CN116777294A - Crowd-sourced quality safety assessment method based on federal learning under assistance of blockchain - Google Patents
Crowd-sourced quality safety assessment method based on federal learning under assistance of blockchain Download PDFInfo
- Publication number
- CN116777294A CN116777294A CN202310807841.2A CN202310807841A CN116777294A CN 116777294 A CN116777294 A CN 116777294A CN 202310807841 A CN202310807841 A CN 202310807841A CN 116777294 A CN116777294 A CN 116777294A
- Authority
- CN
- China
- Prior art keywords
- task
- blockchain
- quality
- data
- crowdsourcing platform
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 93
- 238000011156 evaluation Methods 0.000 claims abstract description 31
- 238000013441 quality evaluation Methods 0.000 claims abstract description 25
- 238000012545 processing Methods 0.000 claims abstract description 21
- 238000001303 quality assessment method Methods 0.000 claims description 58
- 230000008569 process Effects 0.000 claims description 36
- 238000009826 distribution Methods 0.000 claims description 24
- 238000005516 engineering process Methods 0.000 claims description 20
- 238000004364 calculation method Methods 0.000 claims description 19
- 238000012795 verification Methods 0.000 claims description 16
- 238000012549 training Methods 0.000 claims description 14
- 238000005520 cutting process Methods 0.000 claims description 5
- 230000000977 initiatory effect Effects 0.000 claims description 4
- 230000006870 function Effects 0.000 claims description 3
- 230000035945 sensitivity Effects 0.000 claims description 3
- 230000004931 aggregating effect Effects 0.000 claims description 2
- 238000011157 data evaluation Methods 0.000 claims description 2
- 230000007704 transition Effects 0.000 claims description 2
- 229910052723 transition metal Inorganic materials 0.000 claims description 2
- 238000010200 validation analysis Methods 0.000 claims 1
- 230000005540 biological transmission Effects 0.000 description 12
- 230000006399 behavior Effects 0.000 description 4
- 230000007246 mechanism Effects 0.000 description 4
- 238000000586 desensitisation Methods 0.000 description 3
- 238000013210 evaluation model Methods 0.000 description 3
- 238000012854 evaluation process Methods 0.000 description 3
- 238000002372 labelling Methods 0.000 description 3
- 230000002776 aggregation Effects 0.000 description 2
- 238000004220 aggregation Methods 0.000 description 2
- 238000012550 audit Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000001010 compromised effect Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012358 sourcing Methods 0.000 description 2
- 241000269350 Anura Species 0.000 description 1
- 241000282472 Canis lupus familiaris Species 0.000 description 1
- 241000282994 Cervidae Species 0.000 description 1
- 241000282326 Felis catus Species 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 239000003550 marker Substances 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a crowdsourcing quality safety assessment method based on federation learning under the assistance of a blockchain, wherein a task publisher publishes task information on a crowdsourcing platform according to own requirements, selects a corresponding federation learning model, calculates evidence information and uploads the evidence information to the blockchain; task executors select the task of interest to accept, calculate evidence information and upload the evidence information to the blockchain; task executors complete tasks according to task demand description to obtain task data, train a federal learning model by using the task data, encrypt the task data and send the task data to task publishers, and upload the task data to a blockchain by utilizing differential privacy and noise model parameters and calculating evidence information; and the task publisher performs quality evaluation on task data submitted by the task executor and uploads an evaluation ranking result to the blockchain. According to the invention, based on a task model of a specific task type trained by federal learning, important information is recorded by using a block chain, so that a result has traceability and verifiability; meanwhile, the differential privacy is adopted to carry out noise processing on the model updating parameters, so that an attacker is effectively prevented from reversely deducing the original data of the task executor according to the model parameters.
Description
Technical Field
The invention relates to a crowdsourcing service data security technology, in particular to a crowdsourcing quality security assessment method based on federal learning under the assistance of a blockchain.
Background
Crowdsourcing is a business mode of distributing work tasks to a large number of independent individuals for completion through an internet platform, the crowdsourcing platform usually connects task issuers and task executors together to enable the task executors to freely select tasks or match tasks according to requirements of the task executors and the task issuers, the task executors participating in the tasks submit work results according to task requirements, the task issuers can audit and pay corresponding rewards on the results submitted by the task executors, and the crowdsourcing has become a diversified, global business mode due to the freedom and low cost and is continuously developed and developed.
The task execution quality evaluation is a key stage in crowdsourcing application, and refers to judging whether a task meets requirements or not by evaluating and auditing a work result submitted by a task executor and giving corresponding rewards. Quality assessment is critical in crowd sourcing because task performers in crowd sourcing are publicly recruited to the social public, and the skill, knowledge and experience levels of the performers may vary, as may the quality of the completion. In order to ensure the quality of the task, the crowdsourcing platform needs to evaluate and audit the working result of the task performer, select task data with higher quality as a final result, and give corresponding rewards or rewards. The good quality assessment algorithm is beneficial to more fairly exciting task performers and improving the satisfaction of task publishers.
How to evaluate the quality of crowd-sourced task execution (data) has become a research hotspot. Common quality assessment algorithms are a double assessment method, a random sampling method, a grade assessment method, an expert assessment method and a mutual assessment method. The double evaluation method requires a task publisher to distribute two or more task performers to complete the same task, which may increase the cost and time of the task publisher, while if the results of the two task performers are not very different, the task publisher may need to distribute the task again in order to select the best results. The random sampling method only evaluates sampled work results and cannot guarantee the quality of non-sampled results. If the quality of the non-sampled results is poor, the quality of the overall task may be affected. The grade evaluation method decides whether to pass the task according to the score of the task executor and gives corresponding consideration, and the problem that the scoring standard is inconsistent possibly exists; that is, different raters may give different scores to the same work result, making it difficult for the task publisher to determine the final result. Expert evaluation requires the employment of professionals or professionals within the industry, increasing the cost of the task publisher. At the same time, the expert's assessment may be affected by his personal experience and bias. The mutual evaluation method may have a problem that the relationship between crowdsourcing personnel is poor or mutual deliberate scoring is performed, resulting in the unfairness of evaluation.
From the above description, the prior art has the problems of high cost, non-traceability of the process and single evaluation index. More importantly, when the task execution results relate to the privacy of the performer, such as health data, location data, etc., and when any of the participants may exhibit malicious behavior, the assessment algorithm must protect the privacy of the data assessors.
The privacy protection method in the crowdsourcing quality assessment stage mainly comprises data encryption, data anonymization, data desensitization, differential privacy and access control. Data encryption is a common privacy protection method, and in the crowdsourcing quality evaluation stage, an encryption algorithm (such as homomorphic encryption, secure multiparty calculation and the like) can be used for encrypting user data so as to ensure the security of the data in the transmission and processing processes; thus, even if the data is intercepted by an attacker, the original information cannot be directly acquired. Data anonymization is a method by which data is processed such that sensitive information in the data cannot be associated with a particular individual; processing user data using data anonymization techniques (e.g., k-anonymization, l-diversity, etc.) to protect user privacy; thus, even if the data is compromised during the evaluation, the attacker cannot directly identify the individual in the data. The data desensitization is a method for reducing the risk of data leakage by carrying out operations such as replacement, shielding or deletion on sensitive information in the data; processing the user data by using a data desensitization technology to protect user privacy; thus, even if data is compromised during the evaluation process, an attacker cannot obtain complete sensitive information. Differential privacy is a technology for protecting individual privacy in the process of data release and analysis, and the differential privacy technology is used for carrying out noise processing on data so as to protect user privacy; thus, even if data is leaked during the evaluation, an attacker cannot accurately acquire individual information. Access control includes restricting access to sensitive data, granting access to data only to specific users or roles, monitoring and auditing data access, and the like. This effectively prevents unauthorized users from accessing and using sensitive data.
With the development of blockchain technology in recent years, researchers have also actively explored the use of blockchain technology to solve the problems of trust dependence, single point failure, service reliability and the like of the traditional crowdsourcing service on the crowdsourcing platform. The decentralized, publicly transparent and non-tamperable nature of the blockchain can provide more fair and transparent crowdsourcing services, including task allocation, quality assessment and reward allocation, effectively preventing the assessment results from being manipulated by malicious participants.
