Design method for trusted fair data exchange platform for machine learning data

Abstract

The application discloses a trusted and fair data transaction platform design method for machine learning data, which enables a data purchaser to accurately evaluate the quality of data before purchasing the data, and does not disclose the data privacy of a data provider during the evaluation process.A model owner samples data of the data provider, the data provider submits an encrypted data subset, and the two parties cooperatively train by using a related privacy calculation method, and the quality of the data provided by the data owner is evaluated by the accuracy of the obtained model on test data.The model owner can repeat the process for multiple different data providers, select data providers with higher data quality, and perform subsequent transaction operations.When the transaction is completed, the transaction data cannot be forged, that is, the data evaluated before the transaction is indeed the data delivered by the data provider when the transaction is completed.The application can guarantee the trustworthiness and fairness of the transaction data.

Description

Technical Field

[0001] This invention relates to the field of fair data trading market technology, and in particular to a method for designing a trustworthy and fair data trading platform for machine learning data. Background Technology

[0002] As the value of data, especially in machine learning tasks, has become increasingly recognized in the industry, the concept of a data trading market has gained more attention. However, how to assess data quality without directly acquiring the data is an unavoidable issue for the data trading market. This is because data quality directly affects various metrics of the models obtained from machine learning tasks, such as classification accuracy. The model's metrics on the test dataset are crucial quantitative indicators for evaluating the quality of the training data. Therefore, to assess data quality before purchasing data, model training and performance testing on privacy-preserving datasets are necessary.

[0003] Traditional data trading markets typically include a centralized platform where all data for trading is uploaded and submitted, and the platform then evaluates the data. However, with increasing emphasis on data privacy and the enactment of relevant privacy protection laws, directly handing over plaintext data to the platform and unconditionally trusting its protective measures has become unacceptable. Since joint learning is generally more time-consuming than plaintext machine learning with centralized models and datasets, and datasets in data trading markets are usually large, training and evaluating models on the entire dataset would result in excessive time consumption and transmission costs. Homomorphic cryptography is a rapidly developing branch of cryptography in recent years. Its homomorphic property allows operators to perform operations on ciphertext data, with the result being the encrypted ciphertext of the plaintext operation. The CKKS cryptosystem supports approximate fixed-point addition and multiplication operations in the ciphertext space, making it suitable for machine learning systems that do not require precise computation. Furthermore, the CKKS cryptosystem supports vector operations, providing optimization space for parallelization of operations. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] To address this, this invention proposes a design method for a trustworthy and fair data trading platform for machine learning data. It attempts to design a decentralized trading model where plaintext data does not need to be handed over to a trusted third party, and data buyers (model owners) can collaborate with data providers to conduct data quality assessments. Quality assessment requires training on privacy-preserving datasets; therefore, this invention employs a joint learning approach to ensure that data privacy is not leaked to the assessors during the quality assessment process. Furthermore, the protocol designed in this invention ensures that data quality assessment is only performed on a randomly sampled subset of data. Simultaneously, this invention also guarantees that the data ultimately traded is indeed the dataset reflected in the quality assessment, and that data providers cannot falsify data. Specifically, this invention addresses classification problems in machine learning by constructing two data trading models: one where the data owner fully owns the data and corresponding classification labels; and another where the data owner owns the input data, while the data buyer owns the corresponding classification labels.

[0006] Another objective of this invention is to propose a design system for a trustworthy and fair data trading platform for machine learning data.

[0007] To achieve the above objectives, this invention proposes a design method for a trustworthy and fair data trading platform for machine learning data, comprising:

[0008] Obtain the data required for the pre-trained model of the data buyer based on the machine learning task;

[0009] Based on the required data, all data owners participating in the transaction are obtained, and the pre-trained model is trained according to the preset sampling ratio dataset and the preset privacy calculation method in the data point summary values ​​of all data owners to obtain the test model.

[0010] The accuracy of the test model is tested using a preset test set, and data quality evaluation indicators are obtained based on the accuracy test results.

[0011] The optimal data owner is selected based on the data quality assessment indicators, and the optimal data owner is paid a price corresponding to the preset sampling ratio of data points through a data trading platform to obtain the preset sampling ratio of data points from the optimal data owner.

[0012] In addition, the design method for a trustworthy and fair data trading platform for machine learning data according to the above embodiments of the present invention may also have the following additional technical features:

[0013] Furthermore, in one embodiment of the present invention, the method includes:

[0014] The local model is trained using the preset sampling ratio of data points to obtain a trained local model.

