CN-121980623-A - Verification method and device for federal learning model high-dimensional heterogeneous parameters

CN121980623ACN 121980623 ACN121980623 ACN 121980623ACN-121980623-A

Abstract

The invention relates to the technical field of federal learning and blockchain crossing, in particular to a method and a device for verifying high-dimensional heterogeneous parameters of a federal learning model. The intelligent contract comprises the steps that an intelligent contract selects mechanisms meeting the trust level requirement from mechanisms participating in multi-centralized federal learning as verification nodes, the mechanisms upload local model high-dimensional heterogeneous parameters to a blockchain to trigger an intelligent contract verification process, the verification nodes verify all the local model high-dimensional heterogeneous parameters, the intelligent contract collects initial verification results of the verification nodes and determines final consensus verification results through a consensus mechanism, the intelligent contract writes the final consensus verification results, parameter details and node verification records into the blockchain to store the evidence, and the parameters which pass the authorization verification participate in global model aggregation and reject the parameters which do not pass the authorization verification. The invention can solve the problems that federal learning relies on a third party to carry out model parameter aggregation, model parameter verification lacks a decentralization mechanism, and verification results are not tamperable and stored.

Inventors

TANG SONG
CUI NENGXI
Gai Suli
ZHANG TIANLIANG
ZHAO JINGXIN

Assignees

河北省科学院应用数学研究所

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. The method for verifying the high-dimensional heterogeneous parameters of the federal learning model is characterized by comprising the following steps of: The intelligent contract selects an organization meeting the trust level requirement from all organizations participating in multi-centralized federal learning as a verification node to form a verification node set; After the local model training is completed by each mechanism, uploading the high-dimensional heterogeneous parameters of the local model to a blockchain, and triggering an intelligent contract verification process; the verification nodes verify all the uploaded local model high-dimensional heterogeneous parameters, the intelligent contract collects initial verification results of all the verification nodes, and a final consensus verification result is determined through a multi-node consensus mechanism; and the intelligent contract writes the final consensus verification result, the parameter detail and the verification records of all nodes into a blockchain memory card, and the parameters passing the authorization verification participate in global model aggregation to reject the parameters not passing the authorization verification.
2. The method for verifying high-dimensional heterogeneous parameters of a federal learning model according to claim 1, wherein the intelligent integration selects an entity meeting a trust level requirement from among entities participating in multi-centralized federal learning as a verification node, and comprises: the intelligent contract obtains the historical verification accuracy, the calculation force contribution value and the compliance score of each mechanism participating in multi-centralized federal learning; Adopting an improved PBFT algorithm, and calculating the trust value of each mechanism according to the historical verification accuracy, the calculation force contribution value and the compliance score; and when the trust level value is greater than or equal to a trust level threshold value, determining that the mechanism corresponding to the trust level value is a verification node.
3. The method for verifying high-dimensional heterogeneous parameters of a federal learning model according to claim 2, wherein calculating trust values of each institution based on the historical verification accuracy, the calculation force contribution value and the compliance score using a modified PBFT algorithm comprises: According to Calculating the trust value of each organization; Wherein, the Represent the first The trust level value of the individual authorities, 、、 Respectively are weight coefficients, and , Represent the first The history of the individual institutions verifies the accuracy rate, Represent the first The calculated force contribution value of the individual institution, Represent the first Compliance scores for each institution.
4. A method of validating federal learning model high-dimensional heterogeneous parameters according to any one of claims 1-3, wherein the validating node validates all local model high-dimensional heterogeneous parameters uploaded, comprising: And the verification node sequentially executes parameter anomaly dimension pre-screening, parameter standardization and robustness conversion, multidimensional consistency test and fine granularity detection on all uploaded local model high-dimensional heterogeneous parameters.
5. The method for verifying high-dimensional heterogeneous parameters of a federal learning model according to claim 4, wherein the parameter anomaly dimension pre-screening is to perform anomaly score calculation on each dimension of the high-dimensional heterogeneous parameters of the local model by adopting an improved parallel isolated forest algorithm, and when the anomaly score is greater than a preset anomaly threshold, marking the current dimension as an anomaly dimension and deleting the current dimension; The improved parallel isolated forest algorithm is based on the existing isolated forest algorithm, and introduces a time attenuation weight and a sparse self-adaptive sampling mechanism; the calculation formula of the improved parallel isolated forest algorithm is as follows: ; Wherein, the Represent the first The anomaly score for a dimension is determined, Represent the first The first of the mechanism upload A set of parameters for the dimensions, The time decay weight is represented as a function of time, , Representing the difference between the time stamp of the current parameter upload and the system reference time, Indicating the decay period of the preset time, The sparse adaptive sampling rate is represented by, , Represent the first The first of the mechanism upload The sum of the absolute values of all the parameters of the dimension, Representing the maximum value of the sum of the absolute values of all the dimensional parameters.
