Apparatus And Method For Generating Fake Data For Fair Guarantee Of Artificial Intelligence Learning Data

Abstract

The present invention relates to an apparatus and method for generating fake data to ensure fairness in artificial intelligence learning data, and more specifically, to an apparatus and method for generating fake data to ensure fairness in artificial intelligence learning data, comprising: an uploader that uploads an original data set having a plurality of columns; a deep learning module that includes a fake data generator that generates a fake data set using an inverted data set in which the original data set is inverted, and a fake data discriminator that determines the fake data set based on the original data set; and a verification module that compares and verifies the fake data set determined by the data discriminator with the original data set, and extracts it as a fair data set if fairness is improved.

Description

Technology Field

[0001] The present invention relates to an apparatus and method for generating spoofed data for the fair assurance of artificial intelligence training data. Background Technology

[0002] Artificial intelligence algorithms have been growing rapidly recently, and they can be utilized to predict risks and impacts before making decisions, thereby assisting human judgment or automating processes. AI algorithms can learn decision-making models across various fields, including public safety, policy, finance, medicine, and recruitment, and these models often learn even the biases present in society. Decisions made by algorithms, which can have a significant impact on individuals' lives, can be biased due to characteristics that necessitate their indifference, and they can disadvantage individuals belonging to specific groups based on gender, race, religion, etc.

[0003] With the rapid development of artificial intelligence (AI) technology and its application across various industries, there is a need to discuss its adverse effects and impact on society as a whole. In particular, fairness issues, such as biases regarding gender, race, and social groups, as well as a lack of transparency, are becoming significant concerns. Therefore, a key challenge in AI model development is to correct for data set unfairness that degrades model performance.

[0004] For example, in the case of an artificial intelligence model that assesses the risk of heart disease, using a dataset where the ratio of high-risk to low-risk groups is unequal (e.g., high-risk group : low-risk group = 80,000 : 20,000) can lead to incorrect results due to the unfairness of the training data. Furthermore, correcting only the ratio of one sensitive attribute (e.g., whether it is in the high-risk group) in the dataset (e.g., high-risk group : low-risk group = 20,000 : 20,000) can lead to bias in other dependent attributes (e.g., high-risk group-male : high-risk group-female = 15,000 : 5,000), so proper AI training results cannot be expected.

[0005] Furthermore, to date, methods for removing static data sets have been used to ensure fairness, while additional correction techniques for newly collected data are currently insufficient.

[0006] Therefore, there is an urgent need in this technical field for technology that can correct data bias in data sets more flexibly than conventional methods. Prior art literature

[0007] Korean Patent Publication No. 10-2039540

[0008] Son, Young-Shin, "Perception of Fairness in AI Algorithm-Based Decision Making," Seoul National University Graduate School: Department of Psychology, College of Social Sciences, Feb. 2020. The problem to be solved

[0009] The present invention aims to solve the aforementioned problems by providing an apparatus and method for generating fake data to ensure fairness in artificial intelligence training data, wherein the invention generates and identifies a fake data set using an inverted data set in which the original data set is inverted, so as to not only correct bias within the original data set but also increase the amount of data within the process data set for training an artificial intelligence model, and then extracts the fake data set as the process data set if fairness is improved after comparing and verifying the identified fake data set with the original data set.

[0010] The technical problems that the present invention aims to solve are not limited to those mentioned above, and other unmentioned technical problems can be clearly understood by those skilled in the art from the description of the present invention. means of solving the problem

[0011] To achieve the above objective, the device for generating fake data for ensuring fairness of artificial intelligence learning data according to the present invention provides: an uploader that uploads an original data set having a plurality of columns; a fake data generator that generates a fake data set using an inverted data set in which the original data set is inverted; a deep learning module comprising: a fake data discriminator that determines the fake data set based on the original data set; and a verification module that compares and verifies the fake data set determined by the data discriminator with the original data set, and extracts it as a fair data set if fairness is improved.

[0012] To achieve the above objective, the method for generating fake data for ensuring fairness of artificial intelligence learning data according to the present invention provides: an upload step in which an original data set having multiple columns is uploaded by an uploader; an inversion data generation step in which inverted data is generated by an inverted original data set by a deep learning module; a fake data generation step in which a fake data set is generated using the inverted data set by the deep learning module; a fake data determination step in which the fake data set is determined based on the original data set by the deep learning module; and a verification step in which the fake data set determined from the fake data determination step and the original data set are compared and verified by a verification module, and if fairness is improved, the fake data set is extracted as a fair data set. Effects of the invention

[0013] As described above, according to the present invention, by generating and identifying a fake data set using an inverted data set obtained by inverting the original data set, and by comparing and verifying the identified fake data set with the original data set and extracting it as a fair data set when fairness is improved, it is possible not only to correct bias within the original data set but also to significantly increase the amount of data within the fair data set for training an artificial intelligence model, thereby providing a significant effect.