Blockchains can be applied to the quality assessment phase of crowdsourcing platforms in combination with encryption, smart contract, artificial intelligence and reputation system technologies, however, there are still some problems in practical applications: for example, patent CN107871196a discloses a crowdsourcing quality assessment method based on a sliding task window, which does not consider protection of identity information of a task publisher and a task executor, and the design of the sliding task window only considers the task completion condition in a section, and lacks global consideration; such as the CN110472948A method for confirming the quality of task completion through manual auditing, which may lead to inefficiency and high cost; for example, the patent CN112148890a adopts a multi-aspect evaluation mechanism, including self-evaluation, peer evaluation, platform evaluation, etc., to perform multi-dimensional evaluation on task performers to ensure objectivity and accuracy of task completion quality, but the multi-aspect evaluation mechanism adopted by the patent still has a certain subjectivity, and the patent does not relate to specific measures and methods of privacy protection, if sensitive data in evaluation is leaked or tampered, the whole evaluation result will be adversely affected; for example, the patent CN112785162a guarantees the authenticity and credibility of the crowdsourcing task through a blockchain technology, improves the quality and efficiency of the crowdsourcing task, simultaneously weights multiple evaluation indexes to excite enthusiasm and creativity of participants, and improves the operation efficiency of the whole crowdsourcing platform, but how the patent objectively designs and determines the weight of the tokens is not fully discussed, and the evaluation process does not consider the protection of the privacy of the task participants, which may limit the application of the task participants in the large-scale crowdsourcing task; for example, the patent CN112862303a provides a safe and reliable distributed crowdsourcing platform, effectively solves the problems of single-point failure, data leakage and data reliability, obtains more reliable worker nodes through the reputation mechanism of node history behaviors, and further improves the quality of crowdsourcing results.
Disclosure of Invention
The invention aims to: the invention aims to solve the defects in the prior art and provides a crowdsourcing quality safety assessment method based on federal learning under the assistance of a blockchain; the method adopts an information storage mode combining the upper chain and the lower chain, ensures the verifiability of the quality evaluation score of the crowdsourcing task, effectively identifies potential malicious behaviors, considers diversified objective indexes, avoids the inefficiency of artificial subjective evaluation, and has mobility to the tasks of the same type; the federal learning is utilized to enable the evaluation model to be automatically updated in an iterative mode, and cost overhead for designing different evaluation algorithms and maintaining the evaluation model in a quality evaluation stage is reduced.
The technical scheme is as follows: the invention discloses a crowdsourcing quality safety assessment method based on federal learning under the assistance of a blockchain, which relates to a key generation center, a task publisher, a task executor, a crowdsourcing platform and the blockchain, and comprises the following steps:
step (1), system initialization including user registration, parameter generation and key distribution
User registers as task publisher R on crowdsourcing platform ID Or task performer W j ID and j are numbers generated by the crowdsourcing platform for the user; generating initial parameter C by crowdsourcing platform l G, H and TMT ID The method comprises the steps of carrying out a first treatment on the surface of the The crowdsourcing platform maintains q federal learning models M= { M suitable for different task scenes 1 ,m 2 ,…,m q }，C l Initial thresholds, l=1, 2, …, q, used when aggregating data for different models; g and H are base points randomly generated on an elliptic curve and are initial parameters promised by Pedersen; TMT (transition metal-transition metal) ID Task publisher R with successful registration by crowdsourcing platform ID A generated task matching table for recording R ID The published task information is initially empty; the key generation center is the task publisher R with successful registration ID And task performer W j Generating public-private key pairs (PK ID ,SK ID) and (PKj ,SK j )；
Step (2), task publisher R ID Publishing task τ on crowd-sourced platform i ＝(ID,i,m i ,B i ,TR i ,RAlloc i ,Γ i ,κ i ) I is the number of the issuing task, m i Is R ID For task tau i Selected federal learning model, B i For task tau i Remuneration of (TR) i Is tau i Task demand description of R ID There is a need to provide task related information and Pedersen commitments, RAlloc therein i Is tau i And (f) a reward distribution algorithm i Is tau i Task data quality assessment algorithm, κ i Representing task τ i Requiring each W j The number of data submitted;
task publisher R ID Computing task τ i Hash value ζ of (a) i ＝H(τ i ) Using its own private key SK ID For hash value ζ i Signing to obtain psi i ＝Sig(ζ i ,SK ID ) And upload to the blockchain; the crowdsourcing platform is according to R ID Filled-in task information and R ID Is subjected to validity check of related information, such as R ID Whether the filled task information is complete, whether the format is legal or not, R ID If the authority of the issuing task exists, if passing the legitimacy check, the issuing task tau i And at R ID Task matching table TMT of (a) ID Adding an entry { i, W } for recording τ i W represents successful acceptance τ i Is initially empty;
step (3), task executor W j Selecting a task tau of interest i Receiving, wherein the crowdsourcing platform is responsible for recording task matching conditions;
when W is j Accept τ i If τ i Within acceptable time, the crowdsourcing platform will accept the W for the task j Added to task matching table TMT ID In W corresponding to the task list item { i, W }, and calling the intelligent contract to learn model m of federation i Send to W j The method comprises the steps of carrying out a first treatment on the surface of the If task tau i Is not acceptableWithin the time, then W j Accept τ i Failure;
W j accept τ i After success, calculate the acceptance task τ i Hash value ζ 'of (1)' i ＝H(τ i ) Using the private key SK itself j To ζ' i Signing to obtain a signature value psi' i ＝Sig(ζ′ i ,SK j ) And upload to the blockchain; psi' i The signature value obtained by signing the task hash value by a task executor by using a private key is evidence information of a crowdsourcing task matching stage, so that the task hash value is convenient to check in an arbitration stage;
Step (4), W j Completion of task τ according to task demand description i Obtaining a task data set D i,j ＝Task performer W j Will D i,j Kappa of (C) i The hash value of each task data is calculated by splicing the task dataUsing private key SK j Para zeta i,j Signature is carried out to obtain->Task dataDepending on the task, the task publisher will describe the task in the task description when publishing the taskIs explicitly specified; for example, the task publisher publishes a task marked by a picture, and if the size of the picture to be marked is 27 x 27, the task executor submits +.>For a vector of 730 dimensions, the first 729 dimensions represent the noted picture, and the last 1 dimension represents the noted category;
W j based on D i,j Local training federal learning model m i Obtaining the model parameters g to be polymerized i,j For g i,j Cutting to obtainDifferential privacy technology pair is used again>Adding noise to obtain->
W j Task data set D i,j With corresponding task issuers R ID Public key PK of (2) ID Encryption to obtain xi i,j ＝Enc(D i,j ,PK ID ) And send to R ID Will beAnd zeta is i,j Uploading to the blockchain together;
step (5), task publisher R ID To xi i,j Decryption to obtain D' i,j ＝Dec(ξ i,j ,SK ID )＝Will D' i,j Kappa of (C) i The hash value zeta 'of the task data is calculated by splicing the task data' i,j ＝Using self-private key SJ ID To ζ' i,j Signing to obtain ζ' i,j ＝Sig(ζ′ i,j ,SK ID ) And zeta 'is taken as' i,j Uploading to a blockchain;
R ID Invoking a quality assessment algorithm Γ i For D' i,j Quality evaluation is carried out to obtain an evaluation score E i,j The method comprises the steps of carrying out a first treatment on the surface of the The evaluation index includes three items (t i,j ,r j ,l i,j), wherein ti,j For task executionLine time, r j Is W j Reputation value of l i,j Is based on federal learning model m i Calculating the obtained data similarity; r is R ID According to E i,j Ranking from high to low and grouping tasks τ i Quality assessment ranking table QART of (E) i Uploading to a blockchain; at task τ i After completion of R ID The random number { r > of the Pedersen commitment will be verified 1 ,r 2 ,…,r N ,r N+1 ,…,r 2N Upload to blockchain for W j Validating a data evaluation procedure, wherein N is m i Returning to the dimension of the gradient;
step (6), W j According to R ID The model m provided i Pedersen promise { c 1 ,c 2 ,…,c N ,c N+1 ,…c 2N Sum of random numbers { r } and 1 ,r 2 ,…,r N ,r N+1 ,…,r 2N verifying Γ i For evaluation results of (e) such as R ID and Wj When the ranking result is questioned at this stage, an arbitration program is initiated to the crowdsourcing platform, the crowdsourcing platform examines the questionable content, and uploads the arbitration result to the blockchain for disclosure in a specified time, if more than half of the participants agree to the arbitration result, the crowdsourcing platform invokes the intelligent contract to update the quality assessment ranking table QART i ；
Step (7), the crowdsourcing platform calls the intelligent contract to evaluate the ranking list QART according to the latest quality i To execute the reward distribution program according to QART i Upper W j Ranking condition weighted update m i 。
Further, the task demand description TR in step (2) i The internal details of (a) are:
wherein ,representative τ i Start time of (2)(indicating W j Can start to R ID Commit D 1,j Time of (2) and->Representative τ i End time of (indicating R ID No longer accept W j Commit D i,j Time of day) TD i Representative τ i Specific requirements (indicating R ID Hope W j Standard for completing tasks, format of submitted data), { c 1 ,c 2 ,…,c N ,c N+1 ,…c 2N The expression of Pedersen commitment, R ID Tau to be issued for it i Preparing a reference data set SD i (indicating τ i Is included) to assist the quality assessment algorithm in assessing, SD i Format and D of (2) i,j The same; to prevent SD i Leakage R ID R ID No need to publish SD i And also to prevent R ID Random replacement or modification of SD i Thus R is ID Need to provide and SD i Related petersen commitments;
R ID will reference data set SD i Input model m as training data i Obtaining gradient V i,j And to V i,j Providing relevant Pedersen commitments to facilitate task performers W j Subsequent verification of quality assessment score E i,j The method comprises the steps of carrying out a first treatment on the surface of the Introduction of the Pedersen promise ensures both R ID In evaluation D i,j When using the same SD i At the same time utilize the homomorphism of Pedersen promise, W j And can verify R ID Invoking a quality assessment algorithm Γ i Accuracy of (3).