6. The method for verifying high-dimensional heterogeneous parameters of a federal learning model according to claim 5, wherein the parameter normalization and robustness are converted into parameters of the residual dimension after the parameter anomaly dimension is pre-screened, and the normalization is performed by adopting the absolute deviation of median and median; the formula of the normalization process is: ; Wherein, the Represent the first The first of the mechanism upload The standardized parameters of the dimensional parameters are defined, Represent the first The first of the mechanism upload Parameters of the dimensions, Represent the first The first of the mechanism upload The median of the parameter set of the dimension, Represent the first The first of the mechanism upload The median absolute deviation of the parameters of the dimension.
7. The method for verifying high-dimensional heterogeneous parameters of a federal learning model according to claim 6, wherein the multi-dimensional consistency test is to calculate a mahalanobis distance between a standardized parameter and a global parameter by adopting an improved mahalanobis distance introducing regularized covariance estimation and feature selection mechanism, and when the mahalanobis distance is greater than a preset dynamic threshold, determining the current standardized parameter as an abnormal parameter and marking; The improved mahalanobis distance has the following calculation formula: ; Wherein, the Represent the first The mahalanobis distance of the standardized parameters of the individual institution from the global parameters, Represent the first The standardized parameters of the individual institutions are set up, The global parameter mean vector is represented as such, Representing the feature filtered global parameter covariance matrix, The regularization coefficient is represented as a function of the regularization coefficient, Representing the identity matrix.
8. The method for verifying high-dimensional heterogeneous parameters of a federal learning model according to claim 7, wherein the fine-granularity detection is performed by adopting an optimized local anomaly factor algorithm introducing a dynamic neighborhood selection and weighted distance measurement mechanism, calculating a local anomaly factor for the standardized parameters subjected to multi-dimensional consistency test, and determining that the current standardized parameters are anomalous in the local neighborhood when the local anomaly factor is greater than a preset anomaly factor threshold; The calculation formula of the optimization local anomaly factor algorithm is as follows: ; Wherein, the Represent the first The optimized local anomaly factors of the standardized parameters of the individual institutions, Represent the first Standardized parameters of individual institutions A set of neighbors that are close together, Represent the first Standardized parameters of individual institutions The number of elements in the neighbor set, The distance weight is represented as a function of the distance, Representing the first in a neighbor set The normalized parameter vector of the individual institution, Representing the neighborhood scale parameters, Representing the first in a neighbor set The locally attainable density of the individual mechanisms, Represent the first Local reachable densities of standardized parameters of the individual institutions.
9. The method for validating high-dimensional heterogeneous parameters of federal learning model of claim 8, wherein determining the final consensus validation result by a multi-node consensus mechanism comprises: According to Determining a final consensus verification result; Wherein, the Represent the first Final consensus verification of standardized parameters of the authorities, Represent the first The similarity weight of each verification node is determined, , Representing the total number of authentication nodes, The similarity-enhancement factor is represented by a set of similarity-enhancement factors, Represent the first The degree of consistency of the authentication results of each authentication node with the authentication results of other authentication nodes, Representing the historical verification trust coefficient of the verification node, Represent the first Verification node pair number Initial validation of standardized parameters of the individual institution, Represents the contribution adjustment coefficient and, Represent the first The contribution weights of the nodes are verified.
10. The utility model provides a verifying attachment of high dimension heterogeneous parameter of federal learning model which characterized in that includes: The intelligent contract module is used for selecting an organization meeting the trust level requirement from all organizations participating in multi-centralized federal learning as a verification node to form a verification node set; each mechanism terminal is used for uploading the high-dimensional heterogeneous parameters of the local model to the blockchain after the local model training is completed, and triggering an intelligent contract verification process; The verification node terminal is used for verifying all the uploaded local model high-dimensional heterogeneous parameters and sending an initial verification result to the intelligent contract module; The intelligent contract module is also used for collecting initial verification results of all verification nodes, determining final consensus verification results through a multi-node consensus mechanism, writing the final consensus verification results, parameter details and node verification records into a blockchain memory card, authorizing the verification passed parameters to participate in global model aggregation, and rejecting the failed parameters.

