An end-to-end distributed model trading method

Abstract

The present invention relates to the field of data element transaction technology, and provides an end-to-end distributed model transaction method, focusing on privacy protection and market participant incentives. It includes: 1. Model training and quality assessment and data compensation allocation under distributed training. Brokers evaluate model quality according to data information and privacy requirements of data providers under distributed training, and allocate data compensation according to their individual characteristics to motivate them to participate in training; 2. Model bidding and version control under information asymmetry. Brokers issue multi-round dynamic bidding tasks to potential model buyers, and determine the optimal production strategy for each version of quality model based on model characteristics and profit maximization principles. The present invention constructs an end-to-end distributed multi-round dynamic bidding transaction market with models as transaction targets. While considering the strategy game of the three main market entities, the actual characteristics of the participants and the special properties of the model are embedded in the market pricing and production strategies to build a safe, complete and sustainable transaction market.

Description

Technical Field

[0001] The present invention relates to the field of data element transaction technology, and in particular, to an end-to-end distributed model transaction method. Background Art

[0002] Past research has conducted extensive explorations in data pricing and data trading market design. Early researchers compared data assets to commodities or financial assets, and tried to design data trading markets by combining data characteristics (such as data volume, data type, and acquisition cost) through economic pricing theory. However, this method of directly pricing source data may bring the risk of privacy leakage to data providers. In order to consider data privacy security, some researchers have begun to try to price data based on the degree of privacy leakage, or to price data based on the content of specified queries. However, the former faces difficulties due to the difficulty in accurately quantifying the cost of privacy leakage, while the latter is difficult to adapt to real market needs due to its lack of flexibility. It is worth noting that the above-mentioned trading markets are all priced for raw data.

[0003] With the rapid development of digital technology, the predictive value of data has become particularly important. From a data-driven perspective, researchers have proposed a data market with machine learning models as the subject of transactions. Its characteristics are that data providers can separate data ownership and predictive value, and model buyers only need to pay for the predictive value they are interested in. The security and low cost of the model trading market simultaneously achieve effective incentives for both supply and demand. It is worth noting that the existing model trading markets are all centralized training, and data providers need to upload data to the central training point. As the amount of training data increases, centralized training may be difficult to support the huge computing costs due to the limitations of computing power and storage resources. In addition, centralized training requires data providers to upload data to untrusted training centers, and the potential risk of data leakage makes it difficult to motivate them to participate.

[0004] On the other hand, the mechanism design of the model trading market is also insufficient. In terms of incentive mechanism design, existing research has the problem of insufficient depth and breadth of incentive mechanisms. Most studies only focus on the incentives of local individuals in the market. In fact, the model trading market is an interactive whole and does not satisfy incentive compatibility. For example, the cost minimization of brokers and the data compensation of data providers are mutually exclusive. In addition, the incentive mechanism should be more extensive and specific, and it is necessary to consider the multiple properties of market participants at the same time, such as privacy needs, fairness and individual rationality, which can better reflect the real market participants. In terms of the design of pricing mechanisms, although there are many existing pricing mechanisms, including those based on model accuracy, the principle of no arbitrage, and the Shapley value coverage. However, these methods only price for a certain property of the model. In fact, the pricing mechanism of the model trading market should take into account the uniqueness of the model in more aspects, such as replicability, additivity, homogeneity, asymmetric information, etc. These characteristics make model products different from other products.

[0005] Finally, the real trading market is a dynamic and complex system of multi-party games, and each participant pursues a strategy of maximizing benefits under the condition of satisfying individual rationality. However, few existing studies can fully describe the dynamic game process of market players, which limits the understanding of real market participants and hinders the further improvement and optimization of the transaction mechanism of the data trading market. Summary of the invention

[0006] The present invention provides an end-to-end distributed model transaction method, which can overcome the defects of the prior art in terms of privacy protection and participant incentives.

[0007] The present invention provides an end-to-end distributed model transaction method, which comprises the following steps:

[0008] 1. Model training and quality evaluation and data compensation allocation under distributed training. Under distributed training, brokers evaluate model quality based on the data information and privacy requirements of potential data providers, and allocate data compensation based on their individual characteristics to encourage them to participate in training;

[0009] 2. Model bidding and version control under information asymmetry. The broker issues multiple rounds of dynamic bidding tasks to potential model buyers and determines the optimal production strategy for each version of the quality model based on model characteristics and profit maximization principles.

[0010] As a preferred method, the model training and quality assessment process under distributed training is specifically as follows:

[0011] In distributed training, brokers query data providers for data gradients instead of raw data through distributed training. Data providers add noise to the gradients in each round of communication to protect privacy. Brokers determine the quality of the model under distributed training based on the privacy requirements of data providers and the status of their data.

