Internal security detection method for federated learning and privacy computing

Abstract

The invention provides an internal security detection method for federated learning and privacy calculation, and relates to the field of data security. The method comprises the following steps: acquiring a multi-source to-be-detected data packet corresponding to a target scene, performing data connection on the multi-source to-be-detected data packet to generate a to-be-detected data sample, and acquiring a data feature and a user feature corresponding to the to-be-detected data sample; setting a target data training model based on federal learning, obtaining a target data model corresponding to the to-be-detected data sample according to the data features and the user features, further obtaining a target data packet, and performing data mining on the target data packet to obtain security target data; setting a data encryption training mechanism, performing encryption training on the security target data, obtaining encrypted target data, establishing an encrypted target data model according to the encrypted target data, and further obtaining an abnormal behavior; according to the invention, the internal safety is improved.

Description

Technical Field

[0001] The present invention relates to the field of data security, and specifically to an internal security detection method for federated learning and privacy computing. Background Art

[0002] Federated learning refers to a machine learning framework that can effectively help multiple institutions use data and conduct machine learning modeling while meeting the requirements of user privacy protection, data security, and government regulations; Privacy computing refers to a collection of technologies that implement data analysis and computing while protecting the data from external disclosure, so as to achieve the goal of making data "available but invisible"; and to achieve the transformation and release of data value while fully protecting data and privacy security. In the existing technology, even the centralized integration of data between different departments of the same company faces many obstacles. Due to the constraints of data privacy and security such as laws and regulations, policy supervision, commercial secrets, and personal privacy, multiple data sources cannot directly exchange data, forming a "data island" phenomenon, which restricts the further improvement of the capabilities of artificial intelligence models. Therefore, in reality, it is almost impossible to integrate data scattered across various places and institutions, which will cause the following problems in the internal security testing process: First, if the internal data cannot be fully integrated, it is impossible to perform all security tests on the internal data; second, if there is external intrusion data and malicious behavior during the security testing process, the internal data will be leaked; therefore, in order to solve the above problems, an internal security testing method for federated learning and privacy computing is provided. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides an internal security detection method for federated learning and privacy computing; The purpose of the present invention can be achieved by the following technical solution: an internal security detection method for federated learning and privacy computing, the method comprising the following steps: Step 1: Obtain multi-source data packets to be detected corresponding to the target scene, perform data connection on the multi-source data packets to be detected, generate data samples to be detected, and obtain data features and user features corresponding to the data samples to be detected; Step 2: Set up a target data training model based on federated learning, obtain the target data model corresponding to the data sample to be detected according to data characteristics and user characteristics, and then obtain the target data packet, and perform data mining on the target data packet to obtain security target data; Step 3: Set up a data encryption training mechanism, perform encryption training on the security target data, obtain the encrypted target data, establish an encrypted target data model based on the encrypted target data, and then obtain abnormal behavior.

[0004] Furthermore, the process of generating the data sample to be detected includes: Obtain all target departments corresponding to the target location, and number the target departments without duplication, denoted as i, where i is a natural number greater than 2; obtain the data packets to be detected corresponding to the target departments in sequence according to the numbering sequence; and align the data packets to be detected according to the numbering sequence, thereby obtaining multi-source data packets to be detected; The multi-source to-be-detected data packets are encoded according to the target location, which is recorded as location D, where D is a positive integer; The number of target locations is obtained according to the code, and then the multi-source to-be-detected data packets are processed according to the number of target locations, wherein the target locations include 1 or n; wherein n is a natural number greater than 1; If there is one target location, a sub-data chain is set up to connect the data packets to be detected from multiple sources in sequence through the sub-data chain according to the numbering order to generate multiple data samples to be detected; If there are n target locations, a data link is set up to connect the multiple data samples to be detected corresponding to the n multi-source data packets to be detected through the data link to generate a single data sample to be detected; The data samples to be detected include multiple data samples to be detected and single data samples to be detected.