[0014] The effects of the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art from the detailed description and claims. Brief explanation of the drawing

[0015] FIG. 1 is a configuration diagram of a device for generating imitation data for ensuring fairness of artificial intelligence training data according to the present invention. FIG. 2 is a diagram showing an original data set (a), an imitation data set using conventional deep learning technology (b), and a processed data set (c) output after combining the original data set and the imitation data using conventional deep learning technology. FIG. 3 is a diagram showing an original data set (a), an inverted data set (b) generated according to an embodiment of the present invention, an imitation data set (c) generated according to an embodiment of the present invention, and a processed data set (d) extracted according to an embodiment of the present invention. FIG. 4 is a flowchart of a method for generating imitation data for ensuring fairness of artificial intelligence training data according to the present invention. Specific details for implementing the invention

[0016] The terms used in this specification have been selected based on currently widely used general terms whenever possible, taking into account their functions in the present invention; however, these terms may vary depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Additionally, in specific cases, terms have been arbitrarily selected by the applicant, and in such cases, their meanings will be described in detail in the relevant description of the invention. Therefore, the terms used in this invention should be defined not merely by their names, but based on their meanings and the overall content of the invention.

[0017] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which the present invention pertains. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this application.

[0018] Hereinafter, embodiments according to the present invention will be described in detail with reference to the attached drawings. FIG. 1 is a configuration diagram of a device for generating imitation data for ensuring fairness of artificial intelligence learning data according to the present invention. FIG. 2 is a diagram showing an original data set (a), an imitation data set using conventional deep learning technology (b), and a processed data set (c) output after combining the original data set and the imitation data using conventional deep learning technology. FIG. 3 is a diagram showing an original data set (a), an inverted data set (b) generated according to an embodiment of the present invention, an imitation data set (c) generated according to an embodiment of the present invention, and a processed data set (d) extracted according to an embodiment of the present invention. FIG. 4 is a flowchart of a method for generating imitation data for ensuring fairness of artificial intelligence learning data according to the present invention.

[0020] Plunder data generation device for fair assurance of artificial intelligence training data

[0021] Referring to FIG. 1, the device for generating fake data for fair assurance of artificial intelligence learning data according to the present invention includes an uploader (100), a deep learning module (200), and a verification module (300).

[0022] More specifically, the uploader (100) uploads an original data set having multiple columns.

[0023] According to one embodiment of the present invention, the uploader (100) can obtain and upload an original data set in the CSV (Comma-Separated Values) file format from the OpenAPI of the public data portal. The original data set in the CSV file format may consist of a table, a field, a column, and column values.

[0024] The table mentioned in the present invention is a collection of data in a suitable form for quick reference, representing the structure of data in a relational database model as a two-dimensional table. That is, it is managed in the form of column values ​​and columns, and desired data can be found quickly and easily by specifying a key.

[0025] The aforementioned field refers to a piece of information that carries a specific meaning, serving as the smallest unit of processing in a database system. In relational databases, the column value is also called a record or tuple, and it refers to a single structured data item within a table. In simple terms, a database table can be said to consist of column values ​​and columns or fields. The column values ​​of each table represent a series of related data, and all column values ​​within a table have the same structure.

[0026] Next, the deep learning module (200) includes a fake data generator (210) that generates a fake data set using an inverted data set in which the original data set is inverted, and a fake data discriminator (220) that determines the fake data set based on the original data set.

[0027] The deep learning module (200) is characterized by being based on a GAN model. Generally, a GAN model is a network that artificially generates data similar to existing data by utilizing mutually adversarial discriminators and generators, and is a model belonging to unsupervised learning in machine learning. That is, the GAN model within the deep learning module (200) may include the imitation data generator (210) and the imitation data discriminator (220), or may be the same.

[0028] Meanwhile, the present invention is characterized by further including a data structure analysis module (400) that analyzes the data structure of the original data set and divides the original data set into subsets having only the unique values ​​of each column.