[0012] For the privacy protection of data providers, assume that in the t∈T round of brokers’ gradient query to data providers, the i∈N data provider responds to the gradient query in the local data with a gradient response R t (D i ) adds independent and identically distributed Laplace noise, and the noise distribution satisfies Then the noise of the data provider satisfies Differential privacy mechanism, where ρ = maxg (x，y) (w t , D i ), β is the variance of the Laplace noise distribution; where ∈ is the privacy budget, g (x，y) (w t , D i ) represents the sub-gradient response of the i-th data provider data under the current t-round gradient query, g (x，y) The g in the formula refers to the sub-gradient, (x, y) is the specific data point, and w t Refers to the parameters of the model during the t-round gradient query, L i is the amount of data from the i-th data provider, and β is the variance of the Laplace noise distribution; after the data provider determines the privacy protection, the broker determines the performance convergence of the distributed training model based on the data status and privacy requirements of the data provider;

[0013] The performance convergence relationship of model training quality Q under distributed training is: Among them, the increase of data size, the increase of privacy budget ∈ and the increase of the total number of data providers N all help to improve the quality of the training model; further, when there is (∈n +1 , l n+1 ) data provider participates, the marginal impact of the distributed training model quality is in Represents the cumulative sum of the data volume, represents the sum of the reciprocals of the squares of the privacy budget; E(Q n+1 -Q n ) indicates an improvement in model quality.

[0014] As a preference, the data compensation allocation under distributed training is specifically as follows:

[0015] The value of data provided by different data providers has a great impact on model training, and brokers need to select the most valuable data; at the same time, brokers must ensure reasonable remuneration to encourage the active participation of data providers while minimizing costs; finally, the entire game process must reflect the characteristics of real market participants, which requires incorporating the attributes of brokers and data providers, such as individual rationality and fairness, as constraints into the game framework;

[0016] The broker's compensation strategy c(g i , fi) and the compensation strategy s(D i , x i ) should satisfy the following properties:

[0017] a) Non-negativity

[0018] b) Fairness

[0019] c) Monotonicity

[0020] d) individual rationality;

[0021] Data provider rationality: Broker rationality:

[0022] The above individual characteristics show that the goal of the broker is to minimize the cost while motivating the data providers to participate by establishing different compensation schemes based on their data conditions, privacy requirements, and expected compensation; using the greedy selection-based minimization compensation algorithm, i.e., Algorithm 1, to screen the most valuable data providers and perform differentiated pricing for data providers with different data conditions and privacy budgets;

[0023] 1.1) The data providers to be selected shall be The weights are arranged in ascending order;

[0024] 1.2) According to the weight, data providers are added to participate in the training, and Calculate model quality;

[0025] 1.3) Repeat until the target model quality is achieved or the maximum number of data providers is reached;

[0026] 1.4) Construct a set of expected compensation per unit data volume from the selected data providers in descending order of privacy budget;

[0027] 1.5) For the current data provider with ∈i privacy budget, ensure that the compensation per unit of data is not less than the compensation of the previous data provider;

[0028] 1.6) The loop is iterated to traverse all selected data providers.

[0029] As a preferred method, the model bidding under information asymmetry is as follows:

[0030] For the model buyer, assume that the market has V∈N + potential model buyers, who expect to purchase a model that meets performance requirements and is within their budget. For any model buyer j∈, there are two strategies, namely, the expected model quality m j and bid willingness y j , the pricing mechanism of a single model pricing under asymmetric information is Among them, I(m j ≤Q) represents the model mass m j Matches the desired quality of the jth buyer, that is, if m j ≤Q is 1, otherwise it is 0; I(y j ≥p) represents the bid y of the jth potential buyer j If it is greater than the model price p, it is 1, otherwise it is 0; p is the possible bid of each potential buyer. If the model quality is Q, the model buyer whose bid is not less than p and meets the model quality requirements will succeed in bidding and obtain the model at a price of p.

[0031] As a preferred method, version control under information asymmetry is as follows:

[0032] Providing models of differentiated quality at different prices through version control to meet the needs of model buyers with different needs; Different from the single model pricing mechanism, the multi-version pricing mechanism needs to consider the characteristics of the model at the same time, including replicability, homogeneity, additivity and asymmetric information;

[0033] (1) Replicability: For any model with quality Q, its profit E(Q) is expressed by the following formula:

[0034]

[0035] (3) Homogeneity: Each potential buyer of a model is only willing to purchase models of a certain quality. In a multi-round bidding, if a model buyer wins in any round of bidding, they will withdraw from the next round; otherwise, they will continue to participate in the bidding until the end. For brokers, each model of each quality will only bid once.