[0005] Furthermore, the process of acquiring the data features and user features includes: Obtain data features and user features corresponding to the data packets to be detected in the target department corresponding to the target location; Mark the acquired data features and user features, and mark the data features as places ; Mark user features as places ;in, It is represented as the data features corresponding to the data packet to be detected corresponding to the target department i in the target location D; It is represented as the user characteristics corresponding to the data packet to be detected corresponding to the target department i in the target location D; Furthermore, the target data model generation process includes: A sample training unit and a data mining unit are set in the target data training model, wherein the sample training unit is used to perform data training on the data sample to be detected; and the data mining unit is used to perform data mining on the target data; The sample training unit trains two adjacent data packets to be detected connected by a sub-data link in the multiple data samples to be detected, and obtains a data feature set and a corresponding user feature set corresponding to the two adjacent data packets to be detected; and marks them as and ; Wherein, Q and P represent the numbers corresponding to the Q data features and the numbers corresponding to the P user features contained in the data packet to be detected; The target data training model set based on federated learning will aggregate the data packets to be detected from different sources according to the data characteristics and user characteristics, and then perform federated learning based on the size of the overlapping parts of the data characteristics and user characteristics; if the data characteristics The overlap is large and the user characteristics If the overlap is small, horizontal federated learning is performed to increase the total amount of data samples to be tested. The overlap is small and the user characteristics If the overlap is large, vertical federated learning is performed to increase the dimension of the data sample to be tested. and user characteristics If the overlaps are small, transfer federated learning is performed; According to the data feature set and the user feature set, two adjacent data packets to be detected are trained to generate target training data packets; similarly, Continue data training with target training data to generate The training is stopped when one target training data packet is generated, and the corresponding target data model is generated.

[0006] Furthermore, the process of acquiring the target data packet includes: The single data sample to be detected is sent to the target data model to generate n target training data packets. Similarly, the n target training data packets are continuously trained until one target training data packet is generated, then the training is stopped and marked as the target data packet.

[0007] Furthermore, the process of acquiring the security target data includes: Acquire all target data in the target data packet; set a standard parameter conversion criterion in the data mining unit based on federated learning, send the target data packet to the data mining unit, and perform parameter conversion on the target data in the target data packet according to the standard parameter conversion criterion to generate target parameters, and then determine whether the target parameters meet the standard parameter conversion criterion; If the target parameter meets the standard parameter conversion criteria, the corresponding target data is marked as safe target data; Otherwise, the corresponding target data is marked as abnormal target data; and an abnormal behavior recycling bin is set in the data mining unit, and the abnormal target data is sent to the abnormal behavior recycling bin for storage.

[0008] Furthermore, the process of obtaining the encrypted target data includes: The data encryption training mechanism is used to set a number of noise nodes in the security target data, and the noise nodes are used to add noise to the security target data; Obtaining privacy target data corresponding to the security target data, adding noise nodes to the privacy target data, and then generating encrypted target data; Generate a corresponding noise digital signature for the noise node corresponding to the encrypted target data, obtain the private key corresponding to the noise digital signature, and then connect the noise digital signature and the corresponding private key to generate a noise node database; Setting up a data encryption chain to connect encrypted target data in pairs and generate an encrypted target data model; An interaction node is set in the data encryption chain to interact with the encrypted target data and obtain the interaction account corresponding to the interactor who interacts with the data.

[0009] Furthermore, the abnormal behavior acquisition process includes: Establish a data trust model, wirelessly connect and encrypt the target data model, and pre-store the authorized account of the interactor in the data trust model to detect the trustworthiness of the interactor; The interactor enters the encrypted target data model through the interaction node to interact with data. The interaction node obtains the interaction account corresponding to the interactor and sends the interaction account to the data trust model for interaction account detection to determine whether it is an authorized account. If the interaction account successfully matches the pre-stored authorization account of the interactor, it is marked as the authorized account, and the corresponding interaction behavior of the interactor data is marked as the authorized behavior; Otherwise, abnormal behavior is generated; The interactor enters the encrypted target model to interact with the data, obtains the encrypted target data, and matches the private key with several noise digital signatures corresponding to the encrypted target data according to the noise node database; If the match is successful, it is marked as an authorized behavior, and then the noise digital signature is decrypted to obtain the privacy target data corresponding to the noise digital signature; Otherwise, abnormal behavior is generated; Perform trust detection on the authorized accounts of interactors in the data trust model based on the authorized behaviors and abnormal behaviors to obtain trust scores; Set the authorized behavior to add one point and the abnormal behavior to subtract one point. If the trust score of the interactor account is less than three points, the interactor account will be removed from the data trust model, and will no longer be able to interact with the encrypted target data in the encrypted target model. It is necessary to pre-store the interaction authorization account in the data trust model again.