Further, the Pedersen commitment { c ] in the step (2) 1 ,c 2 ,…,c N ,c N+1 ,…c 2N The generation process of the } is as follows:
task publisher R ID Will reference data set SD i As model m i Is input into the training dataset to obtain the gradient
R ID For promise data Calculating corresponding promise value-> wherein r1 ,r 2 ,…,r N ,r N+1 ,…,r 2N Is R ID And (5) generating a random number.
Further, the details of the remuneration distribution algorithm in the step (2) are as follows: r is R ID First determine the rule according to which is to be used for W j Performing remuneration distribution and giving remuneration b i,j The specific calculation formula is as follows:
wherein bi,j Represents W j Completion of task τ i The obtained remuneration; according to W j Quality assessment score E of (2) i,j Proportionally distributing remuneration;
R ID the previously filled reward distribution algorithm or the new reward distribution algorithm can be selected, and the crowd-sourced platform generates an intelligent contract for the new reward distribution algorithm and deploys the intelligent contract on the blockchain after checking the legal.
Further, the step (4) is to perform model parameters g i,j The formula for clipping is as follows:
k is the selected model m i Numbering corresponding to threshold：
Task performer W j Differential privacy technology pairAdding noise to obtain->The formula is as follows:
l is the size of the data set and,is Gaussian noise, sigma 2 Is the noise scale, I is the sensitivity;
task performer W j Will be and />After uploading to the blockchain, task publisher R ID Record W j Submitting task τ i Time of (2)>Judging->Whether or not it is true, if so, W j And if the submission is successful, the submission fails.
Further, the specific process of the step (5) is as follows:
step (5.1), R ID Will beUploading to the blockchain for subsequent traceback, R ID From each W recorded j Task commit time->Calculate each W j Task time of (1)> (assumed task scenario is W j At start time +.>The task can be started after that);
step (5.2), R ID Obtaining the latest reputation value r of each task executor from a blockchain j The method comprises the steps of carrying out a first treatment on the surface of the The crowdsourcing platform uses blockchain for each W j Maintaining a reputation value r j The value depends on W j Historical task completion of W j The better the historical completion of (2), r j The higher; w (W) j D is completed each time i,j E obtained after i,j Will also be used to update W j R of (2) j The update formula is as follows:
r j ' represents W j Updated reputation value, r j Represents W j Reputation value before updating E i,j Represents W j Executing task τ i Quality assessment score of (2);
step (5.3), R ID According to m i For D i,j Calculation of wherein /> Represents m i For R ID Prepared reference data set SD i The gradient produced, Z i,j ＝/>Represents m i For W j Submitted D i,j The gradient generated; the index compares W using gradients generated during federal learning training j Submitted D i,j And R is R ID Similarity between the required data;
R ID calling Γ i Based on the index (t) i,j ,r j ,l i,j ) Calculation D i,j Quality score of (2) wherein />
Step (5.4), task publisher R ID According to E i,j Generating τ from high-to-low ordering i Quality assessment ranking table QART of (E) i Uploading to a blockchain; at τ i After finishing, R ID The random number { r } generated when the Pedersen commitment was previously made 1 ,r 2 ,…,r N ,r N+1 ,…,r 2N Upload to blockchain for W j And verifying the validity of the quality evaluation process and the correctness of the quality evaluation result.
Further, the step (6) verifies l i,j The specific method of calculating the correctness is as follows,
task performer W j First, the task data set D i,j Input to m i The gradient vector obtained in (3) Then calculate +.>Let d=l i,j Task executor W j Generating random number r' 1 ,r′ 2 ,…,r′ N According to r' 1 ,r′ 2 ,…,r′ N Calculate->W j Then calculate promise->C＝G D *H R The method comprises the steps of carrying out a first treatment on the surface of the Last W j Verification equationWhether or not it is true, if so, prove l i,j The calculation flow of the index is legal and the result is correct; w (W) j Verification t can be directly calculated through information on a blockchain i,j and rj If all three indexes are verified to pass, then prove R ID For W i,j The quality evaluation flow is legal and the result is correct; t is t i,j and rj The verification process of correctness of (2) is relatively simple, since t is calculated i,j and rj The required related data are all on the blockchain, so that the verifier can directly check;
W j and RID An arbitration application may be initiated for different reasons, such as R ID May doubt W j Occurrence of plagiarism or suspicion of receipt of a task data set D i,j Incomplete or tampered with, W j It may be suspected that the own ranking result is too low or that the own submitted data is tampered with in the transmission. Once W is j or RID The method comprises the steps that when a malicious behavior possibly exists in another party, an arbitration application can be initiated to a crowdsourcing platform; the specific method for arbitration comprises the following steps:
if the reason for initiating arbitration is preset within the system,the crowdsourcing platform calls the previously written intelligent contract to process, and issues the result and key steps in the process to the blockchain to be disclosed after the processing is completed; suppose that the reason for the application arbitration is W j Verification E i,j Incorrect, intelligent contract check W j and RID Evidence information previously uploaded to the blockchain if verified that E is not found to exist i,j If the calculation is incorrect, publishing the arbitration processing result and key information in the process to a blockchain for disclosure; if it is found after verification that E is present i,j Incorrect calculation, let W according to the error cause j Resubmitting data or letting R ID re-W j Calculation E i,j The crowdsourcing platform issues the arbitration processing result and key information in the process to a blockchain for disclosure;
If the reason for initiating arbitration is not preset, matching corresponding professionals to be responsible for manually processing arbitration application by the crowdsourcing platform, and publishing the processing result and key processing steps in the process to a blockchain for disclosure after processing is completed;
if more than half of the participants agree to the arbitration result, the arbitration passes, and the crowd Bao Ping calls the intelligent contract to update the quality evaluation ranking table QART according to the arbitration result i . If the participants who agree to the arbitration result are less than half, the original ranking result is maintained.
Further, the specific process of paying and updating the model in the step (7) is as follows:
step (7.1), the crowdsourcing platform calls the intelligent contract to evaluate the ranking table QART according to the latest quality i and RID Filled-in remuneration distribution algorithm RAlloc i Calculate and distribute each W j The obtained remuneration b i,j ；
Step (7.2), the crowdsourcing platform calls intelligent contracts to place the latest quality assessment ranking list QART i E on i,j Inputting the weight distribution into a normalized exponential function softmax to calculate and obtain the weight distribution of each client learning of the federal learning
Step (7.3), the crowdsourcing platform calls intelligent contracts to aggregate the gradients to obtain an aggregate result
Step (7.4), the crowdsourcing platform calls intelligent contract to update model parameters
The beneficial effects are that: compared with the prior art, the invention has the following advantages:
(1) The invention provides a new quality assessment algorithm, which greatly reduces the cost overhead of a quality assessment stage in a crowdsourcing platform, so that the quality assessment is not dependent on experts or other human subjective factors any more, the operability and efficiency of the quality assessment stage are improved, and the quality assessment can be automatically carried out without additional data input.