[0036] (3) Additivity: The no-arbitrage principle is embedded in the broker’s profit maximization problem as follows:

[0037]

[0038] Among them, δ is the monotonicity distance, which is usually a small value; a, b, c are any number of bidding rounds; for any t-th bidding round, in order to maximize the broker's profit under the no-arbitrage constraint, Algorithm 2 is designed, as follows:

[0039] 2.1) In the tth round of bidding, the candidate model set Iterate to find the point where maximum profit is generated;

[0040] 2.2) Calculate the current quality Costs;

[0041] 2.3) Find the current quality in the historical bidding information adjacent quality and price;

[0042] 2.4) Calculate the current quality The best price;

[0043] 2.5) Calculate the optimal price pe that satisfies the no-arbitrage pricing. If pe is within the no-arbitrage interval, the current value can be used, otherwise only the boundary value can be used;

[0044] 2.6) Calculate the profit E(Qi) of the optimal pricing and set the winning bidder

[0045] 2.7) Determine whether the production point is optimal by whether the profit increases; at the end of the loop, iterate to the quality of the last candidate model;

[0046] (4) Asymmetric information: Based on the actual behavior of buyers in each round of bidding, a dynamic reserve price strategy is proposed, and the dynamic reserve price is embedded into the broker’s profit maximization problem:

[0047]

[0048] The parameter lim represents the intensity of executing the dynamic price retention strategy. The higher the lim value, the stronger the price protection strategy is and the greater the restriction on the buyer's strategic behavior. The variable p lim represents the price of the lim quantile of the bids of all model buyers in the current round of bidding; in order to maximize the broker's profit under asymmetric information, Algorithm 3 is proposed, as follows:

[0049] 3.1) In the tth round of bidding, the candidate model set Iterate to find the production point with maximum profit;

[0050] 3.2) Calculate the current quality Costs;

[0051] 3.3) Calculate the boundary price of the bidding limit;

[0052] 3.4) Calculate the current quality The best price;

[0053] 3.5) Calculate the price range that meets the bidding restrictions;

[0054] 3.6) Calculate the profit E(Qi) of the optimal pricing and set the winning bidder

[0055] 3.7) Determine whether the production point is optimal by whether the profit increases; at the end of the cycle, iterate to the quality of the last candidate model.

[0056] The present invention designs a model trading market framework based on distributed learning, which takes into account three parties. Its complete model dynamic trading process includes model training, quality assessment, reward distribution, model bidding, version control, etc. The framework is decoupled from two interactive frameworks: in the interaction between brokers and data providers, in order to ensure security and reduce computing pressure, distributed training is introduced. According to the model performance convergence relationship in distributed training, the model quality is evaluated by using the number of data providers, the amount of data of each person, and the privacy protection requirements of each person, and focusing on privacy and incentive issues, a greedy minimization compensation method based on economics that takes into account privacy sensitivity, fairness, and individual rationality attributes is designed to maximize incentives for data providers and achieve a balance between cost and incentive. In the interaction between brokers and model buyers, in order to deal with the problem of information asymmetry, a multi-round dynamic bidding strategy with models as the trading subject is constructed. The strategy takes into account the specific attributes of the model, such as reproducibility, additivity, and homogeneity, and also takes into account the strategic behavior of buyers, achieving a balance between cost and benefit. In summary, the present invention integrates the multifaceted attributes of market participants and the inherent complexity of model pricing to construct a complete, secure and efficient model trading market. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A flowchart of an end-to-end distributed model transaction method in an embodiment;

[0058] Figure 2 This is a schematic diagram of the impact of different data provider selection methods on the total remuneration of brokers when N=1000 in the embodiment;

[0059] Figure 3 This is a schematic diagram of changes in model quality during the process of screening data providers in the embodiment;

[0060] Figure 4 It is a schematic diagram of the impact of no-arbitrage pricing on broker pricing in the embodiment;

[0061] Figure 5 This is a schematic diagram of the impact of the capacity limitation strategy on broker profits in the embodiment;

[0062] Figure 6 Schematic diagram of profit loss caused by applying the no-arbitrage constraint δ=[0.20, 0.25, 0.30] in the embodiment;

[0063] Figure 7 In the embodiment, when τ t Schematic diagram of the impact of different types of buyers in the market on brokers’ production strategies when δ = 0.25 and δ = 0.2;

[0064] Figure 8 Schematic diagram of the impact of buyer allocation on profits in the embodiment. DETAILED DESCRIPTION

[0065] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0066] Example

[0067] like Figure 1 As shown, this embodiment provides an end-to-end distributed model transaction method, which includes the following steps:

[0068] 1. Model training and quality evaluation and data compensation allocation under distributed training. Under distributed training, brokers evaluate model quality based on the data information and privacy requirements of potential data providers, and allocate data compensation based on their individual characteristics (such as individual rationality, fairness, etc.) to encourage them to participate in training;

[0069] 2. Model bidding and version control under information asymmetry. The broker issues multiple rounds of dynamic bidding tasks to potential model buyers and determines the optimal production strategy for each version of the quality model based on model characteristics and profit maximization principles.