[0015] The accuracy metric of the trained local model is evaluated on the test set data, and the accuracy metric is compared with the data quality evaluation metric to obtain the data validity evaluation result.

[0016] Furthermore, in one embodiment of the present invention, the pre-trained model for local model training includes a frozen layer; the step of performing an accuracy test on the test model using preset test set data, and obtaining a data quality evaluation index based on the accuracy test results, includes:

[0017] Based on the data point summary value, obtain the sampling input data with a preset sampling ratio;

[0018] Based on the case where the data owner only contains sampled input data, joint model training is performed to return the first intermediate output result obtained by inputting the first test set data into the frozen layer to the data owner. The first model accuracy index is obtained by comparing the output result obtained by joint model training according to the forward propagation process with the correct classification label.

[0019] Based on the sampled input data and corresponding classification labels contained by the data owner, the sampled input data is used to train the local model. After training, the second test set data is input into the frozen layer to obtain the second intermediate output result, which is then returned to the data owner for use in the training of the second half of the model to obtain the prediction result. The second model accuracy index is obtained based on the comparison result of the prediction result and the correct classification label.

[0020] Furthermore, in one embodiment of the present invention, the pre-trained model further includes a data owner-trainable layer and a data purchaser-trainable layer, wherein the joint model training based on the case where the data owner only contains sampled input data includes:

[0021] The sampled input data is input into the model freeze layer and the data owner trainable layer for training to obtain the third intermediate output result;

[0022] The third intermediate output result is input into the data purchaser's trainable layer for training output to obtain the correct classification label;

[0023] The loss function is calculated based on the correct classification label. The model gradient of the trainable layer of the data buyer is calculated based on the loss calculation result and the backpropagation method. The backpropagation gradient is then propagated to the trainable layer of the data owner for training. The model training is completed after multiple rounds of iteration.

[0024] Furthermore, in one embodiment of the present invention, the data point summary value is stored in a distributed storage system; the step of obtaining the preset sampling ratio data points of the optimal data owner includes:

[0025] Obtain the storage address of the data point corresponding to the data point digest in the distributed storage system;

[0026] The preset sampling ratio of data points for the actual optimal data owner is obtained based on the storage address of the data points.

[0027] To achieve the above objectives, another aspect of the present invention proposes a design system for a trustworthy and fair data trading platform for machine learning data, comprising:

[0028] The task data publishing module is used to obtain the data needed by the data buyer for the pre-trained model based on the machine learning task;

[0029] The model joint training module is used to obtain all data owners participating in the transaction based on the required data, and to train the pre-trained model to obtain the test model according to the preset sampling ratio dataset and the preset privacy calculation method in the data point summary values ​​of all data owners.

[0030] The quality assessment and comparison module is used to perform accuracy tests on the test model using preset test set data, and to obtain data quality assessment indicators based on the accuracy test results.

[0031] The data transaction acquisition module is used to select the optimal data owner based on the data quality assessment indicators, and to pay the optimal data owner a price corresponding to a preset sampling ratio of data points through the data transaction platform, so as to acquire the preset sampling ratio of data points from the optimal data owner.

[0032] The present invention provides a method and system for designing a trustworthy and fair data trading platform for machine learning data. This allows data buyers to accurately assess data quality before purchasing, while ensuring the privacy of the data provider is not compromised during the assessment process. Model owners can repeat this process with multiple different data providers to select the one with the highest data quality for subsequent transactions. The transaction data is tamper-proof upon completion of the transaction; that is, the data assessed before the transaction is indeed the data delivered by the data provider at the time of the transaction.

[0033] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0034] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0035] Figure 1 This is a flowchart illustrating a design method for a trustworthy and fair data trading platform for machine learning data according to an embodiment of the present invention;

[0036] Figure 2 This is an architecture diagram of a data quality assessment method based on joint learning according to an embodiment of the present invention;

[0037] Figure 3 This is a protocol process diagram for joint training of fully connected layers using a homomorphic encryption scheme according to an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram illustrating the operation of a distributed storage system used in the data transaction process according to an embodiment of the present invention;

[0039] Figure 5 This is a schematic diagram illustrating the results of quality assessments conducted on different quality data owners according to an embodiment of the present invention.