(2) According to the invention, a blockchain technology is introduced into the newly proposed quality evaluation algorithm to record important information in the evaluation process, so that the quality evaluation score of the crowdsourcing platform can be verified, and the process can be traced. The invention allows a task publisher or a task executor to issue a question on the quality assessment score, ensures the transparency of the disclosure of the arbitration process through the blockchain, and calls the arbitration program to trace back relevant archiving points in the quality assessment process on the blockchain to detect the correctness and legality of the quality assessment process. Uploading the arbitration result to the blockchain to be shown, and if more than half of the participants accept the shown arbitration result, finishing arbitration. The invention uses the characteristics of decentralization and automation of the blockchain technology to pay the task executor through intelligent contracts.
(3) According to the invention, a federal learning technology is introduced into a newly proposed quality evaluation algorithm to realize iterative updating of an evaluation model, and a model training stage of federal learning occurs at a task executor end, so that leakage of original data is avoided. According to the invention, the gradient returned by the task executor is subjected to noise adding processing by using the differential privacy technology, so that the difficulty of the adversary in deducing useful information from the gradient is greatly increased, and the data privacy of the task executor is protected.
(4) The quality evaluation algorithm designed by the invention considers various indexes and tries to ensure the fairness and objectivity of the evaluation result. Wherein l ij The indexes can describe the similarity between the data submitted by the task executor and the data required by the task publisher, and the subjective evaluation by people is not needed, so that the operability is high and the efficiency is high; r is (r) j The index characterizes the completion condition of the historical task of the task executor, and the updated formula of the index prevents a certain task executor from frequently submitting low-quality task data and prevents the task executor from severely reducing the reputation value because the task executor submits the low-quality task data once; t is t ij The index characterizes the efficiency of task performers to complete tasks, which has two roles:
One is to ensure the effectiveness of the quality assessment algorithm, when the other two indexes are the same, the quality assessment algorithm may be difficult to distinguish task quality submitted by different task performers, and the probability of collision can be greatly reduced by introducing the index;
secondly, the contribution rate of unit time is used for evaluating the task completion condition, so that a task executor is stimulated to complete the task quickly and well. Different from a single index, multiple indexes can comprehensively consider the completion condition of a task executor from multiple dimensions, and the evaluation result is more comprehensive and more convincing.
Drawings
FIG. 1 is a block diagram of the overall system of the present invention;
FIG. 2 is a block diagram of federal learning architecture in an embodiment;
FIG. 3 is a schematic representation of differential privacy mechanism in an embodiment.
Detailed Description
The technical scheme of the present invention is described in detail below, but the scope of the present invention is not limited to the embodiments.
As shown in FIG. 1, the crowd-sourced quality security assessment method based on federal learning with the assistance of a blockchain of the invention relates to a crowd-sourced platform, a key generation center, a plurality of task issuers, a plurality of task executors and a public blockchain. The key generation center is responsible for generating public and private key pairs for successfully registered task publishers and task executors; the crowdsourcing platform is responsible for managing the situation that a task publisher publishes a task, recording the situation that a task executor accepts the task, and processing arbitration requests from the task publisher or the task executor; the blockchain is responsible for assisting the crowdsourcing platform to initialize the system, recording evidence information in the running process of the system, disclosing arbitration results, paying remuneration to task executors and updating a quality assessment algorithm.
The main steps of the invention include: firstly, after a crowdsourcing platform is initialized by a system, a task publisher and a task executor can register to be crowdsourcing platform users, a key generation center can generate public and private key pairs for the users who are registered successfully, and the crowdsourcing platform can generate a task matching table for the task publisher who is registered successfully; when a task publisher has task demands, the task publisher fills the task demands according to a specified format and then sends the task demands to a crowdsourcing platform, and meanwhile, the evidence information of the published task is uploaded to a blockchain; when a task executor is interested in a certain task and hopes to participate, the task can be accepted, the crowdsourcing platform registers the acceptance condition of the task and sends a quality evaluation algorithm to the task executor, and after the acceptance is successful, the task executor uploads evidence information of the accepted task to a blockchain; task executors finish tasks according to requirements, submit task data to task publishers and upload evidence information of submitted task data to block chains; after receiving the data, the task publisher uploads evidence information of the received task data to a block chain; the task publisher uses a quality evaluation algorithm to perform quality evaluation on task data submitted by a task executor and uploads a quality evaluation ranking result to the blockchain; the task publisher and the task executor can challenge the quality evaluation ranking result, the crowdsourcing platform can start arbitration, check according to the evidence information recorded by the blockchain, send the arbitration result to the blockchain for disclosure, and if more than half of the participants agree with the arbitration result, the arbitration is passed; the crowd Bao Ping calls intelligent contracts to update quality ranking results according to arbitration results and pay rewards to task executors according to the latest quality ranking results; and the crowdsourcing platform calls the intelligent contract to aggregate and update the quality assessment model by adopting the federal learning technology according to the latest quality assessment ranking result.
Examples
As shown in fig. 2, the specific implementation steps of this embodiment are as follows:
step (1), system initialization
Step (11), the user 1 is registered as a task publisher R on the crowdsourcing platform ID The number ID generated by the crowdsourcing platform is 3478945789287, and the user 2 is registered as a task executor W on the crowdsourcing platform j The number j generated by the crowdsourcing platform is 9253147457868, and the user 3 is registered as a task executor W on the crowdsourcing platform j The number j generated by the crowdsourcing platform is 3573871724169, and the user 4 is registered as a task executor W on the crowdsourcing platform j The number j generated by the crowdsourcing platform is 4146234890752, and the user 5 is registered as a task executor W on the crowdsourcing platform j The crowdsourcing platform generates a number j of 5224724062752 for it.
Step (12), the crowdsourcing platform maintains q federal learning models M= { M through intelligent adaptation to different task scenes 1 ,m 2 ,…,m q Defining initial threshold C for aggregation data for different models l Where l=1, 2, …, q.
And (13) the crowdsourcing platform performs initialization parameter setting for the Pedersen promise, and randomly generates base points G and H on an elliptic curve.
Step (14), the crowdsourcing platform registers the task publisher R successfully ID Generating a task matching table TMT ID For recording R ID And issuing information of the task, wherein the information is initially empty.
Step (15), the key generation center is the task publisher R with successful registration ID And task performer W j Generating public-private key pairs (PK ID ,SK ID) and (PKj ,SK j )。
Step (2), a task publisher publishes tasks on the crowdsourcing platform
Step (21), 2023-4-4, 15:20:14, task publisher R with ID of 3478945789287 ID Task tau intended to issue a picture marker i Where i= 9347835925163 is the number that the crowdsourcing platform generates for the task. He fills in the task information according to the format, the selected federal learning model m i Pretrained ResNet for CIFAR-10 using the classical dataset; sum of rewards provided to task performers B i :1000 yuan; task demand description TR i Including the task start time Task end time->Specific requirements TD of task i : requiring 3 task performers W j Respectively 3 different picture sets, each data set being 10000 pictures (k i ＝10000)，W j Task data to be submitted-> A vector in 3073 dimensions, the first 3072 dimensions representing RGB channel representations of a picture (the first 1024 dimensions representing Red channel data, the middle 1024 dimensions representing Green channel data, the last 1024 dimensions representing Blue data), the last 1 dimensions representing classification labels of the picture (10 categories in total, 0 representing aircraft, 1 representing automobiles, 2 representing birds, 3 representing cats, 4 representing deer, 5 representing dogs, 6 representing frogs, 7 representing horses, 8 representing boats, 9 representing trucks); task publisher R ID Tau to be issued for it i Preparing a reference data set SD i
(indicating τ i Is included) to assist the quality assessment algorithm in the assessment, SD i Is a lattice of (2)And W is as follows j Task data set D to be submitted i,j The same applies. To prevent SD i Leakage R ID R ID No need to publish SD i And also to prevent R ID Random replacement or modification of SD i Thus R is ID There is a need to provide and SD in the demand description of tasks i Related petersen commitments. Introduction of the Pedersen promise ensures both R ID In evaluation D i,j When using the same SD i At the same time utilize the homomorphism of Pedersen promise, W j And can verify R Id Invoking a quality assessment algorithm Γ i Accuracy of (3). Pedersen commitment { c 1 ,c 2 ,…,c N ,c N+1 ,…c 2N The method of generation of the sequence is as follows, wherein N is m i Dimension of return gradient: r is R ID Will reference data set SD i As m i Input of (a) to obtain a gradient
R ID For promise data Calculating corresponding promise value-> wherein r1 ,r 2 ,…,r N ,r N+1 ,…,r 2N Is R ID A generated random number; ralloc algorithm for remuneration distribution i : according to W j Quality assessment score E of (2) i,j Proportionally distribute remuneration, W j Completion of task τ i The obtained remuneration calculation formulaIs->In addition, R ID Task tau issued to self by SHA256 hash algorithm i Calculating hash value ζ i ＝H(τ i ) And uses its own private key SK ID Para zeta i Signing to obtain psi i ＝Sig(ζ i ,SK ID ) And uploaded to the blockchain for subsequent traceback.