[0010] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention obtains data samples to be detected and performs classified detection on the samples to be detected; splits the data packets to be detected obtained from a single target location to generate multiple data samples to be detected, and more accurately performs security detection on the data to be detected in the data packets to be detected in the single target location; connects and merges the multi-source data packets to be detected obtained from multiple target locations to generate a single data sample to be detected, and more conveniently performs security detection on the multi-source data packets to be detected obtained from multiple target locations, thereby reducing detection time and improving detection efficiency; 2. Set up a target data training model based on federated learning, obtain the target data model corresponding to the data sample to be detected according to data features and user features, and then obtain the target data packet, and perform data mining on the target data packet to obtain safe target data. It cannot be determined that all target data packets in the target location are safe target data. Therefore, based on federated learning, convert them into standard data parameters, and then determine whether the obtained target data packet is abnormal target data. The abnormal target data is eliminated, which further ensures the security of the data; BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0012] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0013] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0014] like Figure 1 As shown, an internal security detection method for federated learning and privacy computing includes the following steps: Step 1: Obtain multi-source to-be-detected data packets corresponding to the target scene, perform data connection on the multi-source to-be-detected data packets, generate to-be-detected data samples, and obtain data features and user features corresponding to the to-be-detected data samples; Step 2: Set up a target data training model based on federated learning, obtain the target data model corresponding to the data sample to be tested according to the data characteristics and user characteristics, and then obtain the target data, and perform data mining on the target data to obtain safe target data; Step 3: Set up a data encryption training mechanism, perform encryption training on the security target data, obtain the encrypted target data, establish an encrypted target data model based on the encrypted target data, and then obtain abnormal behavior; In step 1, the process of generating the data sample to be detected includes: The target place is used to indicate the place where internal safety inspection is required, including but not limited to hospitals, schools, and enterprises; Obtain all target departments corresponding to the target location, and number the target departments without duplication, denoted as i, where i is a natural number greater than 2; obtain the data packets to be detected corresponding to the corresponding target departments in sequence according to the numbering sequence; and align the data packets to be detected according to the numbering sequence, thereby obtaining multi-source data packets to be detected; It should be further explained that, in a specific embodiment, the target department is included in the target place. For example, if the target place is a school, the target department includes but is not limited to a college, a class, and an office, etc.; The multi-source to-be-detected data packets are encoded according to the target location, which is recorded as location D, where D is a positive integer; The number of target locations is obtained according to the code, and then the multi-source to-be-detected data packets are processed according to the number of target locations, wherein the target locations include 1 or n; wherein n is a natural number greater than 1; If there is one target location, a sub-data chain is set up to connect the data packets to be detected from multiple sources in sequence through the sub-data chain according to the numbering order to generate multiple data samples to be detected; If there are n target locations, a data link is set up to connect the multiple data samples to be detected corresponding to the n multi-source data packets to be detected through the data link to generate a single data sample to be detected; The data samples to be detected include multiple data samples to be detected and single data samples to be detected; In one embodiment, a data packet to be detected obtained from a single target location is split to generate multiple data samples to be detected, so that the data to be detected in the data packet to be detected in the single target location can be more accurately security detected; multi-source data packets to be detected obtained from multiple target locations are data-connected and merged to generate a single data sample to be detected, so that the multi-source data packets to be detected obtained from multiple target locations can be more conveniently security detected, thereby reducing detection time and improving detection efficiency; It should be noted that in a specific embodiment, the data samples to be detected have different data, and the feature data, annotation data and user characteristics contained therein are different. It is impossible to directly merge all the data samples to be detected. In addition, due to the constraints of laws, regulations, policy supervision, commercial secrets, personal privacy and other data privacy and security, multiple data sources cannot directly exchange data. Federated learning includes horizontal federated learning, vertical federated learning and migration federated learning. Based on federated learning, multi-source data packets to be detected are integrated and exchanged, thereby completing the security detection of the data samples to be detected. The process of acquiring the data features and user features corresponding to the data sample to be detected includes: Mark the acquired data features and user features, and mark the data features as places ; Mark user features as places ;in, It is represented as the data features corresponding to the data packet to be detected corresponding to the target department i in the target location D; It is represented as the user characteristics corresponding to the data packet to be detected corresponding to the target department i in the target location D; In step 2: the process of generating the target data model includes: A sample training unit and a data mining unit are set in the target data training model, wherein the sample training unit is used to perform data training on the data sample to be detected; and the data mining unit is used to perform data mining on the target data; The sample training unit trains two adjacent data packets to be detected connected by a sub-data link in the plurality of data samples to be detected, obtains a data feature set and a corresponding user feature set corresponding to the two adjacent data packets to be detected, and marks them as and ; Wherein, Q and P represent the numbers corresponding to the Q data features and the numbers corresponding to the P user