[0040] Figure 6 This is a schematic diagram illustrating the accuracy test results of the evaluation model obtained by a data provider with different levels of Gaussian noise on a test dataset according to an embodiment of the present invention.

[0041] Figure 7 This is a schematic diagram of the structure of a trusted and fair data trading platform for machine learning data according to an embodiment of the present invention. Detailed Implementation

[0042] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0043] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0044] The following describes, with reference to the accompanying drawings, a design method and system for a trustworthy and fair data trading platform for machine learning data, according to embodiments of the present invention.

[0045] Figure 1 This is a flowchart illustrating the design method of a trustworthy and fair data trading platform for machine learning data according to an embodiment of the present invention.

[0046] like Figure 1 As shown, the method includes, but is not limited to, the following steps:

[0047] S1, Obtain the data required by the data buyer's pre-trained model for the machine learning task;

[0048] S2, based on the required data, obtain all data owners participating in the transaction, and train the pre-trained model to obtain the test model according to the preset sampling ratio dataset and preset privacy calculation method in the data point summary values ​​of all data owners;

[0049] S3 uses preset test set data to test the accuracy of the test model, and obtains data quality evaluation indicators based on the accuracy test results.

[0050] S4. Select the optimal data owner based on the data quality assessment indicators, and use the data trading platform to pay the optimal data owner the price corresponding to the preset sampling ratio of data points in order to obtain the preset sampling ratio of data points from the optimal data owner.

[0051] The method of this invention allows the data buyer to conduct a quality assessment of the data provided by the data provider without disclosing the data owner's privacy during the assessment process. During the data transaction, it is ensured that the data provided by the data owner corresponds to the data previously assessed, is tamper-proof, and guarantees the fairness of the data transaction. The data assessor and the data owner jointly train a machine learning model using privacy-preserving computation methods. The training data used is a sampled portion of the dataset specified by the data assessor, and no plaintext information of the specific data is disclosed to the data assessor during training. Through joint training, both parties can obtain a jointly owned test model. This model is then used to perform an accuracy test on the evaluation data owned by the data assessor, and the accuracy is used as an indicator of data quality assessment. For neural network training, transfer learning and split learning are used to divide the relatively complex neural network model into three parts: a frozen part with fixed model parameters, a trainable part using data held by the data provider, and a trainable part using the model owner. Only the training process of the first trainable layer in the third part employs a homomorphic encryption scheme. The data provider offers homomorphically encrypted ciphertext of the intermediate layer feature representations output from the first two parts. The data evaluator uses a homomorphic algorithm to calculate the model output and gradients, and provides the backpropagation gradients to the data owner for further backpropagation gradient calculations. This homomorphic algorithm includes specially optimized plaintext-ciphertext matrix multiplication and masking methods to improve computational efficiency.

[0052] Furthermore, before the data quality assessment begins, participating data providers submit all encrypted data points to a distributed storage system (such as the InterPlanetary File System, IPFS) and publicly disclose the digest value of each data point as a commitment to the data. This distributed storage system guarantees the data's immutability after submission and provides availability guarantees, meaning the data can always be retrieved using the data storage address obtained after submission. When evaluating the data, the data buyer samples a small subset of data based on the data point digest values ​​to participate in joint training. Upon completion of the data transaction, the data buyer verifies whether the data provided by the provider matches the original data commitment. If they do not match, the data buyer can hold the data provider accountable and seek compensation.

[0053] Specifically, the participants included in the data trading market scenario of this invention include:

[0054] Data Buyers: Data buyers wish to purchase data in the data market. First, data buyers publicly disclose their needs in the trading market, including the required data format and content requirements. Data owners can participate in the bidding, and data buyers decide to purchase data from one or more data providers after comprehensive evaluation.

[0055] Data Owner: Data owners possess a certain amount of plaintext data with usable value. In the data trading market, data owners can view publicly available data demand information and choose to participate in bids issued by data buyers. In machine learning scenarios, the problem to be solved is typically a classification problem. For classification problems, the data that data owners can offer for sale falls into two modes: one is that the data owner completely owns the data and the corresponding classification labels; the other is that the data owner owns the input data, while the data buyer owns the corresponding classification labels. In the second mode, the correspondence between the input data and the correct classification labels can be pre-obtained using privacy set intersection techniques.