Step (22), the crowdsourcing platform is used for carrying out crowdsourcing according to R ID Filled-in task information and R ID To perform validity check on the related information of (1), such as checking whether all necessary information is complete, R ID Whether the balance is not less than B i ，Whether or not to be later than->RAlloc i Whether the sum exceeds B i 。
Step (23), task τ i Through validity verification, the crowdsourcing platform issues task tau i ＝(ID,i,m i ,B i ,TR i ,RAlloc i ,Γ i ,κ i ) And at the task publisher R ID Task matching table TMT of (a) ID An entry { i, W } = (9347835925163, { }) is added for recording task τ i Is a match of the matching information of (a).
Step (24), the crowdsourcing platform is used for carrying out crowdsourcing according to R ID Filled RAlloc i An intelligent contract is generated and deployed for subsequent invocation.
Step (3), a task executor browses task information and selects the task of interest to accept
Step (31), for example task executor W numbered j 9253147457868 j For task τ with task number i 9347835925163 i Interested, the task acceptance is selected at a time of 2023-4-12 11:43:58;
step (32), checking the task of which the task is registered by the crowdsourcing platformIf the executor has reached the upper limit, an error message is returned if the upper limit is exceeded, if the registered task executor has not reached the upper line, the crowdsourcing platform checks whether the task is within an acceptable time range (assuming that the task receiving time in the scene is not later than the task starting time) ) If the task is within the acceptable time range, the task is matched with the TMT corresponding to the task ID And adding a record containing the task executor, and returning error information to the task executor if the record is not in an acceptable time range. w (w) j When the task is accepted, the registered task executor is not online, and the acceptance time is 2023-4-12:11:43:58, and is within the acceptance time of the task, so the crowdsourcing platform is in R ID Task matching table TMT of (a) ID Registering the W j (i, W) = (9347835925163, {9253147457868 }) and invoke the intelligent contract to model m the federal learning model i (in this example, resNet model) to W j ；
Step (33), W j Successful acceptance of tau i Then, calculate the acceptance τ i Hash value ζ 'of (1)' i ＝H(τ i ) Using its own private key SK j To ζ' i Signing to obtain a signature value psi' i ＝Sig(ζ′ i ,SK j ) And upload to the blockchain;
step (34), e.g. task performer W numbered j 3573871724169 j Interested in task number i being 9347835925163, selecting the task acceptance at time 2023-04-15:21:15:43;
step (35), checked W j When receiving the task, the registered task executor is not on line and is in the receiving time of the task, so the crowdsourcing platform is at the TMT with the corresponding number ID= 9347835925163 ID Registering the W j Information (i, W) =
(9347835925163, {9253147457868,3573871724169 }) and invoking the intelligent contract to model m the federal learning model i (in this example, resNet model) to W j 。W j Successful acceptance of tau i Then, calculate the acceptance τ i Hash value ζ 'of (1)' i ＝H(τ i ) Using its own private key SK j To ζ' i Signing to obtain a signature value psi' i ＝Sig(ζ′ i ,SK j ) And upload to the blockchain;
step (36), e.g. W numbered j 4146234890752 j Of interest to task number i as 9347835925163, the task acceptance is selected at time 2023-4-22 15:41:32. Checked W j When receiving the task, the registered task executor is not on line and is in the receiving time of the task, so the crowdsourcing platform is at the TMT corresponding to the number 9347835925163 ID Register the information of the task performer (i, W) = (9347835925163, {9253147457868,573871724169,9347835925163 }) and call the intelligent contract to learn model m federally i (in this example, resNet model) to W j 。W j Successful acceptance of tau i Then, calculate the acceptance τ i Hash value ζ 'of (1)' i ＝H(τ i ) Using its own private key SK j To ζ' i Signing to obtain a signature value psi' i ＝Sig(ζ′ i ,SK j ) And upload to the blockchain;
step (37), e.g. W numbered j 5224724062752 j Of interest to task number 9347835925163, the task acceptance is selected at time 2023-4-24 10:55:54. The crowdsourcing platform firstly checks whether the task executor of the task registration reaches the upper limit, and the checked registered task executor reaches the online state and can not accept the task any more, so the crowdsourcing platform returns error information to the task executor.
Step (4) the task executor completes the task according to the task description
Step (41), W numbered j 9253147457868 j Description of TD according to task i Labeling and classifying pictures to finish task tau i Obtaining the product/>
Step (42), W j Will D i,j Kappa of (C) i The hash value of each task data is calculated by splicing the task data Using its own private key SK j Para zeta i,j Signature is carried out to obtain->
Step (43), W j Based on D i,j Local training m i Obtaining local model parameter g i,j ；
Step (44) W j According to a preset threshold C k For g i,j Cutting to obtainK is the selected model m i Numbering of the corresponding threshold:
step (45), W j Differential privacy technology pairAdding noise to obtain->
Where L is the size of the data set,is Gaussian noise, sigma 2 Is the noise scale, I is the sensitivity;
step (46), W j Using public key PK ID Encryption D i,j Obtaining ciphertext xi i,j ＝Enc(D i,j ,PK ID ) Will be xi i,j Sent to R ID Will be and />Uploading to the blockchain together; r is R ID Record W j Task commit time->Checking if it is later than the end time of the task +.>If exceed->Refusing to accept the task submission due to W j Is->2023-5-24:18:06:39, earlier than +.>Therefore W is j The submission is successful;
step (47), W with number j of 3573871724169 j Description of TD according to task i Labeling and classifying pictures to finish task tau i Obtaining the productWill D i,j Kappa of (C) i The hash value of each task data is calculated by splicing the task dataUsing its own private key SK j Para zeta i,j Signing to obtain ζ i,j ＝Sig(ζ i,j ,SK j )。W j Based on D i,j Local training m i Obtaining local model parameter g i,j 。W j According to a preset threshold C K For g I,j Cutting to obtainDifferential privacy technology pair->Adding noise to obtain->W j Using public key PK ID Encryption D i,j Obtaining ciphertext xi i,j ＝Enc(D i,j ,PK ID ) Will be xi i,j Sent to R ID Will-> and ζi,j Together to the blockchain. R is R ID Record and check W j Task commit time->Due to W j Is->2023-5-25:16:39:42, earlier than +.>Therefore W is j The submission is successful;
step (48), W numbered j 4146234890752 j Description of TD according to task i Labeling and classifying pictures to finish task tau i Obtaining the productWill D i,j Kappa of (C) i The hash value of each task data is calculated by splicing the task dataUsing its own private key SK j Para zeta i,j Signing to obtain ζ i,j ＝Sig(ζ i,j ,SK j )。W j Based on D i,j Local training m i Obtaining local model parameter g i,j 。W j According to a preset threshold C K For g i,j Cutting to obtainDifferential privacy technology pair->Adding noise to obtain->W j Using public key PK ID Encryption D i,j Obtaining ciphertext xi i,j ＝Enc(D i,j ,PK ID ) Will be xi i,j Sent to R ID Will-> and ζi,j Together to the blockchain. R is R ID Record and check W j Task commit time->Due to W j Is->2023-5-27 12:31:18 earlier than +.>Therefore W is j The commit was successful.