[0070] Model training, quality assessment, and data compensation allocation under distributed training are as follows:

[0071] In distributed training, brokers query data providers for data gradients instead of raw data through distributed training. Data providers add noise to the gradients in each round of communication to protect privacy. Brokers determine the quality of the model under distributed training based on the privacy requirements of data providers and the status of their data.

[0072] Assume that N∈N in the real market + Potential data providers have a dataset set D = {D1, D2, ..., D N}, where the dataset of the i∈Nth data provider is D i ∈D; When the broker publishes a distributed training task, it passes T∈N +Round communication aggregates the data gradients of all data providers, and the data providers return the gradient response R according to the gradient query. t (D i ); data providers can use differential privacy to t (D i ) Add random noise to blur the gradient information to achieve privacy protection;

[0073] If the privacy budget of the i∈Nth data provider is ∈, and the amount of data is l i , then it can be in the gradient response R t (D i ) to protect privacy. When the noise distribution satisfies Then it is satisfied Differential privacy mechanism, where ρ = maxg (x，y) (w t , D i ), β is the variance of the Laplace noise distribution, which realizes the privacy protection requirements of the data provider;

[0074] After determining the privacy protection of the data provider, the broker can determine the performance convergence of the distributed training model based on the data status and privacy requirements of the data provider; specifically, the performance convergence relationship of the model training quality Q under distributed training is: The increase in data size, the increase in privacy budget ∈, and the increase in the total number of data providers N all help to improve the quality of the training model; further, it can be calculated that when the n+1th data provider is added, the marginal change in model quality is E(Q n+1 -Q n ) indicates an improvement in model quality, with Data providers with larger values ​​have greater marginal positive contributions to model training; Represents the cumulative sum of the data volume, represents the sum of the reciprocals of the squares of the privacy budget; E(Q n+1 -Q n ) indicates an improvement in model quality.

[0075] The value of data provided by different data providers has a great impact on model training, and brokers need to select the most valuable data; at the same time, brokers must ensure reasonable remuneration to encourage the active participation of data providers while minimizing costs; finally, the entire game process must reflect the characteristics of real market participants, which requires incorporating the attributes of brokers and data providers, such as individual rationality and fairness, as constraints into the game framework;

[0076] The broker's compensation strategy c(g i , fi) and the compensation strategy s(D i , X i ) should satisfy the following properties:

[0077] a) Non-negative x i ≥0,

[0078] b) Fairness

[0079] c) Monotonicity

[0080] d) individual rationality;

[0081] Data provider rationality: Broker rationality:

[0082] Now consider compensating the data providers who participate in the training according to their training contributions. For any i∈N data provider with a dataset D i =D(∈ i , l i ); Any i-th data provider has an expected compensation x i , expected compensation x i With dataset D i Together they constitute the expected compensation strategy s(D i , x i ), indicating that the i-th data provider expects to learn from its dataset D during training i Get no less than x i compensation; for brokers, c(g i , f i ) represents the compensation strategy of the broker to the i-th data provider; in order to formulate a compensation strategy that satisfies non-negativity, fairness, monotonicity and individual rationality, it can be calculated by running Algorithm 1.

[0083]

[0084] 1.1) The data providers to be selected shall be The weights are arranged in ascending order;

[0085] 1.2) According to the weight, data providers are added to participate in the training, and Calculate model quality;

[0086] 1.3) Repeat until the target model quality is achieved or the maximum number of data providers is reached;

[0087] 1.4) Construct a set of expected compensation per unit data volume from the selected data providers in descending order of privacy budget;

[0088] 1.5) For the current data provider with ∈i privacy budget, ensure that the compensation per unit of data is not less than the compensation of the previous data provider;

[0089] 1.6) The loop is iterated to traverse all selected data providers.

[0090] Model bidding and version control under information asymmetry, specifically:

[0091] Assume that the market has V∈N + potential model buyers, who expect to purchase a model that meets performance requirements and is within their budget. For any model buyer j∈V, there are two strategies: the expected model quality m j and bid willingness y j , together constitute their personal strategy b(y j , m j ), represents the model buyer j is willing to pay no more than y j The payment purchase quality satisfies m j ≥Q; For any model with quality Q, the broker determines the pricing mechanism of a single version of the quality Q model based on the bids of all buyers as Among them, I(m j ≤Q) represents the model mass m j Matches the desired quality of the jth buyer, that is, if m j ≤Q is 1, otherwise it is 0; I(y j ≥p) represents the bid y of the jth potential buyer j If it is greater than the model price p, it is 1, otherwise it is 0; p is the possible bid of each potential buyer. If the model quality is Q, the model buyer whose bid is not less than p and meets the model quality requirements will succeed in bidding and obtain the model at a price of p.

[0092] Brokers use version control to provide models of differentiated quality at different prices to meet the needs of model buyers with different needs. At this time, the pricing mechanism needs to consider the four properties of the model, namely, replicability, homogeneity, additivity, and asymmetric information.

[0093] (1) Replicability: In a pricing mechanism that satisfies replicability, the profit E(Q) of any model with quality Q is expressed by the following formula:

[0094]

[0095] (2) Homogeneity: Each potential buyer of a model is only willing to purchase models of a certain quality. In a multi-round bidding, if a model buyer wins in any round of bidding, they will withdraw from the next round; otherwise, they will continue to participate in the bidding until the end. For brokers, each model of each quality will only bid once.

[0096] (3) Additivity: The no-arbitrage principle is embedded in the broker’s profit maximization problem as follows:

[0097]

[0098]

[0099]

[0100] In order to maximize the broker's profit under the no-arbitrage constraint, Algorithm 2 is designed as follows:

[0101]

[0102] 2.1) In the tth round of bidding, the candidate model set Iterate to find the point where maximum profit is generated;

[0103] 2.2) Calculate the current quality Costs;

[0104] 2.3) Find the current quality in the historical bidding information adjacent quality and price;

[0105] 2.4) Calculate the current quality The best price;

[0106] 2.5) Calculate the optimal price p that satisfies no-arbitrage pricing e , if p e If it is within the no-arbitrage interval, the current value can be taken, otherwise only the boundary value can be taken;

[0107] 2.6) Calculate the profit E(Q) of the optimal pricing i ) and set the winning bidder

[0108] 2.7) Determine whether the production point is optimal by whether the profit increases; at the end of the loop, iterate to the quality of the last candidate model;

[0109] (4) Asymmetric information: Based on the actual behavior of buyers in each round of bidding, a dynamic reserve price strategy is proposed, and the dynamic reserve price is embedded into the broker’s profit maximization problem:

[0110]

[0111] The parameter lim represents the intensity of executing the dynamic price retention strategy. The higher the lim value, the stronger the price protection strategy is and the greater the restriction on the buyer's strategic behavior. The variable p lim represents the price of the lim quantile of the bids of all model buyers in the current round of bidding; in order to maximize the broker's profit under asymmetric information, Algorithm 3 is proposed, as follows:

[0112]

[0113] 3.1) In the tth round of bidding, the candidate model set Iterate to find the production point with maximum profit;

[0114] 3.2) Calculate the current quality Costs;

[0115] 3.3) Calculate the boundary price of the bidding limit;

[0116] 3.4) Calculate the current quality The best price;

[0117] 3.5) Calculate the price range that meets the bidding restrictions;

[0118] 3.6) Calculate the profit E(Qi) of the optimal pricing and set the winning bidder

[0119] 3.7) Determine whether the production point is optimal by whether the profit increases; at the end of the cycle, iterate to the quality of the last candidate model;

[0120] We further integrate the entire multi-round dynamic bidding process into a unified framework and design Algorithm 4. The goal of Algorithm 4 is to maximize the total profit of the broker in all bidding rounds, as follows:

[0121]

[0122] 4.1) Initialize and iterate each round of bidding;

[0123] 4.2) Calculate the optimal production strategy under the current t-th round of bidding;

[0124] 4.3) Update the data provider and candidate model and record the optimal production strategy for each round of bidding; the end condition of the loop is that the profit of the current round of bidding is non-positive.

[0125] experiment

[0126] Numerical simulation experiments are used to verify the effectiveness of the MRDBD model trading market.

[0127] Data providers: Set the number of potential data providers N = [1000, 2000, 3000], the data volume l of each provider follows a uniform distribution of l~U(5,20), and the privacy budget ε follows a uniform distribution of ε~U(0.5,5).

[0128] Broker: In the first interaction, assuming that the number of potential data providers N is [1000, 2000, 3000], the model quality Q constraint set by the broker is [0.0002, 0.00012, 0.00008]; in the second interaction, assuming that the broker sets a total of 30 candidate models, the quality Q of the candidate models is taken from the interval [0.001, 0.007]. The number 30 is taken from the same interval as the model quality setting.

[0129] Model buyers: Assume that there are four types of model buyers in the market, and the number of potential model buyers in the model market is V = 2000. The bid of the first type of buyers is linearly related to the model quality, which belongs to the neutral sensitivity quality preference, that is, P1 = α1-β1*m; the bid of the second type of buyers decreases slowly with the decrease of model quality, which belongs to the low sensitivity quality preference, that is, P2 = α2-β2*m; the bid of the third type of buyers decreases slowly with the decrease of model quality, which belongs to the neutral sensitivity quality preference, that is, P2 = α2-β2*m, P2 = α2-β2*m2; the bid of the third type of buyers decreases rapidly with the decrease of model quality, which belongs to the high sensitivity quality preference, that is, P3 = α3-β3*√m; the fourth type of buyers are only interested in models above a certain quality threshold and are willing to pay a fixed bid, that is The buyer has a fixed sensitivity quality preference. The buyer's parameter settings are all subject to uniform distribution.

[0130] The experiment has two objectives: (1) to test whether Algorithm 1 can help brokers reduce costs and train higher quality models; and (2) to test whether the two pricing strategies are effective in different model buyer distributions.

[0131] a) Training-compensation interaction

[0132] Figure 2 The effect of different data provider selection methods on the total compensation of brokers when N = 1000 is shown. The horizontal axis is the desired model quality, and the vertical axis is the total compensation of brokers: SMC: SMC is the screening method of Algorithm 1. Random: Random means randomly selecting data providers to participate in training. Compensation: Compensation is to select data providers from small to large according to the expected compensation S(D, x) of the data provider.

[0133] It can be seen that as the quality to be achieved in training becomes better and better, the training costs of the three methods gradually increase, but the total compensation cost of the SMC method is much lower than that of the other two methods, and the total compensation cost of the compensation method is always slightly lower than that of the random method, which shows that blindly selecting data providers or selecting data providers based only on expected compensation may incur more costs. In addition, when the model quality is 0.0014, both the compensation method and the random method are crossed out, indicating that the current model quality cannot be achieved through training. This means that the model quality of these two methods will also be worse than that of the SMC method because it does not exclude those data providers with low model training efficiency.

[0134] We also analyzed the changes in model quality during the process of selecting data providers, such as Figure 3 As shown. The horizontal axis represents the number of screened data providers, and the vertical axis represents the model quality value. The smaller the value, the better the model quality. Taking the number of data providers N = 1000 and the target quality Q = 0.002 as an example, we can see that as the number of data providers increases, the model quality value shows a process from rapid decline to slow decline, which is consistent with the actual situation, that is, the model quality decreases with the increase in the amount of data, and the marginal quality improves. On the other hand, the SMC method requires the least number of data providers and reaches the target quality the fastest. The compensation method is second, and the random method is the worst. This is consistent with Figure 2 The costs of the three methods are relatively consistent. In general, the cost increases with the number of participants. In addition, the small window in the figure shows the inflection point (marked with a cross) of the compensation method and the random method at the sampling interval of the number of people. This shows that in reality, when blindly screening data providers, it is difficult to avoid the consequences of selecting some data of poor quality, which leads to increased costs of reducing model quality.

[0135] b) Bidding-Production Interaction

[0136] We study the dynamic bidding process of brokers and model buyers in multiple rounds of bidding. Figure 4 The effect of no-arbitrage pricing on broker pricing is shown. The horizontal axis represents the quality of the model, and the vertical axis represents the pricing of the model. Lines of different colors correspond to the strength of different capacity limit strategies, i.e., limit_buyersτ in Algorithm 3. t =[0.20,0.25,0.30,1.00], where τ t =1 indicates that the capacity restriction strategy is not implemented. The solid line indicates that the no-arbitrage constraint is not implemented. t= [0.20, 0.25, 0.30], the three solid lines corresponding to the quality [0.00617, 0.00596, 0.00307] indicate arbitrage. The dotted line indicates the execution of the no-arbitrage constraint, and the monotonicity distance δ = 0.2 is set in Algorithm 2. After execution, for models of different qualities, pricing is smoother and no arbitrage occurs.

[0137] Figure 5 The effect of the capacity restriction strategy on the broker's profit is illustrated. The horizontal axis represents the number of bidding rounds, the left vertical axis represents the profit per bidding round (corresponding to the line graph), and the right vertical axis represents the total profit (corresponding to the bar graph). It can be seen from the line graph that as the capacity constraint increases, the total number of bidding rounds decreases to [10,8,7,2], but the profit per bidding round increases. This is because the capacity restriction strategy limits the number of bidders in each round. The bar graph shows the effect of the capacity restriction strategy on the total profit. Surprisingly, using the restricted capacity in the right way is expected to generate higher profits. This may be because when the capacity is not restricted, the best bidding price for most buyers tends to be lower, thereby reducing the total profit. In contrast, when capacity restrictions are adopted, models of the same quality are priced higher, and the profits are dispersed in each bidding round, resulting in higher total profits.

[0138] Figure 6 The profit loss caused by applying the no-arbitrage constraint δ = [0.20, 0.25, 0.30] is illustrated. Taking δ = 0.3 as an example, the correction of the arbitrage price occurs when the quality is 0.00307, ​​and therefore corresponds to the profit loss due to the arbitrage price constraint in the second round of bidding. When δ = 0.20 and 0.25, the profit loss due to the arbitrage price constraint occurs after the fifth round of bidding. On the other hand, as the monotonicity distance δ increases, the profit loss of no-arbitrage pricing increases, and the constraints on the arbitrage strategy of the model buyers also increase, resulting in large potential gains that are difficult to estimate.

[0139] Figure 7 It shows the impact of different types of buyers in the market on the broker's production strategy when τt = 0.25 and δ = 0.2. Figure 7 In , U, L, H, and N correspond to the above four types of buyers, where the horizontal axis is the optimal model quality in each round of bidding, and the vertical axis is the optimal price. When all buyers in the market have low-sensitivity quality preferences, since the quality willingness of models with low-sensitivity preferences presents a smooth trend, the optimal price of the model in each round of bidding is higher than the optimal price of the model with uniform and high-sensitivity quality preferences. The other two types of buyers, especially those with high-sensitivity quality preferences, have a sharp downward trend in their vehicle quality willingness. For quality preferences with fixed sensitivity, due to being more picky about model quality, the optimal production quality of each round of bidding is smaller, and models with better quality are models that meet the needs of market buyers. Figure 8 The effect of buyer allocation on profits is explained. When the market buyer is a low-sensitivity quality preference, the profit per bidding round and the total profit are the highest, followed by the fixed-sensitivity quality preference, and the lowest is the high-sensitivity quality preference. The total profit of the fixed-sensitivity quality preference setter is only higher than that of the high-sensitivity quality preference setter.

[0140] This embodiment proposes a machine learning model trading market under distributed training, taking into account three participants: data providers, brokers, and model buyers, and achieving the optimal production strategy and market model pricing under equilibrium through two interactive processes. This embodiment will provide theoretical and practical guidance for data decomposition in the digital economy era, while promoting the sustainable development of the data trading market.

[0141] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.

Claims

1. An end-to-end distributed model transaction method, characterized by: The following steps are involved:

1. Model training and quality evaluation and data compensation allocation under distributed training. Under distributed training, brokers evaluate model quality based on the data information and privacy requirements of potential data providers, and allocate data compensation based on their individual characteristics to encourage them to participate in training; 2. Model bidding and version control under information asymmetry. The broker issues multiple rounds of dynamic bidding tasks to potential model buyers and determines the optimal production strategy for each version of the quality model based on model characteristics and profit maximization principles.

2. The end-to-end distributed model transaction method according to claim 1, characterized in that: The model training and quality assessment process under distributed training is as follows: In distributed training, brokers query data providers for data gradients instead of raw data through distributed training. Data providers add noise to the gradients in each round of communication to protect privacy. Brokers determine the quality of the model under distributed training based on the privacy requirements of data providers and the status of their data. For the privacy protection of data providers, assume that in the t∈T round of brokers’ gradient query to data providers, the i∈N data provider responds to the gradient query in the local data with a gradient response R t (D i ) adds independent and identically distributed Laplace noise, and the noise distribution satisfies Then the noise of the data provider satisfies Differential privacy mechanism, where ρ = maxg (x,y) (w t ,D i ), β is the variance of the Laplace noise distribution; where ∈ is the privacy budget, g (x,y) (w t ,D i ) represents the sub-gradient response of the i-th data provider data under the current t-round gradient query, g (x,y) The g in the formula refers to the sub-gradient, (x, y) is the specific data point, and w t Refers to the parameters of the model during the t-round gradient query, L i is the amount of data from the i-th data provider, and β is the variance of the Laplace noise distribution; after the data provider determines the privacy protection, the broker determines the performance convergence of the distributed training model based on the data status and privacy requirements of the data provider; The performance convergence relationship of model training quality Q under distributed training is: Among them, the increase of data size l, the increase of privacy budget ∈ and the increase of the total number of data providers N are all helpful to improve the quality of training model; further, when there is (∈ n+1 ,l n+1 ) data provider participates, the marginal impact of the distributed training model quality is in Represents the cumulative sum of the data volume, represents the sum of the reciprocals of the squares of the privacy budget; E(Q n+1 -Q n ) indicates an improvement in model quality.

3. The end-to-end distributed model transaction method according to claim 2, characterized in that: The specific data compensation allocation under distributed training is: The value of data provided by different data providers has a great impact on model training, and brokers need to select the most valuable data; at the same time, brokers must ensure reasonable remuneration to encourage the active participation of data providers while minimizing costs; finally, the entire game process must reflect the characteristics of real market participants, which requires incorporating the attributes of brokers and data providers, such as individual rationality and fairness, as constraints into the game framework; The broker's compensation strategy c(g i ,f i ) and the compensation strategy s(D i ,x i ) should satisfy the following properties: a) Non-negativity b) Fairness c) Monotonicity d) individual rationality; Data provider rationality: Broker rationality: The above individual characteristics show that the goal of the broker is to minimize the cost while motivating the data providers to participate by establishing different compensation schemes based on their data conditions, privacy requirements, and expected compensation; using the greedy selection-based minimization compensation algorithm, i.e., Algorithm 1, to screen the most valuable data providers and perform differentiated pricing for data providers with different data conditions and privacy budgets; 1.1) The data providers to be selected shall be The weights are arranged in ascending order; 1.2) According to the weight, data providers are added to participate in the training, and Calculate model quality; 1.3) Repeat until the target model quality is achieved or the maximum number of data providers is reached; 1.4) Construct a set of expected compensation per unit data volume from the selected data providers in descending order of privacy budget; 1.5) The loop iteration is for the current one with ∈ i Data providers with privacy budgets ensure that the compensation per unit of data is not less than the compensation of the previous data provider; 1.6) The loop is iterated to traverse all selected data providers.

4. The end-to-end distributed model transaction method according to claim 3, characterized in that: Model bidding under information asymmetry is as follows: For the model buyer, assume that the market has V∈N + potential model buyers, who expect to purchase a model that meets performance requirements and is within their budget. For any model buyer j∈V, there are two strategies: the expected model quality m j and bid willingness y j , the pricing mechanism of a single model pricing under asymmetric information is Among them, I(m j ≤Q) represents the model mass m j Matches the desired quality of the jth buyer, that is, if m j ≤Q is 1, otherwise it is 0; I(y j ≥p) represents the bid y of the jth potential buyer j If it is greater than the model price p, it is 1, otherwise it is 0; p is the possible bid of each potential buyer. If the model quality is Q, the model buyer whose bid is not less than p and meets the model quality requirements will succeed in bidding and obtain the model at a price of p.

5. The end-to-end distributed model transaction method according to claim 4, characterized in that: Version control under information asymmetry is as follows: Providing models of differentiated quality at different prices through version control to meet the needs of model buyers with different needs; Different from the single model pricing mechanism, the multi-version pricing mechanism needs to consider the characteristics of the model at the same time, including replicability, homogeneity, additivity and asymmetric information; (1) Replicability: For any model with quality Q, its profit E(Q) is expressed by the following formula: (2) Homogeneity: Each potential buyer of a model is only willing to purchase models of a certain quality. In a multi-round bidding, if a model buyer wins in any round of bidding, they will withdraw from the next round; otherwise, they will continue to participate in the bidding until the end. For brokers, each model of each quality will only bid once. (3) Additivity: The no-arbitrage principle is embedded in the broker’s profit maximization problem as follows: Among them, δ is the monotonicity distance, which is usually a small value; a, b, c are any number of bidding rounds; for any t-th bidding round, in order to maximize the broker's profit under the no-arbitrage constraint, Algorithm 2 is designed, as follows: 2.1) In the tth round of bidding, the candidate model set Iterate to find the point where maximum profit is generated; 2.2) Calculate the current quality Costs; 2.3) Find the current quality in the historical bidding information adjacent quality and price; 2.4) Calculate the current quality The best price; 2.5) Calculate the optimal price p that satisfies no-arbitrage pricing e , if p e If it is within the no-arbitrage interval, the current value can be taken, otherwise only the boundary value can be taken; 2.6) Calculate the profit E(Q) of the optimal pricing i ) and set the winning bidder 2.7) Determine whether the production point is optimal by whether the profit increases; at the end of the loop, iterate to the quality of the last candidate model; (4) Asymmetric information: Based on the actual behavior of buyers in each round of bidding, a dynamic reserve price strategy is proposed, and the dynamic reserve price is embedded into the broker’s profit maximization problem: The parameter lim represents the intensity of executing the dynamic price retention strategy. The higher the lim value, the stronger the price protection strategy is and the greater the restriction on the buyer's strategic behavior. The variable p lim represents the price of the lim quantile of the bids of all model buyers in the current round of bidding; in order to maximize the broker's profit under asymmetric information, Algorithm 3 is proposed, as follows: 3.1) In the tth round of bidding, the candidate model set Iterate to find the production point with maximum profit; 3.2) Calculate the current quality Costs; 3.3) Calculate the boundary price of the bidding limit; 3.4) Calculate the current quality The best price; 3.5) Calculate the price range that meets the bidding restrictions; 3.6) Calculate the profit E(Q) of the optimal pricing i ) and set the winning bidder 3.7) Determine whether the production point is optimal by whether the profit increases; at the end of the cycle, iterate to the quality of the last candidate model.

Images