Description

Verification method and device for federal learning model high-dimensional heterogeneous parameters Technical Field The invention relates to the technical field of federal learning and blockchain crossing, in particular to a method and a device for verifying high-dimensional heterogeneous parameters of a federal learning model. Background Digital economy is rapidly developed, data cooperation requirements across institutions are increasingly urgent, and data privacy protection becomes a core requirement. Federal learning becomes a key technology for cross-domain data value mining by virtue of the characteristic that data is available and invisible, and is widely applied to the collaborative fields requiring privacy protection, such as finance, medical treatment, credit investigation and the like. The existing federal learning system generates a global model by aggregating local model parameters of each participating mechanism, part of schemes attempt to introduce a blockchain technology, and assist cooperation by means of the decentralization characteristic of the blockchain technology, but the two are not fused deeply. Although the intelligent contracts of the blockchain have automatic execution capability, the intelligent contracts are not effectively applied to the model parameter collaborative verification scene. The method has the obvious defects that federal learning relies on a third party to aggregate model parameters, single-point faults and model pollution risks exist, model parameter verification lacks a decentralization mechanism, authenticity is difficult to guarantee, data discrimination or model poisoning attack is easily encountered, verification results are not falsified, verification and tracing are difficult, and a robust verification scheme for high-dimensional heterogeneous parameters is lacking. Disclosure of Invention The embodiment of the invention provides a method and a device for verifying high-dimensional heterogeneous parameters of a federal learning model, which are used for solving the problems that in the prior art, federal learning relies on a third party to perform model parameter aggregation, a model parameter verification lacks a decentralization mechanism, and a verification result is not tamperable and evidence-storing. In a first aspect, an embodiment of the present invention provides a method for verifying high-dimensional heterogeneous parameters of a federal learning model, including: The intelligent contract selects an organization meeting the trust level requirement from all organizations participating in multi-centralized federal learning as a verification node to form a verification node set; After the local model training is completed by each mechanism, uploading the high-dimensional heterogeneous parameters of the local model to a blockchain, and triggering an intelligent contract verification process; the verification nodes verify all the uploaded local model high-dimensional heterogeneous parameters, the intelligent contract collects initial verification results of all the verification nodes, and a final consensus verification result is determined through a multi-node consensus mechanism; and the intelligent contract writes the final consensus verification result, the parameter detail and the verification records of all nodes into a blockchain memory card, and the parameters passing the authorization verification participate in global model aggregation to reject the parameters not passing the authorization verification. In one possible implementation manner, the smart contract selects, among institutions participating in multi-centralized federal learning, an institution meeting the trust level requirement as a verification node, and includes: the intelligent contract obtains the historical verification accuracy, the calculation force contribution value and the compliance score of each mechanism participating in multi-centralized federal learning; Adopting an improved practical Bayesian and busy-tolerant algorithm (PRACTICAL BYZANTINE FAULT TOLERANCE, PBFT) algorithm, and calculating the trust value of each institution according to the historical verification accuracy, the calculation force contribution value and the compliance score; and when the trust level value is greater than or equal to a trust level threshold value, determining that the mechanism corresponding to the trust level value is a verification node. In one possible implementation, a modified PBFT algorithm is adopted to calculate a trust value of each institution according to the historical verification accuracy, the calculation force contribution value and the compliance score, including: According to Calculating the trust value of each organization; Wherein, the Represent the firstThe trust level value of the individual authorities,、、Respectively are weight coefficients, and,Represent the firstThe history of the individual institutions verifies the accuracy rate,Represent the firstThe calc