Step (5) R ID Calling Γ i For W j Submitted task dataSet Di ,j Quality assessment is performed
Step (51), R ID Using its own private key SK ID Received ciphertext xi i,j Decrypting to obtain plaintextWill D' i,j Kappa of (C) i The hash value of the task data is calculated by splicing the task data>Using its own private key SK ID To ζ' i,j Signing to obtainUploaded to the blockchain for subsequent traceback. R is R ID Calling Γ i From each W recorded j Task commit time->Calculate each W j Task time of (1)>(the task scenario assumed here is W j At start time +.>The task can be started after that);
TABLE 1 task performer t i,j Index scoring condition
| Task executor ID | 9253147457868 | 3573871724169 | 4146234890752 |
| t i,j | 23.34 | 24.28 | 26.52 |
Step (52), R ID Obtaining the latest reputation value r of each task executor from a blockchain j ；
TABLE 2 task performer r j Index scoring condition
| Task executor ID | 9253147457868 | 3573871724169 | 4146234890752 |
| r j | 93.45 | 79.86 | 84.69 |
Step (53), R ID According to the selected model m i To D i,j Calculation of l i,j The index compares W using gradients generated during federal learning training j Submitted D i,j And R is R ID The provided reference data set SD i Similarity between them. R is R ID Reference data sets SD prepared by the respective method i and Wj Submitted D i,j As training model m i Due to R ID In TD i The format of the submitted data is explicitly specified, so the dimensions of the two are the same. m is m i Will return to our corresponding gradient during training, the former is recorded as The latter is marked as +.>We can then calculate +.>
TABLE 3 task performer l i,j Index scoring condition
| Task executor ID | 9253147457868 | 3573871724169 | 4146234890752 |
| l i,j | 35.17 | 26.82 | 25.06 |
Step (54), R ID Calculating D according to the three indexes i,j Quality assessment score of (2) wherein />
TABLE 4 task performer quality assessment total score E i,j Scoring condition
| Task executor ID | 9253147457868 | 3573871724169 | 4146234890752 |
| t i,j | 23.34 | 24.28 | 26.52 |
| r j | 93.45 | 79.86 | 84.69 |
| l i,j | 35.17 | 26.82 | 25.06 |
| E i,j | 273.45 | 273.43 | 275.89 |
Step (55), R ID According to E i,j Generating τ from high-to-low ordering i Quality assessment ranking table QART of (E) i Uploading to the blockchain. At τ i After finishing, R ID The random number { r } generated when the Pedersen commitment was previously made 1 ,r 2 ,…,r N ,r N+1 ,…,r 2N Upload to the blockchain for the task executor to verify the validity of the quality assessment process and the correctness of the quality assessment score.
Table 5 task performer quality assessment ranking results
| Task executor ID | 4146234890752 | 9253147457868 | 3573871724169 |
| t i,j | 26.52 | 23.34 | 24.28 |
| r j | 84.69 | 93.45 | 79.86 |
| l i,j | 25.06 | 35.17 | 26.82 |
| E i,j | 275.89 | 273.45 | 273.43 |
| Ranking | 1 | 2 | 3 |
Step (6) the task executor and the task publisher can initiate the question according to different reasons, the crowdsourcing platform can initiate an arbitration program to check the content of the question, trace back the program execution process, and finally upload the arbitration result to the blockchain to be disclosed
Step (61) W j Can be according to R ID The uploaded ranking result checks the scores of various indexes of the user, and can be according to R ID Offered Pedersen promise { c 1 ,c 2 ,…,c N ,c N+1 ,…,c 2N Sum of random numbers
{r 1 ,r 2 ,…,r N ,r N+1 ,…,r 2N Verification l i,j The correctness of (2) is verified by the following method, W j Will D i,j Input to m i The gradient vector obtained in (3)Calculate-> Let d=l i,j 。W j Generating random number r' 1 ,r′ 2 ,…,r′ N According to r' 1 ,r′ 2 ,…,r′ N Calculate-> W j Calculation promise-> C＝G D *H R 。W j Verification equationWhether or not it is true, if so, prove l i,j The calculation flow of the index is legal and the result is correct. W (W) j Verification t can be directly calculated through information on a blockchain i,j and rj If all three indexes are verified to pass, then prove R ID For W i,j The quality evaluation flow is legal and the result is correct;
step (62) W with number j of 9253147457868 j According to R ID M provided i And Pedersen promise to verify the rank results of the quality assessment algorithm, find and R ID E uploaded to blockchain i,j Have larger access to challenge self-submitted D i,j Tampered in the transmission process, and requires to initiate arbitration;
step (63) consists ofThe questioning content is a common reason, so the crowdsourcing platform can call the previously deployed intelligent contracts for arbitration. The smart contract may require a W that initiates a challenge j Providing its public key PK j Signature ψ 'previously uploaded to blockchain' i Andr requiring to be questioned ID Providing its public key PK ID Signature ψ previously uploaded onto blockchain i and ζ′i,j . Crowd-sourced platform invoking intelligent contract to use PK j Decrypting a previous signature value ψ' i To obtain ζ' i Use of PK ID Decryption psi i To obtain zeta i Verify ζ' i and ζi Whether equal. If equal, then specify τ i The transmission and reception process is error-free. Checked that both are equal so τ i The transmission and reception process is error-free. Smart contract use PK j Decrypting a previous signature value +.>To obtain zeta i,j Use of PK ID Decryption->To obtain ζ' i,j Verification zeta i,j and ζ′i,j Whether equal. If equal, then describe D i,j The transmission and reception process is error-free. Through inspection that both are equal, D i,j The transmission and receiving process is error-free;
step (64) the crowdsourcing platform uploads the arbitration result to the blockchain to be disclosed, the comments of more than half of the participants are collected, the arbitration result is agreed by more than half of the participants, and the arbitration is finished;
step (65) W with number j of 3573871724169 j According to R ID M provided i And Pedersen promise to verify the rank results of the quality assessment algorithm, find and R ID E uploaded to blockchain i,j Have larger access to challenge self-submitted D i,j Is tampered with during transmission, and requires arbitration to be initiated. Since the content of the question isFor common reasons, the crowdsourcing platform invokes a previously deployed smart contract to arbitrate. The smart contract may require a W that initiates a challenge j Providing its public key PK j Signature ψ 'previously uploaded to blockchain' i and ζi,j Require R to be questioned ID Providing its public key PK ID Signature ψ previously uploaded onto blockchain i Andsmart contract use PK j Decrypting a previous signature value ψ' i To obtain ζ' i Use of PK ID Decryption psi i To obtain zeta i Verify ζ' i and ζi Whether equal. If equal, then specify τ i The transmission and reception process is error-free. Checked that both are equal so τ i The transmission and reception process is error-free. Smart contract use PK j Decrypting a previous signature value +.>To obtain zeta i,j Use of PK ID Decryption->To obtain zeta i ′ ,j Verification zeta i,j and ζ′i,j Whether equal. If equal, then describe D i,j The transmission and reception process is error-free. Inspection shows that the two are inconsistent, say D i,j There is a tamper phenomenon during transmission and reception. So the crowdsourcing platform will require W j Resubmit D i,j To R ID And meet D i,j Hash value of (2) and W j The hash value uploaded onto the blockchain for the first time is the same. R is R ID Re-invoking Γi i Evaluating newly submitted data D i,j And will be up to date E i,j Uploading to the blockchain. And the final crowdsourcing platform uploads the arbitration result to the blockchain to disclose and collect the comments of the participants, and more than half of the participants agree to the arbitration result, so that the arbitration result passes and the arbitration flow is ended. Updating final quality assessment ranking results based on arbitration results . In this case, the updated quality assessment ranking results are as follows:
table 6 updated task executor quality assessment ranking results
| Task executor ID | 4146234890752 | 3573871724169 | 9253147457868 |
| t i,j | 26.52275.89 | 24.28274.16 | 23.34 |
| r j | 84.69 | 79.86 | 93.45 |
| l i,j | 25.06 | 26.54 | 35.17 |
| E ij | 275.89 | 274.16 | 273.45 |
| Ranking | 1 | 2 | 3 |
Step (7), the crowdsourcing platform calls the intelligent contract to evaluate the ranking list QART according to the latest quality i To execute the reward distribution program, and evaluate the ranking table QART according to the latest quality i E on i,j Federal learning model in weighted update quality assessment algorithm
Step (71), the crowdsourcing platform calls the intelligent contract to evaluate the ranking table QART according to the latest quality i and RID Filled-in remuneration distribution algorithm RAlloc i Calculate and distribute each W j The obtained remuneration
Step (72), the crowdsourcing platform calls intelligent contracts to make the latest ranking list QART i E on i,j Inputting the weight distribution into normalized exponential function to calculate and obtain weight distribution of each client learning of federal learning
TABLE 7 weight for computing the participation model update for task performers based on their quality assessment scores
| Task executor ID | 4146234890752 | 3573871724169 | 9253147457868 |
| E ij | 275.89 | 274.16 | 273.45 |
| φ j | 0.79 | 0.14 | 0.07 |
Step (73), the crowdsourcing platform calls intelligent contracts to aggregate the gradients to obtain an aggregation result
