Data transaction market incentive method fusing reputation and auction strategies
By calculating the quality and reputation scores of the seller in the data trading market, designing auction mechanisms and poisoning verification, the problem of clean label poisoning attacks is solved, and market security and interests are improved.
Patent Information
- Application Number
- CN202510533254.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-26
- Publication Date
- 2025-08-01
AI Technical Summary
In the data trading market, clean label poisoning attacks are difficult to detect, affecting the market ecology and transaction security, and this problem is urgently needed.
By calculating the quality score and reputation scores of the enrolled seller dataset, an auction mechanism is designed to match the most suitable seller and perform poison verification, reward or punish reputation scores to incentivize the seller to provide high-quality datasets.
Effectively incentivize sellers to provide high-quality data sets, improve market security and the interests of buyers and sellers, and prevent poisoning attacks.
Smart Images
Figure CN120410697A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of network security technology, and in particular, to an incentive method for a data trading market that integrates reputation and auction strategies. Background Art
[0002] In a poisoning attack, a clean label means that the attacker does not modify the data label and only achieves the attack by fine-tuning the input sample. The key to a clean-label poisoning attack is that they only misclassify specific samples without affecting the overall performance of the model on other samples, making this kind of attack difficult to detect. These attacks are similar to the target image in the feature space, making it difficult for the network to distinguish between normal samples and poisoned samples.
[0003] In the data trading market, a clean-label poisoning attack has a profound impact on the market ecosystem, trading security, and technical trust system by covertly modifying the sample features (rather than directly tampering with the label). Therefore, there is an urgent need to provide a solution to solve the above problems. Summary of the Invention
[0004] The purpose of the present invention is to provide an incentive method for a data trading market that integrates reputation and auction strategies to solve the security problems caused by clean-label poisoning attacks in the data trading market.
[0005] An incentive method for a data trading market that integrates reputation and auction strategies provided by the present invention adopts the following technical solutions:
[0006] Based on the category center distance between the clean poisoning attack and the label value, and the distance of the target sample close to the category center of the poisoned sample in the feature space, obtain the sample feature representation of the candidate seller dataset, and calculate the quality score of the candidate seller dataset based on the sample feature representation;
[0007] Calculate and update the reputation score of the candidate seller dataset based on the quality score of the candidate seller dataset. The candidate seller participates in the bidding and proposes a bid price based on the reputation score;
[0008] Design an auction mechanism based on the reputation score and price. The buyer matches the seller with the highest score within the budget cost based on the auction mechanism and assigns the dataset task to the seller with the highest score;
[0009] Perform poisoning verification on the dataset content of the seller with the highest score. If the verification passes, the buyer and the seller conduct a transaction, calculate the reputation reward score of the bidder and give a reward. If the verification fails, no transaction is conducted, calculate the reputation penalty score of the bidder and deduct it.
[0010] Optionally, when calculating the quality score of the candidate seller dataset based on the sample feature representation, the following formula is used:
[0011]
[0012] f(Q)=a α ×Im β ;
[0013] wherein, is the quality score of the dataset, C represents the adjustment coefficient, n represents the sample size of the dataset, a is the label annotation accuracy, and its value range is (0, 1), Im is the class balance rate, and its value range is (0, 1), and α and β are adjustment parameters used to adjust the influence of accuracy and class balance rate on the final quality score.
[0014] Optionally, in the process of calculating and updating the reputation value score of the candidate seller dataset based on the quality score of the dataset, it includes:
[0015] The initial reputation value score of all newly added dataset sellers is defaulted to 1, and the reputation value score of the candidate seller dataset is calculated and updated based on the historical reputation score and the reputation value scores of other existing dataset sellers. The calculation formula is as follows:
[0016]
[0017] wherein, is the reputation value score of the i-th data seller at the y-th round, is the quality score of the dataset, σ is the historical weight parameter, is the historical reputation score.
[0018] Optionally, when the quality score of the dataset in the y-th round is at the top 50% level of the candidate seller dataset, the value range of σ is (0, 0.5), and when the quality score of the dataset in the t-th round is at the bottom 50% level of the candidate seller dataset, the value range of σ is [0.5, 1).
[0019] Optionally, in the process of the candidate seller participating in the bidding and submitting a bid price based on the reputation value score, it includes:
[0020] Randomly give an initial bid price, and construct the bidding strategy of the seller based on the reputation value score and the initial bid price of the candidate seller;
[0021] Based on the data collection, cleaning, and storage calculation costs, calculate the utility function of the seller based on the cost, bid price, and reputation value score changes;
[0022] Obtain the equilibrium strategies of other sellers, calculate the maximum expected revenue of the seller based on the equilibrium strategies and the utility function, and obtain the optimal bid price.
