A Data Analysis Method and System Based on Federated Learning

Abstract

The present invention discloses a data analysis method and system based on federated learning. The method includes: determining a data requirement set and a training model; transmitting the data requirement set to each data node to extract training samples; training gradient data of each data node based on the training samples, aggregating and analyzing to obtain a training result, and extracting data features; transmitting the data features to each data node for feature matching, and matching to obtain approximate samples in each data node that are adapted to the data features; if the training samples of any data node are consistent with the approximate samples, performing desensitization processing on the training result. In the embodiments of the present invention, by extracting data features from the training result and matching according to the data features with the training samples, if the matching is successful, it proves that there is a risk of privacy leakage in the training result, and accordingly, desensitization is performed on the training result to avoid situations such as deducing the original data from public data, ensuring data security.

Description

Technical Field

[0001] The present invention relates to data encryption and security protection, and in particular to a data analysis method and system based on federated learning. Background Art

[0002] Federated learning refers to training using distributed node models in each distributed data node, and then updating and iterating model parameters such as gradient functions and loss functions obtained by training the distributed node models through a central node model, so as to train and study the data in each distributed data node without collecting and misappropriating off-site data. Since there is no interaction of original data in the process, data security can be improved and privacy leakage can be avoided.

[0003] Currently, federated learning mainly ensures data security by following the principle of keeping the original data in place, and encrypts the data in the distributed data nodes first, and then conducts training after removing patient information that may pose a risk of leakage. However, in medical research, especially in the research of specific diseases, there are characteristics such as a small number of patients, significant pathological features, strong correlation, and loose geographical distribution of patients. This makes the research results highly compatible with the pathological features of users, and there is a risk of derivation and traceability through the result data. Although large-scale data privacy leakage events are difficult to occur, the confidentiality principle followed in the training process fails to ensure data security. Summary of the Invention

[0004] An embodiment of the present invention discloses a data analysis method and system based on federated learning. By extracting data features from the training results and matching the data features with the training samples, if the match is successful, it proves that there is a risk of privacy leakage in the training results, and accordingly, the training results are desensitized to avoid situations such as deriving the original data from the disclosed data and ensure data security.

[0005] A first aspect of an embodiment of the present invention discloses a data analysis method based on federated learning, and the method includes:

[0006] Determine a data requirement set and a training model;

[0007] Transmit the data requirement set to each data node;

[0008] Extract training samples from the structured data stored in each data node according to the data requirement set;

[0009] Train the training samples using the training model to obtain gradient data corresponding to each data node;

[0010] Perform aggregation analysis on each gradient data to obtain a training result;

[0011] Extract the data features of the training results;

[0012] Transmit the data features to each data node for feature matching, and match the approximate samples adapted to the data features in each data node;

[0013] If the training samples of any data node are consistent with the approximate samples, desensitize the training results.

[0014] As an optional implementation manner, in the first aspect of the embodiments of the present invention, the desensitization processing of the training results includes:

[0015] Based on the training samples, divide the association intervals for each field in the training results;

[0016] Using the same deviation index, perform dislocation deviation on the fields in the strong association interval and the fields in the weak association interval to obtain a deviation result.

[0017] Among them, the deviation index is only assigned to the central node to restore the deviation result to obtain the training result.

[0018] As an optional implementation manner, in the first aspect of the embodiments of the present invention, after the desensitization processing of the training results, the method further includes:

[0019] Analyze the contribution ratio of the training data provided by each data node to the training results;

[0020] Based on the contribution ratio, perform incentive distribution on the research benefits of the current time.

[0021] As an optional implementation manner, in the first aspect of the embodiments of the present invention, the method further includes:

[0022] For the training results after desensitization processing, after the incentive distribution is completed, store the incentives;

[0023] When the incentive storage reaches the preset desensitization times, distribute the stored incentives to each data node.

[0024] As an optional implementation manner, in the first aspect of the embodiments of the present invention, the method further includes:

[0025] When a new data node is added, extract and identify the features of the structured data of the new data node;

[0026] Based on the identification features, identify the duplicate data consistent with the new data node in each data node.

[0027] The second aspect of the embodiments of the present invention discloses a data analysis system based on federated learning, and the system includes:

[0028] A building unit for determining a set of data requirements and a training model;

[0029] A transmission unit for transmitting the set of data requirements to each data node;

[0030] A sample extraction unit for extracting training samples from the structured data stored in each data node according to the set of data requirements;

[0031] A distributed training unit for training the training samples using the training model to obtain gradient data corresponding to each data node;

[0032] An aggregation analysis unit for aggregating and analyzing each gradient data to obtain a training result;

[0033] A feature extraction unit for extracting data features of the training result;

[0034] An approximate matching unit for transmitting the data features to each data node for feature matching, and matching to obtain approximate samples in each data node that are adapted to the data features;

[0035] A desensitization unit for performing desensitization processing on the training result when the training samples of any data node are consistent with the approximate samples.

[0036] As an optional implementation manner, in the second aspect of the embodiments of the present invention, the desensitization unit includes:

[0037] An interval division sub-unit for dividing an associated interval for each field in the training result based on the training sample;

[0038] A dislocation deviation sub-unit for performing dislocation deviation on the fields in the strong association interval and the fields in the weak association interval using the same deviation index to obtain a deviation result.

[0039] Wherein, the deviation index is only assigned to the central node to restore the deviation result to obtain the training result.

[0040] As an optional implementation manner, in the second aspect of the embodiments of the present invention, the system further includes:

[0041] A contribution analysis unit for analyzing the contribution ratio of the training data provided by each data node to the training result after the desensitization unit performs desensitization processing on the training result;

[0042] An incentive allocation unit for performing incentive allocation on the research benefits of the current time based on the contribution ratio.

[0043] As an alternative embodiment, in the second aspect of the embodiments of the present invention, the system further includes:

[0044] An incentive storage unit, configured to store the incentives after the incentive distribution is completed for the training results that have been desensitized.

[0045] The incentive distribution unit is further configured to, after the incentive storage reaches a preset desensitization number of times, distribute the stored incentives to each data node.

[0046] As an alternative embodiment, in the second aspect of the embodiments of the present invention, the system further includes:

[0047] The feature extraction unit is further configured to extract identification features from the structured data of the newly added data node when there is a newly added data node accessing.

[0048] The approximate matching unit is further configured to identify duplicate data that is consistent with the newly added data node in each data node based on the identification features.

[0049] The third aspect of the embodiments of the present invention discloses a data analysis system based on federated learning, including:

[0050] A memory storing executable program code;

[0051] A processor coupled to the memory;

[0052] The processor calls the executable program code stored in the memory and executes a data analysis method based on federated learning disclosed in the first aspect of the embodiments of the present invention.

[0053] The fourth aspect of the embodiments of the present invention discloses a computer-readable storage medium storing a computer program, wherein the computer program causes a computer to execute a data analysis method based on federated learning disclosed in the first aspect of the embodiments of the present invention.

[0054] The fifth aspect of the embodiments of the present invention discloses a computer program product, which, when running on a computer, causes the computer to execute some or all of the steps of any one of the methods in the first aspect.

[0055] The sixth aspect of the embodiments of the present invention discloses an application publishing platform for publishing a computer program product, wherein, when the computer program product runs on a computer, it causes the computer to execute some or all of the steps of any one of the methods in the first aspect.

[0056] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0057] In the embodiments of the present invention, data features are extracted from the training results, and the data features are matched with the training samples. If the match is successful, it proves that there is a risk of privacy leakage in the training results. Accordingly, the training results are desensitized to avoid situations such as deducing the original data from the public data, ensuring data security. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0059] Figure 1 is a schematic flowchart of a data analysis method based on federated learning disclosed in the embodiments of the present invention;

[0060] Figure 2 is a schematic structural diagram of a data analysis system based on federated learning disclosed in the embodiments of the present invention;

[0061] Figure 3 is a schematic structural diagram of another data analysis system based on federated learning disclosed in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0062] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0063] It should be noted that the terms "first", "second", "third", "fourth", etc. in the description and claims of the present invention are used to distinguish different objects, rather than to describe a specific order. The terms "including" and "having" in the embodiments of the present invention and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products, or devices.

[0064] An embodiment of the present invention discloses a data analysis method and system based on federated learning. By extracting data features from the training results and matching the data features with the training samples, if the match is successful, it proves that there is a risk of privacy leakage in the training results. Accordingly, the training results are desensitized to avoid situations such as deducing the original data from the publicly available data and ensure data security.

[0065] Embodiment 1

[0066] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of a data analysis method based on federated learning disclosed in an embodiment of the present invention. As Figure 1 shown, the data analysis method based on federated learning may include the following steps.

[0067] 101. Determine the data requirement set and the training model.

[0068] In this embodiment, different research topics have different data requirements. For example, the case data of patients seeking medical treatment is stored in medical institutions. When conducting a special disease study, such as a study only on breast cancer, the case data of breast cancer patients needs to be extracted. If the study is only on breast cancer cases with a certain symptom, further screening is required based on the case data of breast cancer patients. Therefore, the data requirement set and the training model are determined based on the research topic and the research purpose first.

[0069] 102. Transmit the data requirement set to each data node.

[0070] In this embodiment, the data nodes may be medical institutions such as local hospitals. During the federated learning process, to ensure data security, the case data of the data nodes is only distributedly trained using the training model locally, and the gradient / loss data is shared.

[0071] In this embodiment, considering that there may be a phenomenon that the same patient seeks medical treatment in multiple medical institutions successively, the data with the same primary key in the newly added data nodes is screened and identified.

[0072] As an alternative implementation, when a new data node is accessed, identify features are extracted from the structured data of the new data node; based on the identify features, duplicate data consistent with the new data node is identified in each data node. Thus, the training samples are screened in advance before the federated learning process to avoid phenomena such as duplicate samples.

[0073] 103. Extract training samples from the structured data stored in each data node according to the data requirement set.

[0074] In this embodiment, each data node extracts the required training samples according to the data requirements determined in step 101.

[0075] 104. Train the training samples using a training model to obtain gradient data corresponding to each data node.

[0076] In this embodiment, each data node performs distributed training using a training model and trains the corresponding gradient data based on the training samples.

[0077] 105. Perform aggregation analysis on each gradient data to obtain a training result.

[0078] In this embodiment, at the central node, aggregation analysis is performed on the gradient data of each data node, and the training result of the current study is summarized to implement the data analysis process of federated learning.

[0079] 106. Extract the data features of the training result.

[0080] In this embodiment, data features are also analyzed and extracted from the training result to be used as evidence data for reverse derivation of the training samples.

[0081] 107. Transmit the data features to each data node for feature matching, and match the approximate samples in each data node that are adapted to the data features.

[0082] In this embodiment, when approximate samples in the data node are obtained based on data feature matching, it indicates that there is a risk of privacy leakage in the training result. The reason for the risk may be that the pathological features are unique and significant, and specific patient information can be deduced from the publicly available research data.

[0083] Therefore, if the training samples of any data node are the same as the approximate samples, step 108 is executed to perform desensitization processing on the training result to avoid privacy leakage.

[0084] 108. Perform desensitization processing on the training result.

[0085] In this embodiment, desensitization processing is performed on the training result with the risk of leakage to avoid the risk of privacy leakage of specific cases due to data disclosure.

[0086] As an alternative implementation, based on the training samples, association intervals are divided for each field in the training result; using the same deviation index, misalignment deviation is performed on the fields in the strong association interval and the fields in the weak association interval to obtain a deviation result. The deviation index is only assigned to the central node to restore the deviation result to obtain the training result.

[0087] As another alternative implementation, a desensitization database can also be preset to perform data replacement on the training results with a risk of leakage. Alternatively, random data can be used to replace the true values, and the random replacement algorithm is only controlled by the central node.

[0088] In this embodiment, an incentive system is also established to record and issue incentives for the performance of the training samples and gradient data provided by each data node in actual research.

[0089] As an alternative implementation, after desensitizing the training results, analyze the contribution ratio of the training data provided by each data node to the training results; based on the contribution ratio, perform incentive distribution on the benefits of the current research.

[0090] As another alternative implementation, for the desensitized training results, after completing the incentive distribution, store the incentives; when the incentive storage reaches the preset desensitization times, distribute the stored incentives to each data node. In addition, in the case where the training results involve leakage, in order to avoid deducing the contribution degree of each node through the comparison of incentive values and then deducing the specific training samples, the incentives are stored and distributed after multiple storages to ensure that the incentive values do not affect data security.

[0091] In summary, by extracting data features from the training results and matching them with the training samples, if the match is successful, it proves that there is a risk of privacy leakage in the training results, and accordingly, the training results are desensitized to avoid situations such as deducing the original data from the public data and ensure data security.

[0092] Embodiment 2

[0093] Please refer to Figure 2 , Figure 2 which is a schematic structural diagram of a data analysis system based on federated learning disclosed in an embodiment of the present invention. As Figure 2 shown, the data analysis system based on federated learning may include:

[0094] A construction unit 201, configured to determine a data requirement set and a training model;

[0095] A transmission unit 202, configured to transmit the data requirement set to each data node;

[0096] A sample extraction unit 203, configured to extract training samples from the structured data stored in each data node according to the data requirement set;

[0097] A distributed training unit 204, configured to train the training samples using the training model to obtain gradient data corresponding to each data node;

[0098] An aggregation analysis unit 205 for performing aggregation analysis on each gradient data to obtain a training result;

[0099] A feature extraction unit 206 for extracting data features of the training result;

[0100] An approximate matching unit 207 for transmitting the data features to each data node for feature matching, and obtaining approximate samples in each data node that are adapted to the data features;

[0101] A desensitization unit 208 for performing desensitization processing on the training result when the training samples of any data node are consistent with the approximate samples;

[0102] A contribution analysis unit 209 for analyzing the contribution ratio of the training data provided by each data node to the training result after the desensitization unit performs desensitization processing on the training result;

[0103] An incentive allocation unit 210 for performing incentive allocation on the current research income based on the contribution ratio;

[0104] An incentive storage unit 211 for storing the incentives after the incentive allocation is completed for the training result subjected to desensitization processing;

[0105] The incentive allocation unit 210 is further configured to distribute the stored incentives to each data node after the incentive storage reaches a preset desensitization number of times;

[0106] The feature extraction unit 206 is further configured to extract identification features from the structured data of the newly added data node when a newly added data node is accessed;

[0107] The approximate matching unit 207 is further configured to identify duplicate data in each data node that is consistent with the newly added data node based on the identification features.

[0108] Among them, the desensitization unit 208 includes:

[0109] An interval division sub-unit 2081 for dividing an associated interval for each field in the training result based on the training sample;

[0110] A dislocation deviation sub-unit 2082 for performing dislocation deviation on the fields in the strong association interval and the fields in the weak association interval using the same deviation index to obtain a deviation result.

[0111] Among them, the deviation index is only allocated to the central node to restore the deviation result to obtain the training result.

[0112] As an alternative implementation, the interval division subunit 2081 divides the associated intervals for each field in the training result based on the training samples; the dislocation deviation subunit 2082 uses the same deviation index to perform dislocation deviation on the fields in the strong association interval and the fields in the weak association interval to obtain a deviation result. Among them, the deviation index is only assigned to the central node to restore the deviation result to obtain the training result.

[0113] As another alternative implementation, a desensitization database can also be preset to perform data replacement on the training results with the risk of information leakage. Or, random data can also be used to replace the true values, and the random replacement algorithm is only controlled by the central node.

[0114] In this embodiment, an incentive system is also established to record and issue incentives for the performance of the training samples and gradient data provided by each data node in actual research.

[0115] As an alternative implementation, after the training result is desensitized, the contribution analysis unit 209 analyzes the contribution ratio of the training data provided by each data node to the training result; the incentive distribution unit 210 performs incentive distribution on the income of the current research based on the contribution ratio.

[0116] As another alternative implementation, for the desensitized training result, after the incentive distribution is completed, the incentive storage unit 211 stores the incentives; when the incentive storage reaches the preset desensitization times, the incentive distribution unit 210 distributes the stored incentives to each data node. In addition, in the case where the training result involves information leakage, in order to avoid deducing the contribution degree of each node through the comparison of incentive values and then deducing the specific training samples, the incentives are stored and then distributed after multiple storages to ensure that the incentive values do not affect data security.

[0117] In summary, by extracting data features from the training result and matching them with the training samples, if the match is successful, it proves that there is a risk of privacy leakage in the training result, and accordingly, the training result is desensitized to avoid situations such as deducing the original data from the public data and ensure data security.

[0118] Embodiment III

[0119] Please refer to Figure 3 , Figure 3 which is a schematic structural diagram of another data analysis system based on federated learning disclosed in the embodiments of the present invention. As Figure 3 shown, the data analysis system based on federated learning may include:

[0120] A memory 301 storing executable program code;

[0121] A processor 302 coupled to a memory 301;

[0122] wherein, the processor 302 calls the executable program code stored in the memory 301 and executes Figure 1 a data analysis method based on federated learning.

[0123] An embodiment of the present invention discloses a computer-readable storage medium that stores a computer program, wherein the computer program causes a computer to execute Figure 1 a data analysis method based on federated learning.

[0124] An embodiment of the present invention also discloses a computer program product, wherein when the computer program product runs on a computer, it causes the computer to execute some or all of the steps of the methods in the above method embodiments.

[0125] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program. The program can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electrically-erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc memories, a magnetic disk memory, a tape memory, or any other computer-readable medium capable of carrying or storing data.

[0126] The above has introduced in detail a data analysis method and system based on federated learning disclosed in the embodiments of the present invention. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A data analysis method based on federated learning, characterized in that, The method includes: Determine a set of data requirements and a training model; Transmit the set of data requirements to each data node; Extract training samples from the structured data stored in each data node according to the set of data requirements; Train the training samples using the training model to obtain gradient data corresponding to each data node; Perform aggregation analysis on each gradient data to obtain a training result; Extract the data features of the training result; Transmit the data features to each data node for feature matching, and match to obtain approximate samples in each data node that are adapted to the data features; If the training samples of any data node are the same as the approximate samples, perform desensitization processing on the training result.

2. The data analysis method based on federated learning according to claim 1, wherein, The performing desensitization processing on the training result includes: Based on the training samples, divide the associated intervals for each field in the training result; Using the same deviation index, perform dislocation deviation on the fields in the strong association interval and the fields in the weak association interval to obtain a deviation result; Wherein, the deviation index is only assigned to the central node for restoring the deviation result to obtain the training result.

3. A data analysis method based on federated learning according to claim 1, characterized in that After performing desensitization processing on the training result, the method further includes: Analyze the contribution ratio of the training data provided by each data node to the training result; Perform incentive distribution on the current research benefit based on the contribution ratio.

4. A data analysis method based on federated learning according to any one of claims 1 or 3, characterized in that, The method further includes: For the training result subjected to desensitization processing, after completing the incentive distribution, store the incentives; When the incentive storage reaches the preset desensitization times, distribute the stored incentives to each data node.

5. A data analysis method based on federated learning according to claim 1, characterized in that, The method further includes: When a new data node is added, extract identification features from the structured data of the new data node; Based on the identification features, identify duplicate data in each data node that is the same as the new data node.

6. A data analysis system based on federated learning, characterized in that, The system includes: A construction unit for determining a set of data requirements and a training model; A transmission unit for transmitting the set of data requirements to each data node; A sample extraction unit for extracting training samples from the structured data stored in each data node according to the set of data requirements; A distributed training unit for training the training samples using the training model to obtain gradient data corresponding to each data node; An aggregation analysis unit for performing aggregation analysis on each gradient data to obtain a training result; A feature extraction unit for extracting the data features of the training result; An approximate matching unit for transmitting the data features to each data node for feature matching, and matching to obtain approximate samples in each data node that are adapted to the data features; A desensitization unit for performing desensitization processing on the training result when the training samples of any data node are the same as the approximate samples.

7. An analysis system for data based on federated learning according to claim 6, characterized in that The desensitization unit includes: An interval division sub-unit for dividing the associated intervals for each field in the training result based on the training samples; A dislocation deviation sub-unit for performing dislocation deviation on the fields in the strong association interval and the fields in the weak association interval using the same deviation index to obtain a deviation result; Among them, the deviation index is only allocated to the central node to restore the deviation result to obtain the training result.

8. The data analysis system based on federated learning according to claim 6, characterized in that, The system further includes: A contribution analysis unit, configured to analyze the contribution ratio of the training data provided by each data node to the training result after the desensitization unit performs desensitization processing on the training result; An incentive allocation unit, configured to perform incentive allocation on the research income of the current time based on the contribution ratio.

9. A data analysis system based on federated learning according to any one of claims 6 or 8, characterized in that, The system further includes: An incentive storage unit, configured to store the incentive after the incentive allocation is completed for the training result that has undergone desensitization processing; The incentive allocation unit is further configured to distribute the stored incentive to each data node after the incentive storage reaches a preset desensitization times.

10. A data analysis system based on federated learning according to claim 6, characterized in that, The system further includes: The feature extraction unit is further configured to extract identification features from the structured data of the newly added data node when there is a newly added data node accessing; The approximate matching unit is further configured to identify duplicate data that is consistent with the newly added data node in each data node based on the identification features.

Images