[0029] For example, if the structure analysis module (400) has column values ​​A, B, C, and D for any column in the original data set, it can divide the data having column value A for any column into subset A, divide the data having column value B for any column into subset B, divide the data having column value C for any column into subset C, and divide the data having column value D for any column into subset D. Accordingly, the structure analysis module (400) can divide the original data set into subset A, subset B, subset C, and subset D.

[0030] In addition, the present invention is characterized by further including a test data set generation module (500) that generates a test data set by sampling the subset and transmits the test data set to the verification module (300) so that comparative verification of the imitation data set is possible.

[0031] The above test data set generation module (500) can generate the test data set by sampling evenly from each subset, rather than sampling biasedly from one subset, in order to generate the same test data set as the data distribution of the original data set.

[0032] First, the deep learning module (200) can invert the original data set. Most preferably, the deep learning module (200) can generate the inverted data set by inverting a subset of the original data set, excluding the test data set generated by the test data set generation module (500) from the original data set. The inversion referred to in the present invention is an operation of replacing the labels of each subset.

[0033] According to one embodiment of FIG. 3, as shown in FIG. 3 (a), the remaining original data set, excluding the test data set from the original data set, may have 100 data items included in subset A and 30 data items included in subset B. Then, the deep learning module (200) can generate an inverted data set in which 100 data items are included in subset B and 30 data items are included in subset A by inverting subset A and subset B as shown in FIG. 3 (b).

[0034] In addition, the imitation data generator (210) within the deep learning module (200) can generate an imitation data set having the same columns as the inversion data set but artificially generated column values ​​by inputting the inversion data set as a training data set. Therefore, although there are physical limitations to actually obtaining all data, generating artificial data based on certain actual data can supplement the actual data and has a significant effect of increasing the amount of training for the artificial intelligence model.

[0035] In addition, the imitation data discriminator (220) within the deep learning module (200) can verify and discriminate the imitation data set generated from the imitation data generator (210) by inputting the original data set as a training data set.

[0036] Next, the verification module (300) compares and verifies the imitation data set and the original data set identified by the data discriminator (220), and if the processability is improved, extracts it as a process data set.

[0037] First, the verification module (300) can receive the test data set from the test data set generation module (500). That is, the original data set used for comparative verification from the verification module (300) most preferably refers to the test data set among the original data sets.

[0038] In other words, the verification module (300) can derive a fairness indicator value (F1) for a fairness indicator according to a specific conditional expression by adding the imitation data set identified by the data discriminator (220) to the original data set excluding the test data set. And the verification module (300) can derive a fairness indicator value (F2) for a fairness indicator according to the specific conditional expression for the test data set.

[0039] And the verification module (300) can finally extract the original data set with the added imitation data set as the process data set if, through a comparison between fairness indicator values ​​(F1, F2), the fairness of the data set with the added imitation data set is improved compared to the test data set. At this time, the degree of improvement in fairness may be a value pre-set in the verification module (300), but is not limited to a specific value.

[0041] Method for generating sham data to ensure fairness in artificial intelligence training data

[0042] Referring to FIG. 4, the method for generating fake data for fair assurance of artificial intelligence learning data according to the present invention includes an upload step (S100), an inverted data generation step (S400), a fake data generation step (S500), a fake data determination step (S600), and a verification step (S700).

[0043] More specifically, in the upload step (S100), an original data set having multiple columns is uploaded by the uploader (100).

[0044] According to one embodiment of the present invention, in the upload step (S100), an original data set in the format of a CSV (Comma-Separated Values) file may be obtained and uploaded from the OpenAPI of the public data portal. The original data set in the format of a CSV file may consist of a table, a field, a column, and column values.

[0045] The table mentioned in the present invention is a collection of data in a suitable form for quick reference, representing the structure of data in a relational database model as a two-dimensional table. That is, it is managed in the form of column values ​​and columns, and desired data can be found quickly and easily by specifying a key.

[0046] The aforementioned field refers to a piece of information that carries a specific meaning, serving as the smallest unit of processing in a database system. In relational databases, the column value is also called a record or tuple, and it refers to a single structured data item within a table. In simple terms, a database table can be said to consist of column values ​​and columns or fields. The column values ​​of each table represent a series of related data, and all column values ​​within a table have the same structure.

[0047] Meanwhile, the present invention is characterized by further including a data structure analysis step (S200) in which the data structure of the original data set is analyzed by a data structure analysis module (400), and the original data set is divided into subsets having only the unique values ​​of each column.

[0048] For example, the data structure analysis step (S200) may analyze the structure of any column within the original data set to determine whether it has column value A, column value B, column value C, and column value D.

[0049] And, in the above data structure analysis step (S200), only data having column value A for an arbitrary column is collected and divided into subset A, only data having column value B for an arbitrary column is collected and divided into subset B, only data having column value C for an arbitrary column is collected and divided into subset C, and only data having column value D for an arbitrary column is collected and divided into subset D. Accordingly, in the above data structure analysis step (S200), the original data set can be divided into subset A, subset B, subset C, and subset D.

[0050] In addition, the present invention is characterized by further including a test data set generation step (S300) in which the subset is sampled by the test data set generation module (500) to generate a test data set, and the test data set is transmitted to the verification module (300) so that comparative verification of the imitation data set is possible.

[0051] In the above test data set generation step (S300), the test data set can be generated by sampling evenly across each subset rather than by sampling biasedly from one subset, so that the test data set is generated with the same data distribution as the original data set.

[0052] Next, in the inversion data generation step (S400), the inversion data in which the original data set is inverted is generated by the deep learning module (200).

[0053] Most preferably, the inversion data generation step (S400) may generate the inversion data set by inverting a subset of the original data set, excluding the test data set generated from the test data set generation step (S300) from the original data set. The inversion referred to in the present invention is an operation of replacing the labels of each subset.

[0054] According to one embodiment of FIG. 3, as shown in FIG. 3 (a), the remaining original data set, excluding the test data set from the original data set, may have 100 data items included in subset A and 30 data items included in subset B. Then, the inverted data generation step (S400) may generate an inverted data set in which 100 data items are included in subset B and 30 data items are included in subset A by inverting subset A and subset B as shown in FIG. 3 (b).

[0055] Next, in the above-mentioned fake data generation step (S500), a fake data set is generated using the inverted data set by the fake data generator (210) within the deep learning module (200). In the above-mentioned fake data determination step (S600), the fake data set is determined based on the original data set by the fake data determiner (220) within the deep learning module (200).

[0056] Meanwhile, the above-mentioned fake data generation step (S500) and the above-mentioned fake data discrimination step (S600) are characterized by being based on a GAN model. Generally, a GAN model is a network that artificially generates data similar to existing data by utilizing mutually adversarial discriminators and generators, and is a model belonging to unsupervised learning in machine learning. That is, the GAN model within the above-mentioned deep learning module (200) may include the above-mentioned fake data generator (210) and fake data discriminator (220), or may be the same itself.

[0057] Additionally, the above-mentioned imitation data generation step (S500) can generate an imitation data set having the same columns as the inversion data set but artificially generated column values ​​by inputting the inversion data set as a training data set (S602). At this time, the above-mentioned imitation data generation step (S500) can be generated with a bias opposite to that of the original data set, that is, the same bias as the inversion data set, as shown in (c) of FIG. 3.

[0058] Therefore, there are physical limitations to actually acquiring all data, but the present invention has a significant effect of supplementing actual data and increasing the training amount of an artificial intelligence model by generating artificial data based on certain actual data.

[0059] In addition, the above-mentioned fake data determination step (S600) allows the fake data set generated from the fake data generation step (S500) to be verified and determined by inputting the original data set as a training data set.

[0060] Next, in the verification step (S700), the original data set and the imitation data set identified from the imitation data identification step (S600) are compared and verified by the verification module (300), and if the processability is improved, the process data set is extracted.

[0061] The verification step (S700) may receive the test data set generated from the test data set generation step (S300). That is, the original data set used for comparative verification from the verification step (S700) most preferably refers to the test data set among the original data sets.

[0062] In other words, the verification step (S700) can derive a fairness indicator value (F1) for a fairness indicator according to a specific conditional expression by adding the imitation data set determined by the data determination step (S600) to the original data set excluding the test data set, and can derive a fairness indicator value (F2) for a fairness indicator according to the specific conditional expression for the test data set (S701).

[0063] And in the verification step (S700), if the fairness of the data set to which the imitation data set is added to the original data set excluding the test data set is improved compared to the test data set through a comparison between fairness indicator values ​​(F1, F2), the original data set to which the imitation data set is added can be finally extracted as the fairness data set (S702). At this time, the degree of improvement in fairness may be a value pre-set in the verification module (300), but is not limited to a specific value. The extracted fairness data set is as shown in (d) of FIG. 3.

[0065] According to one embodiment of FIG. 2, conventionally, the original data set is divided into subset A and subset B as in FIG. 2 (a), a mock data set of each subset is generated using a deep learning model according to a preset ratio as in FIG. 2 (b), and after combining the original data set and the mock data set as in FIG. 2 (c), a process data set can be output.