[0023] Optionally, in the process of designing the auction mechanism based on the reputation value score and price, it includes:
[0024] Determine the total budget for the buyer's current data set tender. The available budget for a single round is half of the total budget. The data set buyer receives the set of tender information sent by the data set seller, eliminates the tenders in the set with a bid price exceeding the available budget, obtains the set of valid tenders meeting the budget requirements, and based on the buyer's need to customize the weights of reputation and price, with the predefined reputation weight being 0.5, scores each tender in the set of valid tenders to obtain the seller with the highest score.
[0025] Optionally, during the scoring process, the following formula is used:
[0026]
[0027] where, bs i (t) is the matching score of the i-th data seller at the t-th round, w R is the predefined reputation weight, is the reputation value score of the i-th data seller, R min represents the lowest reputation value score among the data sellers participating in this tender, R max represents the highest reputation value score among the data sellers participating in this tender, ∈ represents a positive number infinitely close to 0, take ∈ = 1e - 6, b max represents the highest bid value among the data sellers participating in this tender round, b min represents the lowest bid value among the data sellers participating in this tender round, b i (t) represents the bid price.
[0028] Optionally, during the process of poisoning verification, it includes:
[0029] Classify and train the data set for testing. When the test misclassification rate is lower than the threshold, it is determined as a normal seller and the verification passes. If the test misclassification rate is greater than the threshold, it is determined as an abnormal seller and the verification fails.
[0030] Optionally, during the process of calculating the reputation value reward score of the bidder, the following formula is used for calculation:
[0031] ΔR_i(t) = μ × ln(n + 1) × Im + (1 - μ)ln(n + 1)
[0032] where, ΔR_i(t) is the reputation value reward score of the i-th data seller at the t-th round, μ is the reward coefficient, n is the data set sample size, and Im is the class balance rate.
[0033] Optionally, during the process of calculating the reputation value penalty score of the bidder, the following formula is used for calculation:
[0034]
[0035] Among them, is the reputation value penalty score of the i-th data seller at the t-th round. λ is the penalty coefficient, s is the quantization factor, representing the number of pre-trained models suspected of poisoning, and Im is the class balance rate.
[0036] An incentive method for a data trading market that integrates reputation and auction strategies provided by the present invention has the beneficial effects as follows:
[0037] 1. The present invention calculates the reputation value score of the current seller's dataset by combining the reputation value score of the market and the historical reputation score of the seller, linking the quality of the data with the reputation score of the seller, and effectively motivating the seller to provide higher-quality datasets;
[0038] 2. The present invention designs a reverse auction mechanism, enabling the buyer to select the most suitable dataset seller within the budget, effectively enhancing the interest requirements of both parties while meeting the market demands of both buyers and sellers;
[0039] 3. The present invention can enable the seller to actively provide higher-quality datasets in order to obtain a higher reputation score in this mechanism environment through the reputation value score update mechanism during the market trading process, ensuring the security of the trading market. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 is a flowchart of the incentive method for the data trading market provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein shall have the ordinary meaning as understood by those of ordinary skill in the art in the field to which the present invention belongs. The words such as "including" used herein mean that the elements or items appearing before this word cover the elements or items listed after this word and their equivalents, without excluding other elements or items.
[0042] The embodiments of the present invention provide an incentive method for a data trading market that integrates reputation and auction strategies. Refer to Figure 1 , including:
[0043] S1. Obtain the sample feature representation of the candidate seller dataset based on the category center distance of the clean poisoning attack and the label value, and the distance of the target sample close to the category center of the poisoning sample in the feature space, and calculate the quality score of the candidate seller dataset based on the sample feature representation;
[0044] S2. Calculate and update the reputation value score of the candidate seller dataset based on the quality score of the candidate seller dataset. The candidate seller participates in the bidding and submits a bid price based on the reputation value score;
[0045] S3. Design an auction mechanism based on the reputation value score and the price. The buyer matches the seller with the highest score within the budget cost based on the auction mechanism and assigns the dataset task to the seller with the highest score;
[0046] S4. Conduct a poisoning verification on the dataset content of the seller with the highest score. If the verification passes, the buyer and the seller conduct a transaction, calculate the reputation value reward score of the bidder and give a reward. If the verification fails, no transaction is conducted, calculate the reputation value penalty score of the bidder and deduct it.
[0047] In some embodiments, during the execution of step S1, it includes:
[0048] S1.1. Obtain the sample feature representation of the candidate seller dataset;
[0049] S1.2. Calculate the quality score of the candidate seller dataset.
[0050] Specifically, during the execution of step S1.1, in the process of obtaining the sample feature representation of the candidate seller dataset, it includes: According to the attack characteristics of the clean poisoning attack, compared with the category center distance of the label value, the distance of the target sample is closer to the category center of the poisoning sample in the feature space. Therefore, the sample feature representation of the dataset can be extracted through a pre-trained model, so as to evaluate the quality of the dataset according to the potential poisoning situation of the sample.
[0051] Further, when executing step S1.2 and calculating the quality score of the candidate seller dataset based on the sample feature representation, the following formula is used:
[0052]
[0053] f(Q) = a α ×Im β ;
[0054] Where, is the quality score of the dataset, C represents the adjustment coefficient, n represents the sample size of the dataset, a is the label annotation accuracy, and its value range is (0, 1). Im is the class balance rate, and its value range is (0, 1). α and β are adjustment parameters used to adjust the influence of accuracy and class balance rate on the final quality score.
[0055] In some embodiments, during the execution of step S2, it includes:
[0056] S2.1. Calculate and update the credit value score of the candidate seller dataset based on the quality score of the candidate seller dataset;
[0057] S2.2. The candidate seller participates in the bidding and submits a bid price based on the credit value score.
[0058] Specifically, during the execution of step S2.1, when calculating and updating the credit value score of the candidate seller dataset based on the quality score of the dataset, it includes: Default the initial credit value score of all newly added dataset sellers to 1, and calculate and update the credit value score of the candidate seller dataset based on the historical credit score and the credit value scores of existing other dataset sellers. The calculation formula is as follows:
[0059]
[0060] Wherein, is the credit value score of the i-th data seller at the t-th round, is the quality score of the dataset, σ is the historical weight parameter, is the historical credit score.
[0061] Furthermore, when the quality score of the dataset at the t-th round is at the top 50% level of the candidate seller dataset, the value range of σ is (0, 0.5); when the quality score of the dataset at the t-th round is at the bottom 50% level of the candidate seller dataset, the value range of σ is [0.5, 1).
[0062] In fact, the historical credit value adopts the principle of time decay influence, that is, the more recent historical credit value scores have a greater influence and are more important; the older historical credit scores have a smaller influence and are less important.
[0063] Specifically, during the execution of step S2.2, when the candidate seller participates in the bidding and submits a bid price based on the credit value score, it includes: Randomly give an initial bid price, and construct the bidding strategy of the seller based on the credit value score and the initial bid price of the candidate seller; Calculate the utility function of the seller based on the data collection, cleaning, and storage calculation costs, based on the cost, bid price, and changes in credit value scores; Obtain the equilibrium strategies of other sellers, and calculate the maximum expected revenue of the seller based on the equilibrium strategies and the utility function to obtain the best bidding price.
[0064] In fact, the bidding strategy of the data set seller each time is The goal of the seller is to maximize the expected revenue, and the utility function of the seller is designed as:
[0065]
[0066] where γ is a reputation conversion weight parameter used to convert the amount of reputation improvement into equivalent economic revenue. A higher γ value means that reputation has a greater impact on revenue, and ΔR i (t) is the difference in reputation score values, and b i (t) is the initial bid price, and C i (t) is the cost of data collection, cleaning, storage, etc.
[0067] Furthermore, given the equilibrium strategies of other sellers, in order to maximize their own expected revenue, that is, the seller does not know all market information (such as the true costs, quality, reputation, etc. of other sellers), and can only estimate the expected revenue based on their own information, so as to formulate the best bidding strategy.
[0068] In some embodiments, during the execution of step S3, it includes:
[0069] S3.1. Design an auction mechanism based on the reputation score and price;
[0070] S3.2. The buyer matches the seller with the highest score based on the auction mechanism within the budget cost range and assigns the data set task to the seller with the highest score.
[0071] Specifically, during the execution of step S3.1, in the process of designing an auction mechanism based on the reputation score and price, it includes: determining the total budget for the data set tender of the buyer this time. The available budget for a single round is half of the total budget. The data set buyer receives the set of bidding information sent by the data set seller, eliminates the bids in the set whose bid prices exceed the available budget, obtains a set of valid bids that meet the budget requirements, and based on the need of the buyer to customize the weights of reputation and price, with the predefined reputation weight being 0.5, scores each bid in the set of valid bids to obtain the seller with the highest score.
[0072] Furthermore, during the scoring process, the following formula is used:
[0073]
[0074] where bs i (t) is the matching score of the i-th data seller at the t-th round, w R is the predefined reputation weight, is the reputation score value of the i-th data seller, and R minRepresents the lowest credit value score among the data sellers participating in this tender, R max Represents the highest credit value score among the data sellers participating in this tender. ∈ represents a positive number infinitely close to 0. Take ∈ = 1e-6, b max Represents the value of the highest bid among the data sellers participating in this round of tender, b min Represents the value of the lowest bid among the data sellers participating in this round of tender, b i (t) represents the bid price.
[0075] In some embodiments, during the execution of step S4, it includes:
[0076] S4.1. Poison-check the content of the dataset of the seller with the highest score;
[0077] S4.2. Update the credit value score based on the poison-check.
[0078] Specifically, during the execution of step S4.1, when performing the poison-check on the content of the dataset of the seller with the highest score, it includes: Classify and train-test the dataset. When the test misclassification rate is lower than the threshold, it is determined as a normal seller, the check passes, and the buyer and seller conduct a transaction. If the test misclassification rate is greater than the threshold, it is determined as an abnormal seller, the check fails, and no transaction is conducted.
[0079] Furthermore, during the execution of step S4.2, it includes:
[0080] S4.2.1. Calculate the credit value reward score of the bidder and give a reward;
[0081] S4.2.2. Calculate the credit value penalty score of the bidder and make a deduction.
[0082] Specifically, during the execution of step S4.2.1, when calculating the credit value reward score of the bidder, the following formula is used for calculation:
[0083] ΔR_i(t) = μ × ln(n + 1) × Im + (1 - μ)ln(n + 1)
[0084] Where, ΔR_i(t) is the credit value reward score of the i-th data seller at the t-th round, μ is the reward coefficient, n is the dataset sample size, and Im is the class balance rate.
[0085] Furthermore, add the credit value reward score to the credit value score of the seller, so that the seller can obtain a higher credit value, maintain market competitiveness, increase the probability of successful bidding, and thus obtain economic benefits.
[0086] Specifically, during the process of performing step S4.2.2 to calculate the reputation value penalty score of the bidder, the following formula is used for calculation:
[0087]
[0088] Among them, is the reputation value penalty score of the i-th data seller at the t-th round. λ is the penalty coefficient, s is the quantization factor, representing the number of pre-trained models suspected of poisoning, and Im is the class balance rate.
[0089] Furthermore, based on the reputation value penalty score, the reputation value score of the seller is deducted, resulting in the damage of the seller's income and reputation, making it difficult for the poisoner to maintain market competitiveness and ultimately being eliminated.
[0090] In fact, the goal of both parties participating in the transactions in the dataset trading market is to obtain benefits. The designed mechanism, through the quality scoring and reputation penalty mechanisms, malicious SPs cannot make long-term profits; the verification of the clean label poisoning attack will prevent the transaction from proceeding, which will lead to the damage of the seller's income and reputation, making it difficult for the poisoner to maintain market competitiveness and ultimately being eliminated. Then, the goal of mitigating the clean label poisoning attack can be achieved from the perspective that the clean label poisoning attack will lead to the damage of the seller's income.
[0091] Furthermore, the mechanism design conforms to two constraints: IC (Incentive Compatibility) and IR (Individual Rationality). Because only by providing good dataset services can the dataset seller obtain a higher reputation value and have a greater probability of successfully bidding, so as to obtain economic benefits; the utility of the dataset seller must be greater than or equal to 0, because in order to ensure the profit margin, the asking price b i (t) must be greater than C i (t), and there is also the benefit of reputation. The dataset buyer participates in the auction and can meet the dataset requirements, thus obtaining benefits.
[0092] Although the implementation manners of the present invention have been described in detail above, it is obvious to those skilled in the art that various modifications and changes can be made to these implementation manners. However, it should be understood that such modifications and changes are all within the scope and spirit of the present invention described in the claims. Moreover, the present invention described herein can have other implementation manners and can be implemented or realized in various ways.
Claims
1. An incentive method for a data trading market that integrates reputation and auction strategies, characterized in that, Including the following steps: Obtain the sample feature representation of the candidate seller dataset based on the class center distance between the clean poisoning attack and the label value, and the distance of the target sample close to the class center of the poisoning sample in the feature space, and calculate the quality score of the candidate seller dataset based on the sample feature representation; Calculate and update the reputation value score of the candidate seller dataset based on the quality score of the candidate seller dataset. The candidate seller participates in the bidding and submits a bid price based on the reputation value score; Design an auction mechanism based on the reputation value score and price. The buyer matches the seller with the highest score within the budget cost based on the auction mechanism and assigns the dataset task to the seller with the highest score; Perform poisoning verification on the dataset content of the seller with the highest score. If the verification passes, the buyer and the seller conduct a transaction, calculate the reputation value reward score of the bidder and give a reward. If the verification fails, no transaction is conducted, and the reputation value penalty score of the bidder is calculated and deducted.
2. The data trading market incentive method integrating reputation and auction strategies according to claim 1, characterized in that, When calculating the quality score of the candidate seller dataset based on the sample feature representation, the following formula is used: f(Q) = a α × Im β ; Among them, is the quality score of the data set, C represents the adjustment coefficient, n represents the sample size of the data set, a is the label annotation accuracy, and its value range is (0, 1), Im is the class balance rate, and its value range is (0, 1). α and β are adjustment parameters used to adjust the influence of accuracy and class balance rate on the final quality score.
3. The data trading market incentive method integrating reputation and auction strategy according to claim 1, characterized in that In the process of calculating and updating the reputation value score of the candidate seller dataset based on the quality score of the dataset, it includes: Default the initial reputation value score of all newly added dataset sellers to 1, and calculate and update the reputation value score of the candidate seller dataset based on the historical reputation score and the reputation value scores of other existing dataset sellers. The calculation formula is as follows: Among them, is the credit score of the i-th data seller at the t-th round, is the quality score of the dataset, and σ is the historical weight parameter, is the historical credit score.
4. The data trading market incentive method integrating reputation and auction strategy according to claim 3, characterized in that When the quality score of the dataset in the t-th round is at the top 50% level of the candidate seller dataset, the value range of σ is (0, 0.5). When the quality score of the dataset in the t-th round is at the bottom 50% level of the candidate seller dataset, the value range of σ is [0.5, 1).
5. The data trading market incentive method integrating reputation and auction strategies according to claim 1, characterized in that, In the process of the candidate seller participating in the bidding and submitting a bid price based on the reputation value score, it includes: Randomly give an initial bid price, and construct the bidding strategy of the seller based on the reputation value score and the initial bid price of the candidate seller; Based on the data collection, cleaning, and storage calculation costs, calculate the utility function of the seller based on the costs, bid prices, and changes in reputation value scores; Obtain the equilibrium strategies of other sellers, calculate the maximum expected revenue of the seller based on the equilibrium strategies and the utility function, and obtain the best bid price.
6. The incentive method for a data trading market integrating reputation and auction strategies according to claim 1, wherein In the process of designing an auction mechanism based on the reputation value score and price, it includes: Determine the total budget for the dataset tender of the buyer. The available budget for a single round is half of the total budget. The dataset buyer receives the set of bidding information sent by the dataset seller, eliminates the bids with bid prices exceeding the available budget in the set, obtains a set of valid bids that meet the budget requirements, and based on the weights of reputation and price that the buyer needs to customize, with the predefined reputation weight being 0.5, score each bid in the set of valid bids to obtain the seller with the highest score.
7. The data trading market incentive method integrating reputation and auction strategies according to claim 6, characterized in that In the process of scoring, the following formula is used: Among them, bs i (t) is the matching score of the i-th data seller at the t-th round, w R is the predefined reputation weight, is the reputation value score of the i-th data seller, R min represents the lowest reputation value score among the data sellers participating in this tender, R max represents the highest reputation value score among the data sellers participating in this tender, ∈ represents a positive number infinitely close to 0, take ∈ = 1e - 6, b max represents the value of the highest bid among the data sellers participating in this round of tender, b min represents the value of the lowest bid among the data sellers participating in this round of tender, b i (t) represents the bid price.
8. The data trading market incentive method integrating reputation and auction strategies according to claim 1, characterized in that, In the process of performing poisoning verification, it includes: Conduct classification training and testing on the dataset. When the test misclassification rate is lower than the threshold, it is determined as a normal seller and the verification passes. If the test misclassification rate is greater than the threshold, it is determined as an abnormal seller and the verification fails.
9. The data trading market incentive method integrating reputation and auction strategy according to claim 1, characterized in that, When calculating the reputation value reward score of the bidder, the following formula is used for calculation: ΔR_i(t) = μ × ln(n + 1) × Im + (1 - μ)ln(n + 1) Where, ΔR_i(t) is the reputation value reward score of the i-th data seller at the t-th round, μ is the reward coefficient, n is the dataset sample size, and Im is the class balance rate.
10. The data trading market incentive method integrating reputation and auction strategies according to claim 1, characterized in that, In the process of calculating the reputation value penalty score of the bidder, the following formula is used for calculation: Among them, is the credit value penalty score of the i-th data seller at the t-th round. λ is the penalty coefficient, s is the quantization factor, representing the number of pre-trained models suspected of poisoning, and Im is the class balance rate.