Step (74), the crowdsourcing platform calls intelligent contract to update model parameters/>
Claims (8)
1. A crowdsourcing quality safety assessment method based on federal learning under the assistance of a blockchain is characterized by comprising a key generation center, a task publisher, a task executor, a crowdsourcing platform and the blockchain, and comprises the following steps:
Step (1), system initialization including user registration, parameter generation and key distribution
User registers as task publisher R on crowdsourcing platform ID Or task performer W j ID and j are numbers generated by the crowdsourcing platform for the user; generating initial parameter C by crowdsourcing platform l G, H and TMT ID The method comprises the steps of carrying out a first treatment on the surface of the The crowdsourcing platform maintains q federal learning models M= { M suitable for different task scenes 1 ，m 2 ，...，m q }，C l Initial thresholds used when aggregating data for different models, l=1, 2, q; g and H are base points randomly generated on an elliptic curve and are initial parameters promised by Pedersen; TMT (transition metal-transition metal) ID Task publisher R with successful registration by crowdsourcing platform ID A generated task matching table for recording R ID The published task information is initially empty; the key generation center is the task publisher R with successful registration ID And task performer W j Generating public-private key pairs (PK ID ，SK ID) and (PKj ，SK j )；
Step (2), task publisher R ID Publishing task τ on crowd-sourced platform i ＝(ID，i，m i ，B i ，TR i ，RAlloc i ，Γ i ，κ i ) I is the number of the issuing task, m i Is R ID For task tau i Selected federal learning model, B i For task tau i Remuneration of (TR) i Is tau i Task demand description of R ID There is a need to provide task related information and Pedersen commitments, RAlloc therein i Is tau i And (f) a reward distribution algorithm i Is tau i Task data quality assessment algorithm, κ i Representing task τ i Requiring each W j The number of data submitted;
task publisher R ID Computing task τ i Hash value ζ of (a) i ＝H(τ i ) Using its own private key SK ID For hash value ζ i Signing to obtain psi i ＝Sig(ζ i ，SK ID ) And upload to the blockchain; the crowdsourcing platform is according to R ID Filled-in task information and R ID Legitimacy check is carried out on the related information of (2), and if the related information passes the legitimacy check, a task tau is issued i And at R ID Task matching table TMT of (a) ID Adding an entry { i, W } for recording τ i W represents successful acceptance τ i Is initially empty;
step (3), task executor W j Selecting a task tau of interest i Receiving, wherein the crowdsourcing platform is responsible for recording task matching conditions;
when W is j Accept τ i If τ i Within acceptable time, the crowdsourcing platform will accept the W for the task j Added to task matching table TMT ID In W corresponding to the task list item { i, W }, and calling the intelligent contract to learn model m of federation i Send to W j The method comprises the steps of carrying out a first treatment on the surface of the If task tau i Not within acceptable time, W j Accept τ i Failure;
W j accept τ i After success, calculate the acceptance task τ i Hash value ζ 'of (1)' i ＝H(τ i ) Using the private key SK itself j To ζ' i Signing to obtain a signature value psi' i ＝Sig(ζ′ i ，SK j ) And upload to the blockchain; psi' i The signature value obtained by signing the task hash value by a task executor by using a private key is evidence information of a crowdsourcing task matching stage, so that the task hash value is convenient to check in an arbitration stage;
step (4), W j Completion of task τ according to task demand description i Obtaining a task data set Task performer W j Will D i，j Kappa of (C) i The hash value of the task data is calculated by splicing the task data>Using private key SK j Para zeta i，j Signature is carried out to obtain->Task data->Depending on the task, the task publisher can add +.>Is explicitly specified;
W j based on D i，j Local training federal learning model m i Obtaining the model parameter g to be polymerized i，j For g i，j Cutting to obtainDifferential privacy technology pair is used again>Adding noise to obtain->
W j Task data set D i，j With corresponding task issuers R ID Public key PK of (2) ID Encryption to obtain xi i，j ＝Enc(D i，j ，PK ID ) And send to R ID Will beAnd->Uploading to the blockchain together;
step (5), task publisher R ID To xi i，j Decryption to obtain Will D' i，j Kappa of (C) i The hash value of the task data is calculated by splicing the task data> Using the private key SK itself ID To ζ' i，j Signing to obtain ζ' i，j ＝Sig(ζ′ i,j ，SK ID ) And zeta 'is taken as' i，j Uploading to a blockchain;
R ID invoking a quality assessment algorithm Γ i For D' i，j Quality evaluation is carried out to obtain an evaluation score E i，j The method comprises the steps of carrying out a first treatment on the surface of the The evaluation index includes three items (t i，j ，r j ，l i，j), wherein ti，j For task execution time, r j Is W j Reputation value of l i，j Is based on federal learning model m i Calculating the obtained data similarity; r is R ID According to E i，j Ranking from high to low and grouping tasks τ i Quality assessment ranking table QART of (E) i Uploading to a blockchain; at task τ i After completion of R ID The random number { r > of the Pedersen commitment will be verified 1 ，r 2 ，...，r N ，r N+1 ，...，r 2N Upload to blockchain for W j Validating a data evaluation procedure, wherein N is m i Returning to the dimension of the gradient;
step (6), W j According to R ID The model m provided i Pedersen promise { c 1 ，c 2 ，...，c N ，c N+1 ，...c 2N Sum of random numbers { r } and 1 ，r 2 ，...，r N ，r N+1 ，...，r 2N verifying Γ i For evaluation results of (e) such as R ID and Wj In this stage, if the ranking result is questioned, an arbitration program is initiated to the crowdsourcing platform, the crowdsourcing platform examines the questioning content, and the arbitration result is uploaded to the blockchain for publicity within a specified time, if more than half of the participants agreeArbitration results, the crowdsourcing platform will call the intelligent contract to update the quality assessment ranking table QART i ；
Step (7), the crowdsourcing platform calls an intelligent contract to evaluate a ranking table QARV according to the latest quality i To execute the reward distribution program according to QARV i Upper W j Ranking condition weighted update m i 。
2. The method for crowd-sourced quality safety assessment based on federal learning with blockchain assistance according to claim 1, wherein the method comprises the steps of: task demand description TR in step (2) i The internal details of (a) are:
wherein ,representative τ i Start time of->Representative τ i End Time of (TD) i Representative τ i Specific requirements of { c } 1 ，c 2 ，...，c N ，c N+1 ，...，c 2N The expression of Pedersen commitment, R ID Tau to be issued for it i Preparing a reference data set SD i To assist the quality assessment algorithm in the assessment, SD i Format and D of (2) i，j The same;
R ID will reference data set SD i Inputting a model m as a training dataset i Obtaining a gradient T i，j And to V i，j Providing relevant Pedersen commitments to facilitate task performers W j Subsequent verification of quality assessment score E i，j。
3. The method for crowd-sourced quality safety assessment based on federal learning with blockchain assistance according to claim 1 or 2, wherein the method comprises the following steps: the Pedersen promise { c ] in the step (2) 1 ，c 2 ，...，c N ，c N+1 ，...，c 2N The generation process of the } is as follows:
task publisher R ID Will reference data set SD i As model m i Is input into the training dataset to obtain the gradient
R ID For promise data Calculating corresponding promise value->
wherein r1 ，r 2 ，...，r N ，r N+1 ，...，r 2N Is R ID And (5) generating a random number.
4. The method for evaluating the crowd-sourced quality safety based on federal learning with the assistance of a blockchain according to claim 1, wherein the method is characterized by: the details of the remuneration distribution algorithm in the step (2) are as follows: r is R ID First determine the rule according to which is to be used for W j Performing remuneration distribution and giving remuneration b i，j The specific calculation formula is as follows:
wherein bi，j Represents W j Completion of task τ i The obtained remuneration; according to W j Quality assessment score E of (2) i，j Proportionally distributing remuneration;
R ID the previously filled reward distribution algorithm or the new reward distribution algorithm can be selected, and the crowd-sourced platform generates an intelligent contract for the new reward distribution algorithm and deploys the intelligent contract on the blockchain after checking the legal.
5. The method for crowd-sourced quality safety assessment based on federal learning with blockchain assistance according to claim 1, wherein the method comprises the steps of: the step (4) is to obtain model parameters g i，j The formula for clipping is as follows:
k is the selected model m i Numbering of the corresponding threshold:
task performer W j Differential privacy technology pairAdding noise to obtain->The formula is as follows:
l is the size of the data set and,is Gaussian noise, sigma 2 Is the noise scale, I is the sensitivity;
task performer W j Will be and />After uploading to the blockchain, task publisher R ID Record W j Submitting task τ i Time of (2)Judging->Whether or not it is true, if so, W j The submission succeeds, and otherwise fails.