[0066] However, as shown in Figure 2(c), the prior art generates a mock data set without correcting the bias in the original data set, combines them, and outputs a process data set, so the process data set contains bias, and accordingly, the number of data per subset in the process data set that can be practically used in the future may be significantly reduced.

[0067] However, looking again at an embodiment of FIG. 3, the present invention divides an original data set into subset A and subset B as in FIG. 3 (a), generates an inverted data set through an inversion operation that replaces labels between subsets using the original data set as in FIG. 3 (b), generates a sham data set having the same bias as the inverted data set using the inverted data set as in FIG. 3 (c), and extracts a process data set as in FIG. 3 (d). That is, the present invention has a significant effect of improving processability by removing bias within the process data set. At the same time, the number of data per subset within the process data set that can be practically used in the future may be significantly greater than in the conventional method.

[0068] Therefore, the present invention has a significant effect of not only being able to correct bias within the original data set but also significantly increasing the amount of data within the process data set for training an artificial intelligence model in the future.

[0069] The embodiments may be implemented by hardware, software, firmware, middleware, microcode, a hardware description language, or any combination thereof. Where implemented by software, firmware, middleware, or microcode, program code or code segments that perform the necessary tasks may be stored on a computer-readable storage medium and executed by one or more processors.

[0070] Furthermore, aspects of the subject matter described herein may be described in the general context of computer-executable instructions, such as program modules or components executed by a computer. Generally, program modules or components include routines, programs, objects, and data structures that perform specific tasks or implement specific data types. The aspects of the subject matter described herein may be implemented in distributed computing environments where tasks are performed by remote processing devices linked through a communication network. In a distributed computing environment, program modules may be located on both local and remote computer storage media, including memory storage devices.

[0071] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results may be achieved even if the described techniques are performed in a different order than described, and / or the components of the system, structure, device, circuit, etc. described are combined or assembled in a form different from the described method, or are replaced or substituted by other components or equivalents.

[0072] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below. Explanation of the symbols

[0073] 100.. Uploader 200.. Deep learning module 210.. Pseudo-data generator 220.. Counterfeit Data Detector 300.. Verification module 400.. Data Structure Analysis Module 500.. Test data set creation module

Claims

Claim 1 A deep learning module comprising: an uploader that uploads an original data set having multiple columns; a data structure analysis module that analyzes the data structure of the original data set and divides the original data set into subsets having only the unique values ​​of each column; a test data set generation module that generates a test data set by sampling the subsets and transmits the test data set to a verification module; a fake data generator that generates a fake data set using an inverted data set in which the original data set is inverted; and a fake data discriminator that determines the fake data set based on the original data set. A device for generating imitation data for ensuring fairness in artificial intelligence training data, comprising: a verification module that derives a fairness indicator value (F1) according to a specific conditional expression by adding the imitation data set to the original data set excluding the test data set, derives a fairness indicator value (F2) according to the specific conditional expression for the test data set, and extracts the data set as a fair data set if the degree of improvement in fairness of the original data set with the test data set excluded and the imitation data set added is a preset value compared to the test data set through a comparison between the fairness indicator values ​​(F1, F2). Claim 2 A device for generating sham data for fair assurance of artificial intelligence training data, characterized in that, in claim 1, the deep learning module is based on a GAN model. Claim 3 An upload step in which an original data set having multiple columns is uploaded by an uploader; a data structure analysis step in which the data structure of the original data set is analyzed by a data structure analysis module and the original data set is divided into subsets having only the unique values ​​of each column; a test data set generation step in which the subsets are sampled to generate a test data set by a test data set generation module and the test data set is transmitted to a verification module; an inversion data generation step in which inverted data is generated by a deep learning module in which inverted data is generated by inverting the original data set; a fake data generation step in which a fake data set is generated using the inverted data set by the deep learning module; and a fake data determination step in which the fake data set is determined based on the original data set by the deep learning module. A method for generating imitation data for ensuring fairness in artificial intelligence training data, comprising: a verification step in which, by means of a verification module, the imitation data set is added to the original data set excluding the test data set, and a fairness indicator value (F1) is derived according to a specific conditional expression, and a fairness indicator value (F2) is derived for the test data set according to the specific conditional expression, and through a comparison between the fairness indicator values ​​(F1, F2), if the degree of improvement in fairness of the original data set with the imitation data set added excluding the test data set is a preset value compared to the test data set, the corresponding data set is extracted as a fair data set. Claim 4 delete Claim 5 delete

Images