features contained in the data packet to be detected; According to the data feature set and the user feature set, two adjacent data packets to be detected are trained to generate target training data packets; similarly, The target training data is trained again to generate The training is stopped when one target training data packet is generated and recorded as the target data packet, and then the corresponding target data model is generated. In one embodiment, it is necessary to further explain that the target data training model set based on the federated learning will aggregate the data packets to be detected from different sources according to the data features and user features, and then perform federated learning according to the size of the overlapping part of the data features and user features; if the data features The overlap is large and the user characteristics If the overlap is small, horizontal federated learning is performed to increase the total amount of data samples to be tested. The overlap is small and the user characteristics If the overlap is large, vertical federated learning is performed to increase the dimension of the data sample to be tested. and user characteristics If the overlaps are small, transfer federated learning is performed; However, it should be further explained that the federated learning is to directly generate a data model from the data samples to be detected, but in the present invention, the target data model is generated from the data samples to be detected based on the federated learning, which improves the accuracy and quality while increasing the total amount and dimension of the data samples to be detected; The target data acquisition process includes: The single data sample to be detected is sent to the target data model to generate n target data models. Similarly, the n target data models are trained again until one target training data packet is generated, then the training is stopped and recorded as the target data packet; The process of acquiring the abnormal target data includes: Acquire all target data in the target data packet; set a standard parameter conversion criterion in the data mining unit based on federated learning, send the target data packet to the data mining unit, and perform parameter conversion on the target data in the target data packet according to the standard parameter conversion criterion to generate target parameters, and then determine whether the target parameters meet the standard parameter conversion criterion; If the target parameter meets the standard parameter conversion criteria, the corresponding target data is marked as safe target data; Otherwise, the corresponding target data is marked as abnormal target data; and an abnormal behavior recycling bin is set in the data mining unit, and the abnormal target data is sent to the abnormal behavior recycling bin for storage; It should be further explained that, in one embodiment, it cannot be determined that all target data of a target location are safe target data. Therefore, based on federated learning, it is converted into standard data parameters, and then it is determined whether the target data obtained is abnormal target data, and the abnormal target data is eliminated, thereby further ensuring the security of the data. In step three, the process of obtaining encrypted target data includes: The data encryption training mechanism is used to set a number of noise nodes in the security target data, and the noise nodes are used to add noise to the security target data; Obtaining privacy target data corresponding to the security target data, adding noise nodes to the privacy target data, and then generating encrypted target data; Generate a corresponding noise digital signature for the noise node corresponding to the encrypted target data, obtain the private key corresponding to the noise digital signature, and then connect the noise digital signature and the corresponding private key to generate a noise node database; In one embodiment, the privacy target data includes but is not limited to ID cards, passwords, account numbers, important numbers, and important methods, etc.; Setting up a data encryption chain to connect encrypted target data in pairs and generate an encrypted target data model; An interactive node is set up in the data encryption chain to interact with the encrypted target data and obtain the interactive account corresponding to the interactor who interacts with the data; It should be further explained that, in one embodiment, during the data interaction process, privacy computing does not move the data itself. Through parameter exchange, the data is signed based on privacy computing to obtain the corresponding private key, and a corresponding database is generated. In the data interaction, the privacy of the data can be protected and the data can be better interacted; The abnormal behavior acquisition process includes: Establish a data trust model, wirelessly connect and encrypt the target data model, and pre-store the authorized account of the interactor in the data trust model to detect the trustworthiness of the interactor; The interactor enters the encrypted target data model through the interaction node to interact with data. The interaction node obtains the interaction account corresponding to the interactor and sends the interaction account to the data trust model for interaction account matching to determine whether it is an authorized account. If the interaction account successfully matches the pre-stored authorization account of the interactor, it is marked as the authorized account, and the corresponding interaction behavior of the interactor data is marked as the authorized behavior; Otherwise, abnormal behavior is generated; In one embodiment, the interaction data behavior of the interactor corresponding to the unsuccessful matching of the interaction account with the interactor's pre-stored authorization account is marked as abnormal behavior, and the corresponding interactor is marked as a rejected interactor; if the rejected interactor needs to enter the encrypted target data model for data interaction, the interaction account needs to be pre-stored in the data trust model first; The interactor enters the encrypted target model to interact with the data, obtains the encrypted target data, and matches the private key with several noise digital signatures corresponding to the encrypted target data according to the noise node database; If the match is successful, it is marked as an authorized behavior, and then the noise digital signature is decrypted to obtain the privacy target data corresponding to the noise digital signature; Otherwise, abnormal behavior is generated; Perform trust detection on the authorized accounts of interactors in the data trust model based on the authorized behaviors and abnormal behaviors to obtain trust scores; Set the authorized behavior to add one point and the abnormal behavior to subtract one point. If the trust score of the interactor account is less than three points, the interactor account will be removed from the data trust model, and will no longer be able to interact with the encrypted target data in the encrypted target model. It is necessary to pre-store the interaction authorization account in the data trust model again.

[0015] The features and exemplary embodiments of various aspects of the present application are described in detail above. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail in combination with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application rather than to limit the present application. For those skilled in the art, the present application can be implemented without some of these specific details. The above description of the embodiments is merely to provide a better understanding of the present application by showing examples of the present application.

[0016] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An internal security detection method for federated learning and privacy computing, characterized in that: The method comprises the following steps: Step 1: Obtain multi-source to-be-detected data packets corresponding to the target scene, perform data connection on the multi-source to-be-detected data packets, generate to-be-detected data samples, and obtain data features and user features corresponding to the to-be-detected data samples; Step 2: Set up a target data training model based on federated learning, obtain the target data model corresponding to the data sample to be detected according to the data characteristics and user characteristics, and then obtain the target data packet, and perform data mining on the target data packet to obtain the security target data; Step 3: Set up a data encryption training mechanism, perform encryption training on the security target data, obtain the encrypted target data, establish an encrypted target data model based on the encrypted target data, and then obtain abnormal behavior.

2. The internal security detection method for federated learning and privacy computing according to claim 1, characterized in that: The process of generating the data sample to be detected includes: Obtain all target departments corresponding to the target location, and number the target departments without duplication, denoted as i, where i is a natural number greater than 2; obtain the data packets to be detected corresponding to the target departments in sequence according to the numbering sequence; and align the data packets to be detected according to the numbering sequence, thereby obtaining multi-source data packets to be detected; The multi-source to-be-detected data packets are encoded according to the target location, which is recorded as location D, where D is a positive integer; Obtaining the number of target locations according to the code, and then processing the multi-source to-be-detected data packets according to the number of target locations, wherein the target locations include 1 or n; wherein n is a natural number greater than 1; If there is one target location, a sub-data chain is set up to connect the data packets to be detected from multiple sources in sequence through the sub-data chain according to the numbering order to generate multiple data samples to be detected; If there are n target locations, a data link is set up to connect the multiple data samples to be detected corresponding to the n multi-source data packets to be detected through the data link to generate a single data sample to be detected; The data samples to be detected include multiple data samples to be detected and single data samples to be detected.

3. The internal security detection method for federated learning and privacy computing according to claim 2 is characterized in that: The process of acquiring the data features and user features includes: The data features and user features corresponding to the data packets to be detected in the target department corresponding to the target location are obtained; and the data features and user features are marked.

4. The internal security detection method for federated learning and privacy computing according to claim 3 is characterized in that: The generation process of the target data model includes: A sample training unit and a data mining unit are set in the target data training model, wherein the sample training unit is used to perform data training on the data sample to be detected; and the data mining unit is used to perform data mining on the target data; The sample training unit trains two adjacent data packets to be detected connected by a sub-data link in the plurality of data samples to be detected, and obtains a data feature set and a corresponding user feature set corresponding to the two adjacent data packets to be detected; According to the data feature set and the user feature set, two adjacent data packets to be detected are trained to generate target training data packets; similarly, Continue data training with target training data to generate The training is stopped when one target training data packet is generated, and the corresponding target data model is generated.

5. The internal security detection method for federated learning and privacy computing according to claim 4 is characterized in that: The process of acquiring the target data packet includes: The single data sample to be detected is sent to the target data model to generate n target training data packets. Similarly, the n target training data packets are continuously trained until one target training data packet is generated, then the training is stopped and marked as the target data packet.

6. The internal security detection method for federated learning and privacy computing according to claim 5, characterized in that: The process of obtaining the security target data includes: Acquire all target data in the target data packet; set a standard parameter conversion criterion in the data mining unit based on federated learning, send the target data packet to the data mining unit, and perform parameter conversion on the target data in the target data packet according to the standard parameter conversion criterion to generate target parameters, and then determine whether the target parameters meet the standard parameter conversion criterion; If the target parameter meets the standard parameter conversion criteria, the corresponding target data is marked as safe target data; Otherwise, the corresponding target data is marked as abnormal target data; and an abnormal behavior recycling bin is set in the data mining unit, and the abnormal target data is sent to the abnormal behavior recycling bin for storage.

7. The internal security detection method for federated learning and privacy computing according to claim 6, characterized in that: The process of obtaining the encrypted target data includes: The data encryption training mechanism is used to set a number of noise nodes in the security target data, and the noise nodes are used to add noise to the security target data; Obtaining privacy target data corresponding to the security target data, adding noise nodes to the privacy target data, and then generating encrypted target data; Generate a corresponding noise digital signature for the noise node corresponding to the encrypted target data, obtain the private key corresponding to the noise digital signature, and then connect the noise digital signature and the corresponding private key to generate a noise node database; Setting up a data encryption chain to connect encrypted target data in pairs and generate an encrypted target data model; An interaction node is set in the data encryption chain to interact with the encrypted target data and obtain the interaction account corresponding to the interactor who interacts with the data.

8. The internal security detection method for federated learning and privacy computing according to claim 7, characterized in that: The process of obtaining the abnormal behavior includes: Establish a data trust model, wirelessly connect and encrypt the target data model, and pre-store the authorized account of the interactor in the data trust model to detect the trustworthiness of the interactor; The interactor enters the encrypted target data model through the interaction node to interact with data. The interaction node obtains the interaction account corresponding to the interactor and sends the interaction account to the data trust model for interaction account matching to determine whether it is an authorized account. If the interaction account successfully matches the pre-stored authorization account of the interactor, it is marked as the authorized account, and the corresponding interaction behavior of the interactor data is marked as the authorized behavior; Otherwise, abnormal behavior is generated; The interactor enters the encrypted target model to interact with the data, obtains the encrypted target data, and matches the private key with several noise digital signatures corresponding to the encrypted target data according to the noise node database; If the match is successful, it is marked as an authorized behavior, and then the noise digital signature is decrypted to obtain the privacy target data corresponding to the noise digital signature; Otherwise, abnormal behavior is generated.

Images