6. The method for crowd-sourced quality safety assessment based on federal learning with blockchain assistance according to claim 1, wherein the method comprises the steps of: the specific process of the step (5) is as follows:
Step (5.1), R ID Zeta's of' i，j ＝Sig(ζ′ i，j ，SK ID ) After uploading to the blockchain, R ID From each W recorded j Task commit time of (2)Calculate each W j Task time of (1)>
Step (5.2), R ID Obtaining the latest reputation value r of each task executor from a blockchain j The method comprises the steps of carrying out a first treatment on the surface of the The crowdsourcing platform uses blockchain for each W j Maintaining a reputation value r j The value depends on W j Historical task completion of W j The better the historical completion of (2), r j The higher; w (W) j D is completed each time i，j E obtained after i，j Will also be used to update W j R of (2) j The update formula is as follows:
r′ j represents W j Updated reputation value, r j Represents W j Reputation value before updating E i，j Represents W j Executing task τ i Quality assessment of (a)A score;
step (5.3), R ID According to m i For D i，j Calculation of wherein /> Represents m i For R ID Prepared reference data set SD i Gradient produced-> Represents m i For W j Submitted D i，j The gradient generated;
R ID based on the index (t) i，j ，r j ，l i，j ) Calculation D i，j Quality score of (2) wherein
Step (5.4), task publisher R ID According to E i，j Generating τ from high-to-low ordering i Quality assessment ranking table QART of (E) i Uploading to a blockchain; at τ i After finishing, R ID The random number { r } generated when the Pedersen commitment was previously made 1 ，r 2 ，...，r N ，r N+1 ，...，r 2N Upload to blockchain for W j And verifying the validity of the quality evaluation process and the correctness of the quality evaluation result.
7. The method for crowd-sourced quality safety assessment based on federal learning with blockchain assistance according to claim 1, wherein the method comprises the steps of: the step (6) verifies i，j The specific method of calculating the correctness is as follows,
task performer W j First, the task data set D i，j Input to m i The gradient vector obtained in (3) Then calculate +.>Let d=l i，j Task executor W j Generating random number r' 1 ，r′ 2 ，...，r′ N According to r' 1 ，r′ 2 ，...，r′ N Calculate->W j Then calculate commitmentLast W j Verification equationWhether or not it is true, if so, prove l i，j The calculation flow of the index is legal and the result is correct; w (W) j Direct computation of validation t through information on a blockchain i，j and rj If all three indexes are verified to pass, then prove R ID For W i，j The quality evaluation flow is legal and the result is correct; the specific method for arbitrating in the step (6) comprises the following steps:
if the reason for initiating arbitration is preset in the system, the crowdsourcing platform calls the previously written intelligent contract to process, and issues the result and key steps in the process to the region after the processing is completedThe block chain is shown; suppose that the reason for the application arbitration is W j Verification E i，j Incorrect, intelligent contract check W j and RID Evidence information previously uploaded to the blockchain if verified that E is not found to exist i，j If the calculation is incorrect, publishing the arbitration processing result and key information in the process to a blockchain for disclosure; if it is found after verification that E is present i，j Incorrect calculation, let W according to the error cause j Resubmitting data or letting R ID re-W j Calculation E i，j The crowdsourcing platform issues the arbitration processing result and key information in the process to a blockchain for disclosure;
if the reason for initiating arbitration is not preset, matching corresponding professionals to be responsible for manually processing arbitration application by the crowdsourcing platform, and publishing the processing result and key processing steps in the process to a blockchain for disclosure after processing is completed;
if more than half of the participants agree to the arbitration result, the arbitration passes, and the crowd Bao Ping calls the intelligent contract to update the quality evaluation ranking table QART according to the arbitration result i The method comprises the steps of carrying out a first treatment on the surface of the If the participants who agree to the arbitration result are less than half, the original ranking result is maintained.
8. The method for crowd-sourced quality safety assessment based on federal learning with blockchain assistance according to claim 1, wherein the method comprises the steps of: the specific process of paying and updating the model in the step (7) is as follows:
step (7.1), the crowdsourcing platform calls the intelligent contract to evaluate the ranking table QART according to the latest quality i and RID Filled-in remuneration distribution algorithm RAlloc i Calculate and distribute each W j The obtained remuneration b i，j ；
Step (7.2), the crowdsourcing platform calls intelligent contracts to place the latest quality assessment ranking list QART i E on i，j Inputting the weight distribution into a normalized exponential function softmax to calculate and obtain the weight distribution of each client learning of the federal learning
Step (7.3), the crowdsourcing platform calls intelligent contracts to aggregate the gradients to obtain an aggregate result
Step (7.4), the crowdsourcing platform calls intelligent contract to update model parameters
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310807841.2A CN116777294A (en) | 2023-07-04 | 2023-07-04 | Crowd-sourced quality safety assessment method based on federal learning under assistance of blockchain |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310807841.2A CN116777294A (en) | 2023-07-04 | 2023-07-04 | Crowd-sourced quality safety assessment method based on federal learning under assistance of blockchain |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116777294A true CN116777294A (en) | 2023-09-19 |
Family
ID=87985828
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310807841.2A Pending CN116777294A (en) | 2023-07-04 | 2023-07-04 | Crowd-sourced quality safety assessment method based on federal learning under assistance of blockchain |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116777294A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117291273A (en) * | 2023-11-24 | 2023-12-26 | 合肥微观纪元数字科技有限公司 | Quantum Computing Blockchain System |
| CN117473559A (en) * | 2023-12-27 | 2024-01-30 | 烟台大学 | Two-party privacy protection method and system based on federal learning and edge calculation |
-
2023
- 2023-07-04 CN CN202310807841.2A patent/CN116777294A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117291273A (en) * | 2023-11-24 | 2023-12-26 | 合肥微观纪元数字科技有限公司 | Quantum Computing Blockchain System |
| CN117291273B (en) * | 2023-11-24 | 2024-02-13 | 合肥微观纪元数字科技有限公司 | Quantum Computing Blockchain System |
| CN117473559A (en) * | 2023-12-27 | 2024-01-30 | 烟台大学 | Two-party privacy protection method and system based on federal learning and edge calculation |
| CN117473559B (en) * | 2023-12-27 | 2024-05-03 | 烟台大学 | Two-party privacy protection method and system based on federal learning and edge calculation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN116777294A (en) | Crowd-sourced quality safety assessment method based on federal learning under assistance of blockchain | |
| US11334882B1 (en) | Data access management on a distributed ledger system | |
| US11588804B2 (en) | Providing verified claims of user identity | |
| US11532195B2 (en) | System and method for cryptographic choice mechanisms | |
| US20200265514A1 (en) | Recording medium recording communication program and communication apparatus | |
| Lu et al. | On enabling machine learning tasks atop public blockchains: A crowdsourcing approach | |
| CN114363043B (en) | Asynchronous federal learning method based on verifiable aggregation and differential privacy in peer-to-peer network | |
| CN111899023A (en) | Crowd sourcing safety method and system for crowd sensing machine learning based on block chain | |
| CN113626866B (en) | Federal learning-oriented localization differential privacy protection method, system, computer equipment and storage medium | |
| CN113553377B (en) | Data sharing method and device based on block chain and federal learning | |
| CN112307331A (en) | Block chain-based college graduate intelligent recruitment information pushing method and system and terminal equipment | |
| CN115952532A (en) | Privacy protection method based on federation chain federal learning | |
| CN110365671B (en) | Crowd sensing incentive mechanism method supporting privacy protection | |
| Ragot et al. | IP lifecycle management using blockchain and machine learning: Application to 3D printing datafiles | |
| CN114091953A (en) | Credibility evaluation method and system based on heterogeneous block chain | |
| US11627125B1 (en) | Apparatus and methods for credentialing users across user devices | |
| US20240161064A1 (en) | System and method for improving fairness among job candidates | |
| CN111861737B (en) | Block chain-based wind control model optimization method and device and computer equipment | |
| CN112287040B (en) | Rights and interests combination method, device, equipment and medium based on block chain | |
| CN112308720B (en) | Method and device for determining participants in federal learning alliance | |
| CN117251882A (en) | Distributed federal learning method with data privacy protection and high model quality | |
| CN112085051A (en) | Image classification method and system based on weighted voting and electronic equipment | |
| US11907874B2 (en) | Apparatus and method for generation an action validation protocol | |
| CN114422147B (en) | Multi-party safety calculation method based on block chain | |
| CN112702159B (en) | Online expert scoring method and system based on block chain |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |