Data authenticity verification method and system for federal learning
By introducing a trusted party to compare data features, the problem of data authenticity judgment in federated learning is solved, ensuring that only valid data is used for model training, thereby improving the accuracy and security of data value mining and model training.
Patent Information
- Application Number
- CN202511380046.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-25
- Publication Date
- 2026-01-06
AI Technical Summary
In federal learning, the authenticity of data provided by participating parties cannot be effectively verified, which may endanger social and national security.
A first trusted party and a second trusted party are introduced to generate first and second copy data of the data participants. The second trusted party performs feature extraction and similarity calculation to determine the authenticity of the data. The first trusted party determines the actual participants and distributes valid copy data for model training.
Ensure that only valid real data is used for model training to improve the accuracy and security of data value mining and model training in federated learning.
Smart Images
Figure CN121278786A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of federated learning technology, specifically relating to a data authenticity verification method and system for federated learning. Background Technology
[0002] Federated Learning (FL) is a distributed training method for machine learning that aims to protect data privacy while enabling multiple participants to collaborate in building a common machine learning model. In traditional centralized machine learning, all data is concentrated on a central server for model training, which may raise privacy and security issues. Federated Learning addresses some of these data privacy and security concerns by distributing model training across multiple local devices or data centers, thus avoiding centralized data storage.
[0003] Unlike classical machine learning, federated learning achieves "usable but not visible" data in order to protect the privacy of the original data. However, if we only protect data privacy without paying attention to the authenticity of the data itself, it will endanger social and national security. Therefore, for model training and data value mining in the context of federated learning, it is necessary to focus on whether the data provided by the participants in federated learning is real data. Thus, in the context of federated learning, how to identify the authenticity of the data provided by the participants has become a key problem that current research urgently needs to solve. Summary of the Invention
[0004] The purpose of this invention is to provide a data authenticity verification method and system for federated learning, in order to solve the problem that existing technologies cannot determine the authenticity of data provided by federated learning participants.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] Firstly, a method for verifying the authenticity of data in federated learning is provided, including:
[0007] Each data participant generates its own first copy data and second copy data based on its own original data, and sends its first copy data to the first trusted party and its second copy data to the second trusted party.
[0008] The second trusted party extracts features from each received second copy data to obtain data feature points for each second copy data, and generates calibration feature points based on all received second copy data.
[0009] The second trusted party verifies the authenticity of each second copy data based on the calibrated feature points and the data feature points of each second copy data, obtains the authenticity verification result, and broadcasts the authenticity verification result to the first trusted party.
[0010] Based on the authenticity verification result, the first trusted party determines at least one valid copy data from all the received first copy data, and takes the data participants corresponding to each valid copy data as the actual participants;
[0011] The first trusted party distributes the corresponding valid copy data to each actual participant, so that each actual participant can use the received valid copy data as training data for model training in federated learning.
[0012] Based on the above disclosure, before performing data authenticity verification, this invention introduces two trusted parties: a first trusted party and a second trusted party. Then, each data participant generates its own first and second copy data based on its original data, sending the first copy data to the first trusted party for storage and the second copy data to the second trusted party for authenticity verification. Upon receiving the second copy data from each data participant, the second trusted party can extract features from each second copy data to obtain data feature points, and simultaneously generate calibration features based on all received second copy data. First, the first trusted party verifies the authenticity of each second copy of data based on the calibrated feature points and the data feature points of each second copy. Then, it sends the authenticity verification result to the first trusted party. The first trusted party can then use this result to identify valid copy data from its stored first copy data and designate the data participants corresponding to the valid copy data as actual participants. Finally, it distributes the corresponding valid copy data to each actual participant, enabling each participant to use their valid copy data as local training data, thereby completing the training of the federated learning model.
[0013] Through the above design, this invention introduces a trusted party and generates a copy of the original data, which is then sent to the trusted party to verify the authenticity of the data. Specifically, the trusted party uses data feature comparison to verify the authenticity of the data from the federated learning participants. This ensures that only valid and authentic data is used for subsequent federated learning model training and data value mining. Therefore, using data authenticity as a crucial basis for data value mining and model training provides a fundamental guarantee for deep learning and data element circulation in the context of federated learning. Thus, this invention is highly suitable for large-scale application and promotion.
[0014] In one possible design, the second trusted party performs authenticity verification on each second copy of the data based on the calibrated feature points and the data feature points of each second copy, obtaining authenticity verification results, including:
[0015] Based on the calibration feature points and the data feature points of each second copy data, the similarity between the data distribution features of each second copy data and the calibration distribution features is calculated, wherein the calibration distribution features are the distribution features of the calibration feature points;
[0016] For any second copy data, determine whether the similarity between the data distribution characteristics of the second copy data and the calibration distribution characteristics is greater than or equal to a similarity threshold.
[0017] If so, then any of the second copy data is taken as the real data, and the authenticity verification result is obtained after all the second copy data has been polled.
[0018] In one possible design, based on the calibrated feature points and the data feature points of each second copy data, the similarity between the data distribution characteristics of each second copy data and the calibrated distribution characteristics is calculated, including:
[0019] Using the calibration feature points and the data feature points of each second replica data, the KL divergence between the data distribution characteristics of each second replica data and the calibration distribution characteristics is calculated;
[0020] The KL divergence between the data distribution characteristics of each second replica data and the calibration distribution characteristics is used as the similarity between the data distribution characteristics of each second replica data and the calibration distribution characteristics.
[0021] In one possible design, calibration feature points are generated based on all received second copy data, including:
[0022] All second-copy data are combined in pairs to obtain several data pairs;
[0023] For any data pair, the correlation between the two second replica data in the data pair is calculated based on the data feature points of the two second replica data in the data pair, and the correlation of each data pair is obtained after all data pairs have been polled.
[0024] Based on the correlation between the data pairs, irrelevant data is identified from all data pairs;
[0025] From all the second copy data, the irrelevant data is removed to obtain the valid data;
[0026] The calibration feature points are generated using the valid data.
[0027] In one possible design, during model training in federated learning, the method further includes:
[0028] During the current training iteration of the model, the first trust side selects the same feature points from the data feature points of the valid replica data corresponding to each actual participant;
[0029] For any actual participant, the first trusted party calculates the total training reward corresponding to the effective copy data used by any actual participant in the current training iteration for model training based on the same feature points.
[0030] The first trust side determines whether the total training reward in the current training iteration is greater than the total training reward in the previous training iteration.
[0031] If so, the first trustee will obtain several valid feature points as supplementary feature points;
[0032] The first trusted party uses all feature points in the valid replica data corresponding to any actual participant and the supplementary feature points to form the actual training data for the current training iteration. After polling all actual participants in the current training iteration, it determines the actual training data of all actual participants in the current training iteration, so as to use the actual training data of each actual participant to perform local federated learning model training.
[0033] In one possible design, based on the same feature points, the total training reward corresponding to the effective copy data used by any actual participant for model training in the current training iteration is calculated, including:
[0034] Calculate the training reward for the same feature point in the effective replica data of any actual participant at the current training iteration, and the training reward for each target feature point, wherein each target feature point is a data feature point other than the same feature point among all data feature points of the effective replica data corresponding to any actual participant.
[0035] The target reward is calculated based on the training reward of each target feature point.
[0036] The training rewards for the same feature points and the target reward are summed to obtain the total training reward for each actual participant in the current training iteration.
[0037] In one possible design, the first trusted party acquires several valid feature points as supplementary feature points, including:
[0038] From all actual participants in the current training iteration, determine the target participant, wherein the target participant is any actual participant other than the aforementioned actual participant among all actual participants in the current training iteration;
[0039] Feature extraction information is sent to the target participant so that after receiving the feature extraction information, the target participant selects at least p data feature points from the data feature points of the effective replica data of the target participant in the current training iteration and sends them to the first trusted party, where p is the maximum value among all data feature points in the effective replica data of any actual participant.
[0040] The first trusted party receives at least p data feature points sent by the target participant as supplementary feature points.
[0041] Secondly, a data authenticity verification system for federated learning is provided, including:
[0042] Each data participant generates its own first copy data and second copy data based on its own original data, and sends its first copy data to the first trusted party and its second copy data to the second trusted party.
[0043] The second trusted party is used to extract features from each of the received second copy data, obtain the data feature points of each second copy data, and generate calibration feature points based on all the received second copy data.
[0044] The second trusted party is also used to perform authenticity verification on each second copy data based on the calibrated feature points and the data feature points of each second copy data, obtain the authenticity verification result, and broadcast the authenticity verification result to the first trusted party.
[0045] The first trusted party is used to determine at least one valid copy data from all the received first copy data based on the authenticity verification result, and to take the data participants corresponding to each valid copy data as the actual participants.
[0046] The first trusted party is also used to distribute corresponding valid copy data to each actual participant, so that each actual participant can use the received valid copy data as training data for model training in federated learning.
[0047] Thirdly, a data authenticity verification device for federated learning is provided. Taking an electronic device as an example, it includes a memory, a processor, and a transceiver that are sequentially and communicatively connected. The memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the data authenticity verification method for federated learning as described in the first aspect or any possible design of the first aspect.
[0048] Fourthly, a storage medium is provided, on which instructions are stored, which, when executed on a computer, perform the data authenticity verification method for federated learning as described in the first aspect or any possible design of the first aspect.
[0049] Fifthly, a computer program product containing instructions is provided that, when executed on a computer, causes the computer to perform the data authenticity verification method for federated learning as described in the first aspect or any possible design of the first aspect.
[0050] Beneficial effects:
[0051] (1) This invention introduces a trusted party and generates a copy of the original data, which is sent to the trusted party to verify the authenticity of the data. That is, the trusted party uses data feature comparison to verify the authenticity of the data of the federated learning participants. In this way, it can be ensured that only valid and real data is used for subsequent federated learning model training and data value mining. Based on this, data authenticity is used as an important basis for data value mining and model training, which can provide a basic guarantee for deep learning and data element circulation in the context of federated learning. Therefore, this invention is very suitable for large-scale application and promotion.
[0052] (2) In the federated learning process, the present invention introduces a reward mechanism, namely, firstly, the total training reward is calculated when the effective copy data used by each actual participant in the current iteration is used for model training; then, it is determined whether the total training reward in the current iteration is greater than the total reward in the previous iteration. If so, effective feature points are supplemented, and the supplemented effective feature points and the feature points of the current data are used to form the training data in the current iteration, thereby increasing the amount of effective training data. In this way, by introducing a reward mechanism, the participants can be encouraged to provide as much effective data as possible to participate in federated learning, thereby ensuring the effectiveness and accuracy of federated learning. Attached Figure Description
[0053] Figure 1 This is a flowchart illustrating the steps of a data authenticity verification method for federated learning provided in an embodiment of the present invention.
[0054] Figure 2This is a schematic diagram of the structure of a data authenticity verification system for federated learning provided in an embodiment of the present invention;
[0055] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation
[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.
[0057] It should be understood that although the terms first, second, etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit, without departing from the scope of the exemplary embodiments of the invention.
[0058] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.
[0059] Example:
[0060] See Figure 1As shown, the data authenticity verification method for federated learning provided in this embodiment introduces a trusted party and generates a copy of the original data, which is sent to the trusted party for data authenticity verification. That is, the similarity of data distribution characteristics is calculated in the trusted party to verify the data authenticity of the federated learning participants. In this way, it can be ensured that only valid data is used for subsequent federated learning model training and data value mining. Based on this, using data authenticity as an important basis for data value mining and model training can provide a basic guarantee for deep learning and data element circulation in the context of federated learning. For example, this method can be run on the data participants (including data holders and data users), the first trusted party, and the second trusted party. Optionally, the aforementioned execution subject can be, but is not limited to, a personal computer (PC), an industrial computer, or a server. It is understood that the aforementioned execution subject does not constitute a limitation on the embodiments of this application. Accordingly, the operation steps of this method can be, but are not limited to, the steps S1 to S5 below.
[0061] S1. Each data participant generates its own first and second copy data based on its original data, and sends its first copy data to the first trusted party and its second copy data to the second trusted party. In specific applications, the first and second trusted parties are trusted third-party verification participants. These third-party verification participants can be selected from, but are not limited to, the data participants ranked first and second in the electronic credit scoring system. Alternatively, additional third parties can be used as trusted parties, such as a central server. Examples of various... Data participants may include, but are not limited to, data holders and data users, and there is a corresponding relationship between data holders and data users (e.g., data holder A corresponds to data user D, etc.). The data holder performs two sampling processes on its own original data to obtain two sample data, and uses the two sample data as the first copy data and the second copy data (the first copy data generated by a data holder corresponds one-to-one with the second copy data), and sends them to the first trusted party and the second trusted party respectively. The data user then uses the data after subsequent authenticity verification to perform model training and / or data mining in federated learning.
[0062] After the data participants send their respective second copy data to the second trusted party, the second trusted party can verify the authenticity of the data. In this embodiment, the authenticity of each second copy data is verified by extracting the feature points of each second copy data and comparing the distribution features of the feature points. Optionally, the aforementioned authenticity verification process can be, but is not limited to, the steps S2 and S3 below.
[0063] S2. The second trusted party extracts features from each of the received second copy data to obtain the data feature points of each second copy data, and generates labeled feature points based on all the received second copy data. In specific implementation, deep learning methods can be used, but are not limited to, to extract the data feature points of each second copy data, such as using a convolutional neural network to extract features from each second copy data. Of course, other neural networks can also be used, and are not limited to the examples mentioned above.
[0064] Thus, after extracting the data feature points of each second copy data, all the received second copy data can be used to generate calibration feature points so that the distribution characteristics of the calibration feature points can be compared with the distribution characteristics of the data feature points of each second copy data to obtain the authenticity verification result; wherein, the process of generating calibration feature points can be, but is not limited to, as shown in steps S21 to S25 below.
[0065] S21. Perform pairwise combination processing on all the second copy data to obtain several data pairs; in this embodiment, assuming there are four second copy data, namely A1, A2, A3 and A4, then A1 and A2, A1 and A3, A1 and A4, A2 and A3, A2 and A4, and A3 and A4 are each treated as a data pair, thus obtaining 6 data pairs; of course, when the number of second copy data is different, the principle of pairwise combination is the same as the previous example, and will not be repeated here.
[0066] After obtaining several data pairs, the correlation between the data pairs can be calculated, as shown in step S22 below.
[0067] S22. For any data pair, based on the data feature points of the two second replica data in the data pair, calculate the correlation between the two second replica data in the data pair, and obtain the correlation of each data pair after polling all data pairs. In specific implementation, the KL divergence between the two second replica data in the data pair is calculated based on the data feature points of the two second replica data in the data pair. The KL divergence is then used as the correlation of the distribution characteristics of the two second replica data in the data pair (i.e., the smaller the distribution difference, the more similar the two are, and the more related the data are). Of course, KL divergence is a commonly used method to measure the difference in data distribution, and its principle will not be elaborated here.
[0068] After obtaining the correlation of each data pair, irrelevant data can be removed based on the correlation, as shown in steps S23 and S24 below.
[0069] S23. Based on the correlation of each data pair, irrelevant data is identified from all data pairs. In specific applications, data pairs with a correlation lower than the correlation threshold are selected from several data pairs. Then, the second copy data in the selected data pairs is taken as irrelevant data. After obtaining the irrelevant data, data removal can be performed, as shown in step S25 below.
[0070] S24. Remove the irrelevant data from all the second copy data to obtain valid data.
[0071] After removing irrelevant data from all the second copy data, the remaining data can be used as valid data, and based on this, calibration feature points are generated, as shown in step S25 below.
[0072] S25. Using the effective data, generate the calibration feature points; in this embodiment, the calibration feature points can be formed using the data feature points of the effective data, but are not limited to.
[0073] Therefore, after determining the calibration data points through the aforementioned steps S21 to S25, the authenticity of each second copy of the data can be verified based on these data points, as shown in step S3 below.
[0074] S3. The second trusted party performs authenticity verification on each second copy data based on the calibrated feature points and the data feature points of each second copy data, obtains the authenticity verification result, and broadcasts the authenticity verification result to the first trusted party. In specific applications, this embodiment performs authenticity verification on each second copy data based on statistical distribution characteristics. The process can be, but is not limited to, the steps S31 to S33 below.
[0075] S31. Based on the calibration feature points and the data feature points of each second copy data, calculate the similarity between the data distribution characteristics of each second copy data and the calibration distribution characteristics, wherein the calibration distribution characteristics are the distribution characteristics of the calibration feature points; in specific implementation, for example, but not limited to, first using the calibration feature points and the data feature points of each second copy data to calculate the KL divergence between the data distribution characteristics of each second copy data and the calibration distribution characteristics; then, use the KL divergence between the data distribution characteristics of each second copy data and the calibration distribution characteristics as the similarity between the data distribution characteristics of each second copy data and the calibration distribution characteristics.
[0076] After calculating the KL divergence between the distribution characteristics of the calibration feature points and the distribution characteristics of each second copy data, the authenticity of each second copy data can be verified based on this, as shown in steps S32 and S33 below.
[0077] S32. For any second copy data, determine whether the similarity between the data distribution characteristics of the second copy data and the calibration distribution characteristics is greater than or equal to a similarity threshold. In this embodiment, if the similarity is greater than the similarity threshold, it indicates that the second copy data is real data; otherwise, it is determined to be fake data. That is, the verification result of the second copy data is real data or fake data. In this way, by performing similarity comparison in the aforementioned manner, the authenticity verification of each second copy data can be completed, thereby obtaining the authenticity verification result. The process is shown in step S33 below.
[0078] S33. If so, then any of the second copy data is taken as real data, and the authenticity verification result is obtained after all the second copy data has been polled; in this embodiment, the authenticity verification result records the authenticity of each second copy data (such as whether it is real data or fake data).
[0079] After obtaining the authenticity verification results of each second copy of the data through the aforementioned steps S31 to S33, the authenticity can be broadcast so that the first trusted party can determine the actual participants in the federated learning based on the authenticity verification results. The process is shown in steps S4 and S5 below.
[0080] S4. Based on the authenticity verification result, the first trusted party determines at least one valid copy data from all received first copy data, and takes the data participants corresponding to each valid copy data as the actual participants. In specific implementation, the first trusted party can determine the valid copy data from all first copy data based on the authenticity of each second copy data in the authenticity verification result and based on the correspondence between the second copy data and the first copy data. For example, if the second copy data is real data, then the first copy data corresponding to the second copy data is valid copy data. At this time, the actual participants can be determined based on the valid copy data. Here, it is assumed that the second copy data and the first copy data are sent by data holder A. As mentioned above, there is a correspondence between data holder and data user, that is, data holder A corresponds to data user D. Therefore, the data participant corresponding to the valid copy data is data user D. Of course, the method for determining the data participants corresponding to the other valid copy data is the same as the example above, and will not be repeated here.
[0081] Based on this, after the actual participants are identified through the aforementioned method, the valid copy data can be broadcast, so that each valid data user (i.e., the actual participant) can train the federated learning model based on the received valid copy data, as shown in step S5 below.
[0082] S5. The first trusted party distributes the corresponding valid copy data to each actual participant, so that each actual participant can use the received valid copy data as training data for model training in federated learning. In this embodiment, the authenticity of the data is verified during each model training iteration, so that the data used by federated learning can be guaranteed to be valid and authentic data in each iteration.
[0083] Furthermore, for example, the second trusted party can broadcast the authenticity verification results to each data user, so that each data user knows whether it will participate in the federated learning model training. At the same time, the authenticity verification results can be recorded in the electronic credit chain of the data holder and the second trusted party, and the screening results of the first trusted party can be recorded in its corresponding electronic credit chain for data retention.
[0084] Therefore, through the data authenticity verification method for federated learning described in detail in steps S1 to S5 above, this invention introduces a trusted party and generates a copy of the original data, which is sent to the trusted party to verify the authenticity of the data. That is, the similarity of the data distribution characteristics is calculated in the trusted party to verify the authenticity of the data of the federated learning participants. In this way, it can be ensured that only valid data is used for subsequent model training and data value mining in federated learning. Based on this, using data authenticity as an important basis for data value mining and model training can provide a basic guarantee for deep learning and data element circulation in the context of federated learning.
[0085] In one possible design, the second aspect of this embodiment, based on the first aspect of this embodiment, provides a model training optimization scheme. In this second aspect, a reward mechanism is introduced during the model training process to encourage data participants to provide as much effective data as possible for model training.
[0086] Optionally, the process of enabling data participants to increase effective feature points based on a reward mechanism is shown in steps S6 to S10 below.
[0087] S6. During the current training iteration of the model, the first trusted party selects the same feature points from the data feature points of the valid replica data corresponding to each actual participant. In this embodiment, the valid replica data during the current training iteration is determined by steps S1 to S5 described in the first aspect of the embodiment. Thus, after obtaining the valid replica data of each actual participant during the current iteration, the same feature points in the valid replica data of the actual participants can be extracted. For example, if actual participant B and actual participant C provide 100 and 200 feature points respectively, then the common feature points of the two need to be selected from the 100 and 200 feature points.
[0088] After obtaining the same feature points, the reward can be calculated, as shown in step S7 below.
[0089] S7. For any actual participant, the first trusted party calculates the total training reward corresponding to the effective copy data used by any actual participant for model training in the current training iteration based on the same feature points; in specific implementation, for example, but not limited to, the following steps S71 to S73 can be used to calculate the total training reward in the current training iteration.
[0090] S71. Calculate the training reward for the same feature point in the effective replica data of any actual participant at the current training iteration, and the training reward for each target feature point, wherein each target feature point is a data feature point other than the same feature point among all data feature points of the effective replica data corresponding to any actual participant.
[0091] In practice, for any target feature point, the training reward for that target feature point can be calculated using, for example, but not limited to, the following formula.
[0092] R i = (A×F i )
[0093] In the formula, R i The training reward for any of the target feature points is A, where A is the reward coefficient and F is the training reward for each target feature point. i This indicates the number of actual participants providing any of the target features, that is, the number of actual participants containing any of the target feature points.
[0094] Furthermore, the training reward for the same feature points can be set to different reward values based on their quantity.
[0095] Thus, after calculating the training reward for the same feature point and the training reward for each target feature point, the total training reward for any actual participant in the current training iteration can be calculated based on this, as shown in steps S72 and S73 below.
[0096] S72. Calculate the target reward based on the training reward of each target feature point; in this embodiment, select the training rewards of several target feature points from the training rewards of each target feature point and sum them to obtain the target reward.
[0097] In this embodiment, it is assumed that the rewards for each target feature point are, in order: R2,...,R p-q+1 Where p is the maximum value among all data feature points in the valid copy data of any actual participant, and q is the total number of identical feature points, then R2 to R can be summed. qTo obtain the target reward.
[0098] After obtaining the target reward, the training reward for the same feature points can be combined to obtain the total training reward, as shown in step S73 below.
[0099] S73. Summing the training rewards of the same feature points and the target reward, to obtain the total training reward of any actual participant in the current training iteration.
[0100] After obtaining the total training reward of any actual participant in the current training iteration through the aforementioned steps S71 to S73, it can be compared with the total training reward in the previous iteration, as shown in step S8 below.
[0101] S8 The first trusted party determines whether the total training reward in the current training iteration is greater than the total training reward in the previous training iteration. In this embodiment, if the total training reward in the current iteration is greater than the total training reward in the previous training iteration, feature points can be supplemented to increase the amount of effective training data. The process is shown in step S9 below.
[0102] S9. If so, the first trusted party acquires several valid feature points as supplementary feature points. In specific implementation, the target participant is first determined from all actual participants in the current training iteration (in this embodiment, the target participant is any actual participant other than the aforementioned actual participant among all actual participants in the current training iteration). Then, feature extraction information is sent to the target participant so that after receiving the feature extraction information, the target participant selects at least p data feature points from the data feature points of the target participant's valid copy data in the current training iteration and sends them to the first trusted party. Finally, the first trusted party can receive at least p data feature points sent by the target participant as supplementary feature points.
[0103] Thus, after obtaining the supplementary feature points, the actual training data used in the current training iteration can be generated, as shown in step S10 below.
[0104] S10. The first trusted party uses all the feature points in the valid copy data corresponding to any actual participant and the supplementary feature points to form the actual training data for the current training iteration. After polling all actual participants in the current training iteration, the first trusted party determines the actual training data of all actual participants in the current training iteration, so as to use the actual training data of each actual participant to perform local federated learning model training.
[0105] Thus, during each iteration of model training, steps S6 to S10 are performed to supplement effective feature points until the model converges or reaches the maximum number of iterations. Based on this, data participants can be encouraged to add effective feature points during each model training iteration to improve the quality of effective training data.
[0106] Therefore, through the aforementioned design, this invention introduces a reward mechanism to encourage participants to provide as much effective data as possible for federated learning, thus ensuring the effectiveness and accuracy of federated learning.
[0107] like Figure 2 As shown, the third aspect of this embodiment provides a hardware system for implementing the data authenticity verification method for federated learning described in the first aspect of the embodiment, comprising:
[0108] Each data participant generates its own first copy data and second copy data based on its own original data, and sends its first copy data to the first trusted party and its second copy data to the second trusted party.
[0109] The second trusted party is used to extract features from each of the received second copy data, obtain the data feature points of each second copy data, and generate calibration feature points based on all the received second copy data.
[0110] The second trusted party is also used to perform authenticity verification on each second copy data based on the calibrated feature points and the data feature points of each second copy data, obtain the authenticity verification result, and broadcast the authenticity verification result to the first trusted party.
[0111] The first trusted party is used to determine at least one valid copy data from all the received first copy data based on the authenticity verification result, and to take the data participants corresponding to each valid copy data as the actual participants.
[0112] The first trusted party is also used to distribute corresponding valid copy data to each actual participant, so that each actual participant can use the received valid copy data as training data for model training in federated learning.
[0113] The working process, working details and technical effects of the system provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.
[0114] like Figure 3As shown, the fourth aspect of this embodiment provides a data authenticity verification device for federated learning. Taking an electronic device as an example, it includes: a memory, a processor, and a transceiver that are connected in sequence. The memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the data authenticity verification method for federated learning as described in the first and second aspects of the embodiments.
[0115] For specific examples, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or first-in-last-out (FILO) memory, etc.; specifically, the processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor may be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), PLA (Programmable Logic Array). The processor may also include a main processor and a coprocessor. The main processor, also known as the CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state.
[0116] In some embodiments, the processor may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. For example, the processor may not be limited to microprocessors of the STM32F105 series, reduced instruction set computer (RISC) microprocessors, x86 architecture processors, or processors with integrated neural network processing units (NPUs). The transceiver may be, but is not limited to, a Wi-Fi transceiver, a Bluetooth transceiver, a General Packet Radio Service (GPRS) transceiver, a ZigBee transceiver (a low-power LAN protocol based on the IEEE 802.15.4 standard), a 3G transceiver, a 4G transceiver, and / or a 5G transceiver. Furthermore, the device may also include, but is not limited to, a power module, a display screen, and other necessary components.
[0117] The working process, working details and technical effects of the electronic device provided in this embodiment can be found in the first and second aspects of the embodiment, and will not be repeated here.
[0118] The fifth aspect of this embodiment provides a storage medium that stores instructions for a data authenticity verification method for federated learning as described in the first and second aspects of the embodiments. That is, the storage medium stores instructions that, when executed on a computer, perform the data authenticity verification method for federated learning as described in the first and second aspects of the embodiments.
[0119] The storage medium refers to a carrier for storing data, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or Memory Sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.
[0120] The working process, working details and technical effects of the storage medium provided in this embodiment can be found in the first and second aspects of the embodiment, and will not be repeated here.
[0121] The sixth aspect of this embodiment provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform the data authenticity verification method for federated learning as described in the first and second aspects of the embodiments, wherein the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.
[0122] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A data authenticity verification method for federated learning, characterized in that, The method comprises the following steps: Each data participant generates first copy data and second copy data based on original data of the data participant, and sends the first copy data to a first trusted party and the second copy data to a second trusted party; The second trusted party extracts features from the received second copy data to obtain data feature points of the second copy data, and generates calibration feature points based on all the received second copy data; The second trusted party performs authenticity verification on the second copy data based on the calibration feature points and the data feature points of the second copy data to obtain authenticity verification results, and broadcasts the authenticity verification results to the first trusted party; The first trusted party determines at least one valid copy data from all the received first copy data based on the authenticity verification results, and regards the data participant corresponding to each valid copy data as an actual participant; The first trusted party issues the corresponding valid copy data to each actual participant, so that each actual participant uses the received valid copy data as training data for model training in federated learning.
2. The method of claim 1, wherein, The second trusted party performs authenticity verification on the second copy data based on the calibration feature points and the data feature points of the second copy data to obtain authenticity verification results, comprising: Based on the calibration feature points and the data feature points of the second copy data, the similarity between the data distribution feature of each second copy data and the calibration distribution feature is calculated, wherein the calibration distribution feature is the distribution feature of the calibration feature points; For any second copy data, it is judged whether the similarity between the data distribution feature of the second copy data and the calibration distribution feature is greater than or equal to a similarity threshold value; If yes, the second copy data is regarded as real data, and the authenticity verification results are obtained after all the second copy data are polled.
3. The method of claim 2, wherein, Based on the calibration feature points and the data feature points of the second copy data, the similarity between the data distribution feature of each second copy data and the calibration distribution feature is calculated, comprising: The KL divergence between the data distribution feature of each second copy data and the calibration distribution feature is calculated using the calibration feature points and the data feature points of the second copy data; The KL divergence between the data distribution feature of each second copy data and the calibration distribution feature is regarded as the similarity between the data distribution feature of each second copy data and the calibration distribution feature.
4. The method of claim 1, wherein, The calibration feature points are generated based on all the received second copy data, comprising: All the second copy data are combined in pairs to obtain a plurality of data pairs; For any data pair, the correlation of the two second copy data in the data pair is calculated based on the data feature points of the two second copy data in the data pair, and the correlation of each data pair is obtained after all the data pairs are polled; Based on the correlation of each data pair, unrelated data is determined from all the data pairs; The unrelated data is removed from all the second copy data to obtain valid data; Generate the calibration feature points by using the effective data.
5. The method of claim 1, wherein, In the model training of the federated learning, the method further includes: In the current training iteration, the first trusted party screens the same feature points from the data feature points of the effective copy data of each actual participant; For any actual participant, the first trusted party calculates the training total reward corresponding to the model training of the effective copy data used by the actual participant in the current training iteration based on the same feature points; The first trusted party determines whether the training total reward in the current training iteration is greater than the training total reward in the last training iteration; If yes, the first trusted party obtains a plurality of effective feature points as supplementary feature points; The first trusted party uses all the feature points in the effective copy data of the actual participant and the supplementary feature points to form the actual training data in the current training iteration, and determines the actual training data of all the actual participants in the current training iteration after all the actual participants in the current training iteration are polled, so as to use the actual training data of each actual participant to perform the model training of the federated learning locally.
6. The method of claim 5, wherein, The calculation of the training total reward corresponding to the model training of the effective copy data used by the actual participant in the current training iteration based on the same feature points includes: The first trusted party calculates the training reward of the same feature points in the effective copy data of the actual participant in the current training iteration, and the training reward of each target feature point, wherein each target feature point is a data feature point in the effective copy data of the actual participant except the same feature point; The first trusted party calculates the target reward based on the training reward of each target feature point; The first trusted party sums the training reward of the same feature points and the target reward to obtain the training total reward of the actual participant in the current training iteration.
7. The method of claim 5, wherein, The first trusted party obtains a plurality of effective feature points as supplementary feature points, including: The first trusted party determines a target participant from all the actual participants in the current training iteration, wherein the target participant is any one of the actual participants in the current training iteration except the actual participant; The first trusted party sends feature extraction information to the target participant, so that the target participant screens at least p data feature points from the data feature points of the effective copy data of the target participant in the current training iteration after receiving the feature extraction information, and sends the at least p data feature points to the first trusted party, wherein p is the maximum value of all the data feature points in the effective copy data of the actual participant; The first trusted party receives the at least p data feature points sent by the target participant as the supplementary feature points. 8.A data authenticity verification system for federated learning, characterized in that, The method includes: Each data participant generates first copy data and second copy data based on its own original data, sends the first copy data to the first trusted party, and sends the second copy data to the second trusted party; The second trusted party is configured to perform feature extraction on each of the received second copy data to obtain data feature points of each of the second copy data, and generate calibration feature points according to all the received second copy data; The second trusted party is further configured to perform authenticity verification on each of the second copy data based on the calibration feature points and the data feature points of each of the second copy data, to obtain authenticity verification results, and broadcast the authenticity verification results to the first trusted party; The first trusted party is configured to determine at least one valid copy data from all the received first copy data according to the authenticity verification results, and take a data participant corresponding to each valid copy data as an actual participant; The first trusted party is further configured to issue a corresponding valid copy data to each actual participant, so that each actual participant takes the received valid copy data as training data for model training in federated learning.
9. An electronic device, comprising: Comprise: A memory, a processor and a transceiver connected in sequence, wherein the memory is configured to store a computer program, the transceiver is configured to receive and send messages, and the processor is configured to read the computer program and execute the data authenticity verification method for federated learning according to any one of claims 1-7.
10. A computer program product comprising instructions, characterized in that, When the instructions run on the computer, the computer is caused to execute the data authenticity verification method for federated learning according to any one of claims 1-7.