parameters by weighted average aggregation.

[0058] Specifically, in the system described in the present invention, the authorization of the model is no longer owned by a single central server, but is implemented through multi-party authorization. No party can decide the right to use the model alone, which ensures the fairness and reliability of the authorization. Only after collecting a sufficient number of authorization credentials can the user use the model, thus avoiding unauthorized use. By implementing a multi-party authorization mechanism, access to the model becomes more difficult and restricted. Users need to spend a certain amount of time and energy to collect authorization credentials, which increases the scarcity and value of the model. At the same time, the multi-party authorization mechanism can also dynamically adjust the number of authorizations and the acquisition threshold according to demand, further enhancing the value of the model. Through the multi-party authorization mechanism, the system effectively prevents unauthorized illegal use and dissemination. Only those users with legal authorization can use the model, which can prevent piracy, infringement and unauthorized use.

[0059] Specifically, unlike the traditional centralized authorization system, this system adopts a distributed authorization credential generation and management mechanism for multiple participants. This distributed approach protects the authorization information while reducing the risk of single point failure, while improving the scalability and robustness of the system. In the system, the generation and management of authorization credentials are distributed among multiple participants. Even if a node is attacked or fails, other nodes can still continue to operate, maintaining the availability and stability of the system. The authorization credentials and data in this system are scattered across multiple nodes. Attackers need to attack multiple nodes at the same time to obtain complete information, which increases the difficulty of the attack.

[0060] In implementation, the system of the present invention sets a model aggregation module, a verification module and a control module, and adjusts the transmission rate of the authorization credential according to the byte ratio of the user's authorization credential during verification. Due to network fluctuations at the user end, errors occur in the data transmission process, resulting in data loss in the authorization credential, which in turn causes the system to fail to identify the credential, thereby failing authorization. By reducing the transmission rate of the authorization credential, the probability of data transmission errors caused by network fluctuations is reduced. By adjusting the running memory ratio of the secret share generator according to the average number of secret shares generated, since the secret share generator may need to rely on other components or services, incompatible changes may be introduced when the component is updated, resulting in incompatibility between the secret share generator and the component, which in turn causes the generator to be unable to generate the correct number of shares. By increasing the running memory ratio of the secret share generator, the resource utilization of the generator is improved, and the error caused by the generator is reduced. In order to solve the problem caused by incompatibility, the training cycle of the federated model is adjusted according to the byte difference before and after the decoding of the authorization credential. Since the character encoding of the credential data may be different, the credential decoder fails to process it correctly, resulting in parsing errors, improper handling of the boundary conditions of the credential data, and then problems with the credential decoder, which leads to unsuccessful verification. By increasing the training cycle of the federated model, the model further optimizes the credential data, reducing the impact of verification failures caused by different character encodings of the credential data. The training cycle of the federated model is adjusted secondary according to the average training time of the federated model for the same amount of data. Since an excessively large training cycle will extend the training time, users need to wait longer to transmit the local model to the central server, which extends the overall training time. By reducing the training cycle of the federated model, the processing efficiency of the federated model is improved, thereby improving the operational security of the multi-party authorization system.

[0061] Specifically, the model aggregation module also includes a trusted server for adjusting the federated model process.

[0062] Specifically, the verification module also includes a backdoor information encoder for generating the backdoor information code.

[0063] Specifically, the process of verifying the user's usage rights is that the user provides the secret share to the trusted server. The trusted server first decrypts the secret share, and then performs secret sharing recovery on the secret share to obtain the backdoor information code. The backdoor information code is used as the input of the credential decoder. The credential decoder generates a set of backdoor instances for verification based on the input backdoor information code and returns it. The trusted server inputs the backdoor instance into the model and returns the confidence level, and judges the user's usage rights based on the confidence level.

[0064] Specifically, the control module is connected to the authorization verification unit, and is used to calculate the byte volume ratio of the user's authorization credential when verifying according to the byte volume of the user's authorization credential when verifying and the original byte volume of the user's authorization credential when verifying, and to determine that the operation security of the multi-party authorization system does not meet the requirements when the byte volume ratio of the user's authorization credential when verifying satisfies the first ratio condition or the second ratio condition;

[0065] The control module is connected to the secret share generator, and is used to preliminarily determine that the multi-party authorization stability does not meet the requirements when the byte ratio of the user's authorization credential during the verification only meets the first ratio condition, and to perform a secondary determination on whether the multi-party authorization stability meets the requirements according to the average number of secret shares generated;

[0066] The control module is connected to the credential transmission unit, and is used to reduce the transmission rate of the authorization credential when the byte ratio of the user's authorization credential during the verification only meets the second ratio condition;

[0067] Among them, the first proportion condition is that the byte ratio of the user's authorization credentials when performing verification is greater than the preset first ratio and less than or equal to the preset second ratio; the second proportion condition is that the byte ratio of the user's authorization credentials when performing verification is greater than the preset second ratio.

[0068] Optionally, a preferred embodiment of the preset first ratio Q1 is Q1=0.7, and a preferred embodiment of the preset second ratio Q2 is Q2=0.75.

[0069] Specifically, the byte ratio of the user's authorization certificate during verification is recorded as Q, and the difference between the byte ratio of the user's authorization certificate during verification and the preset second ratio is recorded as ΔQ, and ΔQ=Q-Q2 is set.

[0070] In implementation, the system of the present invention determines the operational security of the multi-party authorization system by setting a preset first ratio and a preset second ratio, thereby reducing the impact of the decreased operational stability of the federal model due to inaccurate determination of the operational security of the multi-party authorization system, and further improving the operational security of the multi-party authorization system.

[0071] Specifically, the process of user verification can be as follows: the user wants to use a service provided by a federated training model. Because the model is publicly owned by each federated node, assuming that there are N nodes in total, the user needs to send a usage request to no less than n (threshold) federated nodes. After accepting the user's request, these nodes provide the encrypted secret sharing share to the user.

[0072] The user calls the credential recovery interface of the trusted server and inputs no less than n secret shares to the trusted server. The trusted server calls the verification module for processing, and the module outputs a confidence level to determine whether the user has permission to use it.

[0073] Specifically, the trusted server determines the authority of the node as follows: if the model presents backdoor behavior with high confidence, it means that the user has the right to use the model; if the user only obtains a permission share that does not meet the threshold number (that is, only partial authorization) or directly forges the permission share, the model will present backdoor behavior with low confidence and the user will not be able to pass the verification.

[0074] Specifically, the calculation formula for the byte ratio of the user's authorization certificate during verification is:

[0075]

[0076] Among them, S is the byte ratio of the user's authorization certificate when verifying, T1 is the byte size of the user's authorization certificate when verifying, and T2 is the original byte size of the user's authorization certificate.

[0077] Specifically, the reduced transmission rate of the authorization credential is determined by the difference between the byte ratio of the user's authorization credential during verification and a preset second ratio.

[0078] Specifically, the specific process of determining the transmission rate of the authorization certificate by the difference between the byte ratio of the user's authorization certificate during verification and the preset second ratio is:

[0079] If ΔQ≤ΔQ0, the control module adjusts the transmission rate of the authorization certificate to the first transmission rate using a preset second transmission rate adjustment coefficient;

[0080] If ΔQ>ΔQ0, the control module uses a preset first transmission rate adjustment coefficient to adjust the transmission rate of the authorization certificate to a second transmission rate;

[0081] Wherein, the preset first transmission rate adjustment coefficient is smaller than the preset second transmission rate adjustment coefficient.

[0082] Optionally, a preferred embodiment of the preset ratio difference ΔQ0 is ΔQ0=0.1.

[0083] Specifically, the preset first transmission rate adjustment coefficient is recorded as α1, and α1 is set to 0.8. The preset second transmission rate adjustment coefficient is recorded as α2, and α2 is set to 0.9. The transmission rate of the authorization certificate is recorded as V, wherein 0<α1<α2<1. The adjusted transmission rate of the authorization certificate is recorded as V', and V' is set to V×(1+αi) / 2, wherein αi is the preset i-th transmission rate adjustment coefficient, and i is set to 1, 2.

[0084] In implementation, the system of the present invention adjusts the transmission rate of the authorization credential by setting a preset first ratio and a preset second ratio. Due to fluctuations in the network on the user side, errors occur in the data transmission process, resulting in data loss in the authorization credential, which in turn causes the system to fail to recognize the credential, resulting in authorization failure. By reducing the transmission rate of the authorization credential, the probability of data transmission errors caused by network fluctuations is reduced, thereby further achieving improved operational security for the multi-party authorization system.

[0085] Specifically, the control module is connected to the secret share generator, and is used to make a secondary determination that the stability of multi-party authorization does not meet the requirement when the average number of secret shares generated meets the first number condition or the second number condition, and to make a preliminary determination that the validity of the authorization verification does not meet the requirement when the average number of secret shares generated only meets the second number condition, and to make a secondary determination on whether the validity of the authorization verification meets the requirement according to the byte difference before and after the decoding of the authorization credential;

[0086] The control module is connected to the secret share generator, and is used to increase the operating memory ratio of the secret share generator when the average generated quantity of the secret share only meets the first quantity condition;

[0087] Among them, the first quantity condition is that the average generated number of secret shares is greater than the preset first number and less than or equal to the preset second number; the first quantity condition is that the average generated number of secret shares is greater than the preset second number.

[0088] Optionally, a preferred embodiment of the preset first quantity P1 is P1=10, and a preferred embodiment of the preset second quantity P2 is P2=12.

[0089] Specifically, the average number of secret shares generated is recorded as P, the difference between the average number of secret shares generated and the preset first number is recorded as ΔP, and ΔP=P-P1 is set.

[0090] Specifically, the average number of secret shares generated is calculated as:

[0091]

[0092] Among them, P is the average number of secret shares generated, X a is the number of secret shares generated in the ath generation cycle, n is the number of generation cycles, and n is a natural number greater than or equal to 1.

[0093] In implementation, the system of the present invention performs a secondary determination on the stability of multi-party authorization by setting a preset first number and a preset second number, thereby reducing the impact of a decrease in the operational security of the multi-party authorization system due to inaccurate secondary determination of the stability of the multi-party authorization, and further improving the operational security of the multi-party authorization system.

[0094] Specifically, the increased operating memory ratio of the secret share generator is determined by the difference between the average number of secret shares generated and the preset first number.

[0095] Specifically, the specific process of determining the running memory proportion of the secret share generator by the difference between the average number of secret shares generated and the preset first number is:

[0096] If ΔP≤ΔP0, the control module uses a preset first proportion adjustment coefficient to adjust the operating memory proportion of the secret share generator to a first proportion;

[0097] If ΔP>ΔP0, the control module uses a preset second proportion adjustment coefficient to adjust the operating memory proportion of the secret share generator to the second proportion;

[0098] Among them, the preset first proportion adjustment coefficient is smaller than the preset second proportion adjustment coefficient.

[0099] Optionally, a preferred embodiment of the preset quantity difference ΔP0 is ΔP0=3.

[0100] Specifically, the preset first proportion adjustment coefficient is recorded as β1, and β1 is set to 1.2. The preset second proportion adjustment coefficient is recorded as β2, and β2 is set to 1.4. The running memory proportion of the secret share generator is recorded as H, wherein 1<β1<β2. The running memory proportion of the secret share generator after adjustment is recorded as H', and H'=H×(1+2βj) / 3, wherein βj is the preset j-th proportion adjustment coefficient, and j is set to 1, 2.

[0101] In implementation, the system of the present invention adjusts the running memory ratio of the secret share generator by setting a preset first number and a preset second number. Since the secret share generator may need to rely on other components or services, incompatible changes may be introduced when the components are updated, resulting in incompatibility between the secret share generator and the components, which in turn causes the generator to be unable to generate the correct number of shares. By increasing the running memory ratio of the secret share generator, the resource utilization of the generator is improved, the impact of problems caused by incompatibility is reduced, and the operational security of the multi-party authorization system is further improved.

[0102] Specifically, the control module is connected to the secret share generator and the model training unit respectively, and is used to secondarily determine that the validity of the authorization verification does not meet the requirements when the byte difference before and after decoding the authorization credential is greater than a preset difference, and to increase the training cycle of the federal model.

[0103] Optionally, a preferred embodiment of the preset difference amount Y0 is Y0=200 Byte.

[0104] Specifically, the byte difference before and after the authorization certificate is decoded is recorded as Y, and the difference between the byte difference before and after the authorization certificate is decoded and the preset difference is recorded as ΔY, and ΔY=Y0-Y is set.

[0105] Specifically, the calculation formula for the byte difference before and after the authorization credential is decoded is:

[0106] Y=|E1-E2|

[0107] Among them, Y is the byte difference before and after the authorization certificate is decoded, E1 is the byte difference before the authorization certificate data is decoded, and E2 is the byte difference after the authorization certificate data is decoded.

[0108] In implementation, the system of the present invention performs a secondary determination on the validity of the authorization verification by setting a preset difference amount, thereby reducing the impact of the reduced operational security of the multi-party authorization system due to inaccurate secondary determination of the validity of the authorization verification, and further improving the operational security of the multi-party authorization system.

[0109] Specifically, the training cycle of the increased federated model is determined by the difference between the byte difference before and after the decoding of the authorization credential and the preset difference.

[0110] Specifically, the specific process of determining the training cycle of the federation model by the difference between the byte difference before and after the authorization credential is decoded and the preset difference is as follows:

[0111] If △Y≤△Y0, the control module uses a preset first cycle adjustment coefficient to adjust the training cycle of the federated model to the first cycle;

[0112] If △Y>△Y0, the control module uses a preset second cycle adjustment coefficient to adjust the training cycle of the federated model to the second cycle;

[0113] Wherein, the preset first period adjustment coefficient is smaller than the preset second period adjustment coefficient.

[0114] Optionally, a preferred embodiment of the preset difference value △Y0 is △Y0=50Byte.

[0115] Specifically, the preset first cycle adjustment coefficient is recorded as γ1, and γ1 is set to 1.1. The preset second cycle adjustment coefficient is recorded as γ2, and γ2 is set to 1.3. The training cycle of the federal model is recorded as L, where 1<γ1<γ2. The training cycle of the adjusted federal model is recorded as L', and L'=L×(1+3γm) / 4, where γm is the preset m-cycle adjustment coefficient, and m is set to 1, 2.

[0116] In implementation, the system of the present invention adjusts the training cycle of the federal model by setting a preset difference amount. Since the character encoding of the credential data may be different, the credential decoder may not process it correctly, resulting in parsing errors, improper handling of the boundary conditions of the credential data, and further causing problems with the credential decoder, resulting in unsuccessful verification. By increasing the training cycle of the federal model, the model further optimizes the credential data, reducing the impact of verification failures caused by different character encodings of the credential data, and further achieving improved operational security for the multi-party authorization system.

[0117] Specifically, the control module is connected to the secret share generator and the model training unit, respectively, to determine that the processing efficiency of the federal model does not meet the requirements when the average training time of the federal model for the same amount of data is greater than the preset training time, and to reduce the training cycle of the increased federal model.

[0118] Optionally, a preferred embodiment of the preset training duration R0 is R0=3s.

[0119] Specifically, the average training time of the federated model for the same amount of data is denoted as R, and the difference between the average training time of the federated model for the same amount of data and the preset training time is denoted as △R, and △R=R-R0 is set.

[0120] Specifically, the calculation formula for the average training time of the federated model for the same amount of data is:

[0121]

[0122] Among them, R is the average training time of the federated model for the same amount of data, B f is the training duration of the federated model for the same amount of data in the f-th running cycle, c is the number of running cycles, and c is a natural number greater than or equal to 1.

[0123] In implementation, the system of the present invention determines the processing efficiency of the federal model by setting a preset training duration, thereby reducing the impact of the decline in the operational security of the multi-party authorization system due to inaccurate judgment of the processing efficiency of the federal model, and further improving the operational security of the multi-party authorization system.

[0124] Specifically, the reduced training cycle of the federated model is determined by the difference between the average training time of the federated model for the same amount of data and the preset training time.

[0125] Specifically, the specific process of determining the training cycle of the federated model by the difference between the average training time of the federated model for the same amount of data and the preset training time is as follows:

[0126] If △R≤△R0, the control module uses a preset fourth period secondary adjustment coefficient to secondary adjust the training period of the federated model to the third period;

[0127] If ΔR>ΔR0, the control module uses a preset third period secondary adjustment coefficient to secondary adjust the training period of the federated model to the fourth period;

[0128] Among them, the third period quadratic adjustment coefficient is smaller than the fourth period quadratic adjustment coefficient.

[0129] Optionally, a preferred embodiment of the preset training duration difference ΔR0 is ΔR0=1s.

[0130] Specifically, the preset third period quadratic adjustment coefficient is recorded as γ3, and γ3=0.7 is set. The preset fourth period quadratic adjustment coefficient is recorded as γ4, and γ4=0.8 is set, wherein 0<β3<β4<1. The training period of the adjusted federal model is recorded as L”, and L”=L'×(1+3γw) / 4 is set, wherein γw is the preset w-th period quadratic adjustment coefficient, and w=3,4 is set.

[0131] In implementation, the system of the present invention performs secondary adjustment on the training cycle of the federated model by setting a preset training duration. Since an excessively long training cycle will extend the training duration, users will need to wait longer to transmit the local model to the central server, thereby extending the overall training duration. By reducing the training cycle of the federated model, the processing efficiency of the federated model is improved, further achieving an improvement in the operational security of the multi-party authorization system.

[0132] Example 1

[0133] This embodiment 1 uses a multi-party authorization system for federated learning to provide services to multiple federated nodes, and the control module adjusts the transmission rate of the authorization credential according to the difference between the byte ratio of the user's authorization credential during verification and the preset second ratio, wherein the preset first transmission rate adjustment coefficient is recorded as α1, the preset second transmission rate adjustment coefficient is recorded as α2, and the transmission rate of the authorization credential is recorded as V, wherein 0<α1<α2<1, α1=0.8, α2=0.9, △Q0=0.1, V=20Mbps, Q=0.9, Q2=0.75, △Q=Q-Q2.

[0134] In the first embodiment, ΔQ=0.15 is obtained, the control module determines that ΔQ>ΔQ0 and uses the preset first transmission rate adjustment coefficient to adjust the transmission rate of the authorization certificate to the second transmission rate, and calculates that V'=20Mbps×(1+0.8) / 2=18Mbps.

[0135] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. A multi-party authorization system for federated learning, characterized by: include: A model aggregation module, used for processing the local models of the participants to output a federated model, including a model training unit used for training the local models of several participants to output a model to be aggregated and a model aggregation unit connected to the model training unit to perform an aggregation operation on the corresponding number of models to be aggregated to output a federated model; A verification module, which is connected to the model aggregation module and is used to verify the user's use rights according to the authorization certificate of the participant, including a secret share generator for generating corresponding secret shares according to the backdoor information code of the participant, an authorization certificate distributor connected to the secret share generator for distributing the secret shares to the participant, and a certificate restorer for verifying the user's use rights; A control module, which is connected to the model aggregation module and the verification module respectively, and is used to adjust the transmission rate of the authorization certificate or determine the operating memory ratio of the secret share generator according to the average number of secret shares generated when it is determined that the operating security of the multi-party authorization system does not meet the requirements according to the byte ratio of the user's authorization certificate, and to initially adjust the training cycle of the federal model according to the byte difference before and after the authorization certificate is decoded, and to secondary adjust the training cycle of the federal model according to the average training time of the federal model for the same amount of data after the federal model runs a single cycle with the training cycle after the initial adjustment.

2. The multi-party authorization system for federated learning according to claim 1, characterized in that: The model aggregation module also includes a trusted server for coordinating the federated model process.

3. The multi-party authorization system for federated learning according to claim 1, characterized in that: The verification module also includes a backdoor information encoder for generating the backdoor information code.

4. The multi-party authorization system for federated learning according to claim 1, characterized in that: The process of verifying the user's usage rights is that the user provides the secret share to the trusted server. The trusted server first decrypts the secret share, and then recovers the secret share through secret sharing to obtain the backdoor information code. The backdoor information code is used as the input of the credential decoder. The credential decoder generates a set of backdoor instances for verification based on the input backdoor information code and returns it. The trusted server inputs the backdoor instance into the model and returns the confidence level, and judges the user's usage rights based on the confidence level.

5. The multi-party authorization system for federated learning according to claim 1, characterized in that: The control module is connected to the authorization verification unit, and is used to calculate the byte volume ratio of the user's authorization credential when verifying according to the byte volume of the user's authorization credential when verifying and the original byte volume of the user's authorization credential when verifying, and to determine that the operation security of the multi-party authorization system does not meet the requirements when the byte volume ratio of the user's authorization credential when verifying satisfies the first ratio condition or the second ratio condition; The control module is connected to the secret share generator, and is used to preliminarily determine that the multi-party authorization stability does not meet the requirements when the byte ratio of the user's authorization credential during the verification only meets the first ratio condition, and to perform a secondary determination on whether the multi-party authorization stability meets the requirements according to the average number of secret shares generated; The control module is connected to the credential transmission unit, and is used to reduce the transmission rate of the authorization credential when the byte ratio of the user's authorization credential during the verification only meets the second ratio condition; Among them, the first proportion condition is that the byte ratio of the user's authorization credentials when performing verification is greater than the preset first ratio and less than or equal to the preset second ratio; the second proportion condition is that the byte ratio of the user's authorization credentials when performing verification is greater than the preset second ratio.

6. The multi-party authorization system for federated learning according to claim 5, characterized in that: The calculation formula for the byte ratio of the user's authorization certificate during verification is: Among them, S is the byte ratio of the user's authorization certificate when verifying, T1 is the byte size of the user's authorization certificate when verifying, and T2 is the original byte size of the user's authorization certificate.

7. The multi-party authorization system for federated learning according to claim 6, characterized in that: The reduced transmission rate of the authorization credential is determined by the difference between the byte ratio of the user's authorization credential during verification and a preset second ratio.

8. The multi-party authorization system for federated learning according to claim 7, characterized in that: The control module is connected to the secret share generator, and is used to make a secondary determination that the stability of multi-party authorization does not meet the requirement when the average number of secret shares generated meets the first number condition or the second number condition, and to make a preliminary determination that the validity of the authorization verification does not meet the requirement when the average number of secret shares generated only meets the second number condition, and to make a secondary determination on whether the validity of the authorization verification meets the requirement according to the byte difference before and after the decoding of the authorization credential; The control module is connected to the secret share generator, and is used to increase the operating memory ratio of the secret share generator when the average generated quantity of the secret share only meets the first quantity condition; The increased operating memory ratio of the secret share generator is determined by the difference between the average number of secret shares generated and the preset first number; Among them, the first quantity condition is that the average generated number of secret shares is greater than the preset first number and less than or equal to the preset second number; the first quantity condition is that the average generated number of secret shares is greater than the preset second number.

9. The multi-party authorization system for federated learning according to claim 8, characterized in that: The control module is connected to the secret share generator and the model training unit respectively, and is used to secondarily determine that the validity of the authorization verification does not meet the requirements when the byte difference before and after decoding the authorization credential is greater than a preset difference, and to increase the training cycle of the federation model; The training cycle of the increased federated model is determined by the difference between the byte difference before and after the decoding of the authorization credential and the preset difference.

10. The multi-party authorization system for federated learning according to claim 9, characterized in that: The control module is connected to the secret share generator and the model training unit respectively, and is used to determine that the processing efficiency of the federation model does not meet the requirements when the average training time of the federation model for the same amount of data is greater than the preset training time, and to reduce the training cycle of the increased federation model; The reduced training cycle of the federated model is determined by the difference between the average training time of the federated model for the same amount of data and the preset training time.

Images