[0056] Furthermore, the functions that a data trading market should provide include: allowing data buyers to assess the quality of data from multiple data owners bidding on it before paying the purchase price, but during the quality assessment process, the data buyer should not obtain plaintext data or access data-related privacy information. In machine learning scenarios, training data quality is defined as follows: using this batch of training data to train a machine learning model, and then using the trained model to test accuracy or other quality indicators on a test dataset (owned by the data buyer), the resulting test indicator results constitute the data quality of this batch of data. The data trading market ensures the fairness of transactions. That is, when a transaction is completed and the data buyer pays the transaction amount, the data provider must provide corresponding data products, and the quality of the provided data should be consistent with the quality indicators obtained in the pre-transaction assessment stage.

[0057] To address the above scenarios and problems, this invention proposes a data transaction process. This process mainly includes the following steps:

[0058] The first stage is the submission phase: the data buyer prepares a pre-trained model M. init This is used for transfer learning. The purpose of using transfer learning is that further training of the pre-trained model can converge to better training results faster than a randomized initial model. The pre-trained model prepared by the data buyer can be a general model trained on a relatively generalized dataset (e.g., ImageNet). This pre-trained model is divided into three parts: a frozen part, a data owner-trainable part, and a data buyer-trainable part, denoted as θ. froz θ do θ mo .like Figure 2 As shown, the data buyer publishes the pre-trained model and a description of the data required for its machine learning task, as well as its offer for the purchased data, to solicit data providers.

[0059] Data owners participate in the transaction process. Owners of data relevant to machine learning tasks can decide to participate in the bidding after understanding the buyer's needs. Those who decide to participate generate a summary of each data point they own and submit all data to a distributed data system responsible for managing the data. The data point summaries are made public to the data buyer for sampling. This summary ensures the verifiability and immutability of the transaction data. After the bidding period pre-set by the data buyer ends, the data buyer receives a list of participating data owners and conducts a data quality assessment with each owner.

[0060] Further, there is the evaluation phase: the data buyer conducts a data evaluation of each data owner involved in the transaction process.

[0061] As one embodiment of the invention, for a single data owner: the data purchaser specifies a sampling point of CF% (e.g., sampling 5% of the data) based on the summary value disclosed by the data provider. The two cases of data and tag ownership are discussed separately.

[0062] If the input data and labels are held by the data owner and the data buyer, respectively, then the two parties perform joint training: for the sampled input data x, the data owner uses the frozen part of the model and the trainable part of the data owner (θ) froz θ do The intermediate layer output B is obtained, and this intermediate layer output is homomorphically encrypted before being sent to the data buyer. Using a joint training method, the data buyer further propagates this intermediate result forward through (θ). moThis process continues until the network output is obtained. If the correct classification label belongs to the data owner, they will send the corresponding label to the data buyer (the data label cannot contain valid information from the original input). Otherwise, the correct classification label already belongs to the data buyer. The data buyer calculates the loss function based on the correct classification label and calculates the model gradient of the trainable part of the data buyer's model using the backpropagation method. The backpropagation gradient is then propagated to the data owner's model using a joint training method, and the data owner continues the entire backpropagation process. The entire calculation flow is as follows: Figure 3 As shown. This protocol primarily details the algorithm for calculating the matrix multiplication of plaintext and ciphertext. After multiple iterations and model training, the data buyer and data owner jointly evaluate the model's accuracy metrics. Specifically, the data buyer will provide the test dataset through a publicly available frozen layer portion θ. froz The intermediate output is obtained and sent to the data owner. Subsequently, both parties perform a forward propagation in the same manner as described in the training steps above. The data buyer obtains the model output and compares it with the actual correct labels to obtain the model accuracy metric.

[0063] As an embodiment of the present invention, for the implementation of plaintext-ciphertext matrix multiplication in the above-described process, the present invention designs an efficient computation algorithm to leverage the advantages of parallel operation of vector computation in the CKKS cryptosystem. Specifically, let the matrix multiplication to be computed be:

[0064] [C] = [A]·B

[0065] Here, [A] represents the encryption of plaintext A, and A and B are real matrices of size M×R and R×N, respectively, with elements A = (a ij ), B = (b ij ).Pick

[0066]

[0067] The vector has only non-zero elements from the Mkth to (M+1)kth (including the left endpoint but excluding the right endpoint, element indices start from 0, the same applies below), further denote:

[0068]

[0069] The ciphertext of the calculation result C can be calculated as follows:

[0070]

[0071] Wherein, represents the Hadamard product (vector element-wise product of ciphertext and plaintext) naturally supported by the CKKS system, and Rotate([c], k) represents cyclically shifting the vector elements of ciphertext [c] to the right by k positions. The time complexity of this algorithm is O(N+R), which significantly improves computational efficiency compared to the time complexity O(MNR) of the naive algorithm by leveraging the parallel computing capabilities of CKKS.

[0072] Furthermore, if both the input data and labels belong to the data owner, the data owner trains the model locally using the sampled data points. After training, the data buyer tests the model using its own test dataset. Specifically, the data buyer inputs its test data through a fixed part θ of the model. froz The intermediate output is then obtained. The data buyer sends this intermediate output to the data owner, who uses it in the forward propagation process of the latter part of the model to obtain prediction results. The data owner then sends the prediction results back to the data buyer, who compares them with the correct labels to obtain the model accuracy metric.

[0073] Further, in the completion phase: (1) The data buyer selects one (or several) data owners with higher model metrics and pays them a predetermined price of CF% (sampling percentage). The data owners then provide the data buyer with the data points they have specified for sampling. Specifically, the data owners provide the storage addresses of the corresponding data points in the distributed storage system, and the data buyer can obtain the actual data based on these addresses. Since the file system cannot be tampered with after file submission, this method of provision ensures that the data owners cannot forge low-quality data when the transaction is completed. At the same time, the data buyer can verify the consistency between the submitted data and the sampled summary values ​​by calculating the summary values ​​of the original data obtained. After obtaining the data, the data buyer uses this part of the data for local model training. The overall model architecture used is the same as that used in the evaluation phase, including a portion of frozen layers, but all trainable layers are trained by the data buyer at this time. After training is completed, the data buyer evaluates the model accuracy metrics on the test set and confirms that the accuracy metrics are similar to those obtained in the evaluation phase to ensure the effectiveness of the data evaluation.

[0074] (2) If the data provided by the data owner passes the verification in step (1), the data buyer continues to pay the full amount to purchase the complete data, and the data transaction is completed. The data provider provides the complete data in the same way as the sampled data provided in step (1) above, that is, by submitting the storage address of the corresponding data point on the distributed storage system. After the transaction is completed, the data buyer can use the purchased data to independently train machine learning tasks, etc.

[0075] It is understandable that the distributed storage system used in the above work process needs to meet the following conditions: When data is submitted (stored) to the system, the system generates a corresponding data point storage address. Based on the storage address, the data can be read in subsequent operations. That is, the system guarantees the availability of data obtained from the data storage address. After data is submitted to the system, further modifications to the data are not allowed. That is, the system guarantees the immutability of the data after submission. The IPFS (Interplanetary File System) system proposed in existing work meets the above requirements. Operations involving this distributed storage system during transactions are as follows... Figure 4 As shown.

[0076] Based on the above process, this invention proposes a design method for a trustworthy and fair data trading platform for machine learning data. The invention will be explained in detail below with two specific implementation examples, including the aforementioned two cases: the data owner holds both training data and labels, and the data owner only holds training data without labels.

[0077] 1. The training data and labels are owned by the data owner.

[0078] In this embodiment, the data buyer wants to train a model for the image classification task CIFAR10 test set (the test set contains 10,000 images, and the training set contains 50,000 images, all 32×32 RGB color images) and purchases data from the highest quality data provider. Ten data owners participate in the transaction, each possessing 1 / 10 of the CIFAR10 training dataset, i.e., 5,000 images. However, their datasets have varying proportions of mislabeled samples, with each mislabeling proportion being 10%. This setup simulates potential data quality degradation due to mislabeling or noise during data collection, which may occur in real-world scenarios. The data owners directly use their data for local training, employing a data sampling ratio of 4%. The pre-trained model is InceptionV3, with 78 frozen layers (the number of layers here refers to the number of frozen trainable parameter tensors, approximately 40% of the total model structure). Upon completion of training, an evaluation model is obtained to assess data quality.

[0079] After 50 training cycles, the accuracy of the evaluation models trained by data providers with different mislabeling ratios on the test dataset is as follows: Figure 5As shown, it can be observed that those with a lower rate of misclassification (i.e., higher data quality) have higher quality assessment results (i.e., test accuracy). This confirms that data buyers can select high-quality data providers using the quality assessment method described in this invention. Specifically, in this example, if a data buyer decides to purchase data from the top 5 data providers during the quality assessment process, they will have precisely selected the data from the 5 providers with the highest quality (fewest mislabeled data) with 100% accuracy. After the assessment process, the data buyer continues to transact with the selected data owners and acquires all the data.

[0080] Second, all training data belong to the data owner, while the corresponding labels belong to the data purchaser.

[0081] In this embodiment, the CIFAR10 dataset is used for the experiment, with 10 data owners and the data purchaser holding the corresponding labels for their input data. The input image data from each data owner contains Gaussian noise of varying scales; this setup simulates situations where data quality deteriorates due to interference during data collection in real-world scenarios. First, the data purchaser performs privacy set intersection with each data owner to determine the correspondence between data labels and input image data. Then, the data purchaser and each data owner jointly train the model to evaluate data quality. The model used is still the InceptionV3 architecture, with 78 layers frozen, except for the last fully connected layer, which is owned by the data owner and participates in the joint training process. Upon completion of training, both parties obtain the data quality evaluation model.

[0082] After 50 training cycles, the accuracy of the evaluation model on the test dataset obtained from data providers with different levels of Gaussian noise is as follows: Figure 6 As shown, data with less noise yields higher quality assessment metrics (test accuracy), confirming that data buyers can select high-quality data providers using the quality assessment method described in this invention. After the assessment process, the data buyer continues the transaction with the selected data owner and acquires all the data.

[0083] In summary, the joint learning scheme involved in this invention employs split learning and homomorphic encryption methods. Split learning divides the structurally complex deep learning model into multiple parts, each owned by a different participant. In this invention, the model used for quality assessment is divided into three parts: a frozen layer with fixed parameters that does not participate in training; a trainable layer owned by the data owner; and a trainable layer owned by the model owner. These three parts are referred to as θ. froz θ do θ moOnly the first trainable layer of the third part is computed using homomorphic cryptography. Compared to using homomorphic cryptography to compute the entire model's forward and backward propagation, this method improves computational efficiency and saves computational and transmission costs.

[0084] In summary, this invention designs a fair and reasonable data trading market practice scheme, ensuring that data buyers can assess and confirm the quality of the data to be traded before purchasing, and that data providers cannot provide false or low-quality data. This invention can alleviate the data silo problem caused by current data privacy issues, allowing data from all parties to be rationally allocated and utilized more efficiently. The joint training scheme designed in this invention uses a small amount of sampled data for data quality assessment, which is easily extended to real-world scenarios with large amounts of data. Simultaneously, this invention uses transfer learning and designs an efficient plaintext-ciphertext matrix multiplication mechanism under homomorphic encryption to accelerate the convergence speed of model training during data quality assessment, ensuring the efficiency of data assessment. This invention focuses on data trading problems in deep learning scenarios, protecting the privacy of all parties during data assessment. Furthermore, the joint learning concept introduced by this method can be extended to other privacy-related machine learning fields (such as privacy-preserving machine learning prediction), which is beneficial for the further promotion of machine learning applications. This invention utilizes the immutability of files submitted in the IPFS file system and the mathematical property that digest values ​​are difficult to forge, ensuring that data providers cannot forge and provide low-quality data, thus establishing the credibility of data market transactions.

[0085] The trusted and fair data trading platform design method for machine learning data according to embodiments of the present invention enables data buyers to accurately assess data quality before purchasing data, while ensuring the privacy of data providers is not compromised during the assessment process. Model owners can repeat this process with multiple different data providers to select the one with the highest data quality for subsequent transaction operations. The transaction data is tamper-proof upon completion of the transaction; that is, the data assessed before the transaction is indeed the data delivered by the data provider at the time of the transaction.

[0086] To achieve the above embodiments, such as Figure 7 As shown, this embodiment also provides a trusted and fair data trading platform design system 10 for machine learning data. The system 10 includes a task data publishing module 100, a model joint training module 200, a quality assessment and comparison module 300, and a data trading acquisition module 400.

[0087] The task data publishing module 100 is used to obtain the data required by the data buyer's pre-trained model for machine learning tasks.

[0088] The model joint training module 200 is used to obtain all data owners participating in the transaction based on the required data, and to train the pre-trained model to obtain the test model according to the preset sampling ratio dataset and the preset privacy calculation method in the data point summary values ​​of all data owners.

[0089] The quality assessment and comparison module 300 is used to perform accuracy tests on the test model using preset test set data, and obtain data quality assessment indicators based on the accuracy test results.

[0090] The data transaction acquisition module 400 is used to select the optimal data owner based on data quality assessment indicators, and to use the data transaction platform to pay the optimal data owner the price corresponding to the preset sampling ratio of data points in order to acquire the preset sampling ratio of data points from the optimal data owner.

[0091] Furthermore, the aforementioned system 10 also includes an effectiveness evaluation module, used for:

[0092] The local model is trained using data points with a preset sampling ratio to obtain a trained local model.

[0093] The accuracy metrics of the trained local model are evaluated on the test set data, and the accuracy metrics are compared with data quality evaluation metrics to obtain the data validity evaluation results.

[0094] Furthermore, the aforementioned pre-trained model includes a frozen layer; the model joint training module 200 is also used for:

[0095] Obtain the sampling input data with a preset sampling ratio based on the data point summary value;

[0096] The joint model is trained based on the case where the data owner only has the sampled input data. The first intermediate output result obtained by inputting the first test set data into the frozen layer is returned to the data owner. The first model accuracy index is obtained by comparing the output result obtained by the joint model training according to the forward propagation process with the correct classification label.

[0097] Based on the sampled input data and corresponding classification labels contained by the data owner, the sampled input data is used to train the local model. After training, the second test set data is input into the frozen layer to obtain the second intermediate output result, which is then returned to the data owner for use in the training of the second half of the model to obtain the prediction result. The second model accuracy index is obtained based on the comparison between the prediction result and the correct classification label.

[0098] Furthermore, the pre-trained model also includes a data owner-trainable layer and a data purchaser-trainable layer. The aforementioned model joint training module 200 is also used for:

[0099] The sampled input data is fed into the frozen layer and the data owner's trainable layer for training to obtain the third intermediate output result;

[0100] The third intermediate output is input into the data purchaser's training layer for training output to obtain the correct classification label;

[0101] The loss function is calculated based on the correct classification label. The model gradient of the trainable layer for the data purchaser is calculated based on the loss calculation result and the backpropagation method. The backpropagation gradient is then propagated to the trainable layer for the data owner for training. The model training is completed after multiple rounds of iteration.

[0102] Furthermore, the data point summary values ​​are stored in a distributed storage system; the data transaction acquisition module 400 is also used for:

[0103] Obtain the storage address of the data point corresponding to the data point digest in the distributed storage system;

[0104] The optimal sampling ratio of data points for the actual data owner is obtained based on the storage address of the data points.

[0105] The present invention provides a trusted and fair data trading platform system for machine learning data, enabling data buyers to accurately assess data quality before purchasing, while ensuring the privacy of data providers is not compromised during the assessment process. Model owners can repeat this process with multiple different data providers to select the one with the highest data quality for subsequent transactions. The transaction data is tamper-proof upon completion of the transaction; that is, the data assessed before the transaction is indeed the data delivered by the data provider at the time of the transaction.

[0106] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0107] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for designing a trustworthy and fair data trading platform for machine learning data, characterized in that, Includes the following steps: Obtain the data required for the pre-trained model of the data buyer based on the machine learning task; Based on the required data, all data owners participating in the transaction are obtained, and the pre-trained model is trained according to the preset sampling ratio dataset and the preset privacy calculation method in the data point summary values ​​of all data owners to obtain the test model. The accuracy of the test model is tested using a preset test set, and data quality evaluation indicators are obtained based on the accuracy test results. The optimal data owner is selected based on the data quality assessment indicators, and the optimal data owner is paid a price corresponding to the preset sampling ratio of data points through the data trading platform in order to obtain the preset sampling ratio of data points from the optimal data owner. The pre-trained model used for local model training includes a frozen layer; The step of using a preset test set data to perform an accuracy test on the test model and obtaining data quality evaluation indicators based on the accuracy test results includes: Based on the data point summary value, obtain the sampling input data with a preset sampling ratio; Based on the case where the data owner only contains sampled input data, joint model training is performed to return the first intermediate output result obtained by inputting the first test set data into the frozen layer to the data owner. The first model accuracy index is obtained by comparing the output result obtained by joint model training according to the forward propagation process with the correct classification label. Based on the sampled input data and corresponding classification labels contained by the data owner, the sampled input data is used to train the local model. After training, the second test set data is input into the frozen layer to obtain the second intermediate output result, which is then returned to the data owner for use in the training of the second half of the model to obtain the prediction result. The second model accuracy index is obtained based on the comparison result of the prediction result and the correct classification label.

2. The method according to claim 1, characterized in that, After obtaining the preset sampling ratio data points of the optimal data owner, the method further includes: The local model is trained using the preset sampling ratio of data points to obtain a trained local model. The accuracy metric of the trained local model is evaluated on the test set data, and the accuracy metric is compared with the data quality evaluation metric to obtain the data validity evaluation result.

3. The method according to claim 2, characterized in that, The pre-trained model further includes a data owner-trainable layer and a data purchaser-trainable layer. The joint model training based on the sampled input data only included by the data owner includes: The sampled input data is input into the frozen layer and the data owner training layer for training to obtain the third intermediate output result; The third intermediate output result is input into the data purchaser's trainable layer for training output to obtain the correct classification label; The loss function is calculated based on the correct classification label. The model gradient of the trainable layer of the data buyer is calculated based on the loss calculation result and the backpropagation method. The backpropagation gradient is then propagated to the trainable layer of the data owner for training. The model training is completed after multiple rounds of iteration.

4. The method according to claim 1, characterized in that, The data point summary values ​​are stored in a distributed storage system; The step of obtaining the preset sampling ratio data points of the optimal data owner includes: Obtain the storage address of the data point corresponding to the data point digest in the distributed storage system; The preset sampling ratio of data points for the actual optimal data owner is obtained based on the storage address of the data points.

5. A system for designing a trustworthy and fair data trading platform for machine learning data, characterized in that, include: The task data publishing module is used to obtain the data needed by the data buyer for the pre-trained model based on the machine learning task; The model joint training module is used to obtain all data owners participating in the transaction based on the required data, and to train the pre-trained model to obtain the test model according to the preset sampling ratio dataset and the preset privacy calculation method in the data point summary values ​​of all data owners. The quality assessment and comparison module is used to perform accuracy tests on the test model using preset test set data, and to obtain data quality assessment indicators based on the accuracy test results. The data transaction acquisition module is used to select the optimal data owner based on the data quality assessment indicators, and to pay the optimal data owner a price corresponding to a preset sampling ratio of data points through the data transaction platform, so as to acquire the preset sampling ratio of data points from the optimal data owner. The pre-trained model includes a frozen layer; the model joint training module is further used for: Based on the data point summary value, obtain the sampling input data with a preset sampling ratio; Based on the case where the data owner only contains sampled input data, joint model training is performed to return the first intermediate output result obtained by inputting the first test set data into the frozen layer to the data owner. The first model accuracy index is obtained by comparing the output result obtained by joint model training according to the forward propagation process with the correct classification label. Based on the sampled input data and corresponding classification labels contained by the data owner, the sampled input data is used to train the local model. After training, the second test set data is input into the frozen layer to obtain the second intermediate output result, which is then returned to the data owner for use in the training of the second half of the model to obtain the prediction result. The second model accuracy index is obtained based on the comparison result of the prediction result and the correct classification label.

6. The system according to claim 5, characterized in that, The system also includes an effectiveness evaluation module, used for: The local model is trained using the preset sampling ratio of data points to obtain a trained local model. The accuracy metric of the trained local model is evaluated on the test set data, and the accuracy metric is compared with the data quality evaluation metric to obtain the data validity evaluation result.

7. The system according to claim 6, characterized in that, The pre-trained model further includes a data owner-trainable layer and a data purchaser-trainable layer. The model joint training module is also used for: The sampled input data is input into the frozen layer and the data owner training layer for training to obtain the third intermediate output result; The third intermediate output result is input into the data purchaser's trainable layer for training output to obtain the correct classification label; The loss function is calculated based on the correct classification label. The model gradient of the trainable layer of the data buyer is calculated based on the loss calculation result and the backpropagation method. The backpropagation gradient is then propagated to the trainable layer of the data owner for training. The model training is completed after multiple rounds of iteration.

8. The system according to claim 5, characterized in that, The data point summary value is stored in a distributed storage system; the data transaction acquisition module is further configured to: Obtain the storage address of the data point corresponding to the data point digest in the distributed storage system; The preset sampling ratio of data points for the actual optimal data owner is obtained based on the storage address of the data points.

Images