Quality evaluation method and system for data cleaning and labeling oriented to AI (Artificial Intelligence) training

Abstract

The invention provides an AI training-oriented data cleaning and labeling quality evaluation method and system, and relates to the technical field of information security, and the method comprises the steps: obtaining an encrypted sample set which is from an untrusted domain and is used for AI training and standard reference data from a trusted domain; converting the standard reference data into a standard reference vector matched with the feature dimension of the encrypted sample subset based on the time information of each ciphertext sample; in the encryption state, obtaining a plurality of ciphertext feature descriptors; respectively calculating a distribution deviation value between each ciphertext feature descriptor and the standard reference vector; and when the distribution deviation value is greater than a preset safety threshold value, determining that the corresponding encrypted sample subset has distribution abnormity and removing the distribution abnormity in the encrypted sample set, thereby realizing quality evaluation of data cleaning and labeling, and improving the quality and reliability of a downstream AI model training sample while ensuring data privacy safety.

Description

Technical Field

[0001] This application relates to the field of information security technology, and in particular to a method and system for quality assessment of data cleaning and labeling for AI training. Background Technology

[0002] With the widespread application of artificial intelligence technology in fields such as financial risk control, smart healthcare, and government integration, high-quality training data has become a key element in ensuring model performance.

[0003] In existing technologies, a centralized data quality assessment scheme is usually adopted, which requires data from all parties to be aggregated to a central server, and abnormal data and labeling errors are detected by calculating the statistical distribution characteristics of the samples in plaintext.

[0004] However, in cross-institutional collaboration scenarios involving privacy-sensitive data, plaintext data aggregation faces severe privacy leakage risks and compliance challenges, rendering traditional plaintext-based evaluation methods ineffective. Directly processing encrypted data using conventional methods fails because the encryption mechanism destroys the original statistical characteristics of the data, making it impossible to directly apply existing distribution measurement algorithms. This makes it difficult for participating parties to effectively evaluate the quality of data and labels without data visibility. Therefore, existing technologies suffer from the technical problem of failing to effectively evaluate and clean encrypted training samples and labels while ensuring data privacy and security. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for quality assessment of data cleaning and labeling for AI training, in order to solve the technical problem that existing technologies cannot effectively assess and clean encrypted training samples and labels while ensuring data privacy and security.

[0006] Firstly, this application provides a quality assessment method for data cleaning and labeling for AI training, including:

[0007] Acquire an encrypted sample set for AI training from an untrusted domain and standard reference data from a trusted domain. The encrypted sample set includes multiple ciphertext samples and multiple ciphertext labels. The standard reference data includes statistical summary data of benign data corresponding to the encrypted sample set and the ciphertext labels.

[0008] Based on the time information of each ciphertext sample, the encrypted sample set is constructed by block indexing to obtain multiple encrypted sample subsets, and the standard reference data is transformed into a standard reference vector that matches the feature dimensions of the encrypted sample subsets.

[0009] In the encrypted state, a polynomial approximation calculation is performed on each encrypted sample subset to obtain the ciphertext feature information of each encrypted sample subset. Then, using the standard reference vector as the standard, the dimension alignment of each ciphertext feature information is performed to map the position of the encrypted sample subset in the statistical distribution space, resulting in multiple ciphertext feature descriptors.

[0010] Calculate the distribution offset value between each ciphertext feature descriptor and the standard reference vector;

[0011] When the distribution offset value is greater than the preset security threshold, it is determined that the corresponding encrypted sample subset has a distribution anomaly and is removed from the encrypted sample set to achieve quality assessment of data cleaning and labeling.

[0012] Optionally, the method further includes:

[0013] The subset of encrypted samples whose distribution offset value is less than or equal to a preset security threshold is considered a qualified subset of encrypted samples.

[0014] Perform ciphertext summation on the ciphertext feature descriptors corresponding to all qualified encrypted sample subsets to obtain updated statistical information;

[0015] Based on the updated statistical information, the standard reference vector is updated using a homomorphic encryption algorithm to obtain the updated standard reference vector.

[0016] Optionally, the method further includes:

[0017] Obtain a feature importance vector derived from the trusted domain, the feature importance vector including weight values ​​corresponding to the feature dimensions of each ciphertext feature descriptor;

[0018] Using a homomorphic encryption algorithm, the standard reference vector and the feature importance vector are multiplied in ciphertext to obtain a weighted standard reference vector;

[0019] Each feature element in the ciphertext feature descriptor is multiplied by the corresponding weight value in the feature importance vector to obtain multiple weighted ciphertext feature descriptors.

[0020] The step of calculating the distribution offset value between each ciphertext feature descriptor and the standard reference vector includes:

[0021] Calculate the distribution offset value between each weighted ciphertext feature descriptor and the weighted standard reference vector.

[0022] Optionally, calculating the distribution offset between each weighted ciphertext feature descriptor and the weighted standard reference vector includes:

[0023] Using the subtraction property of homomorphic encryption, the difference vector between each weighted ciphertext feature descriptor and the weighted standard reference vector in the corresponding dimension is calculated.

[0024] Utilizing the multiplication property of homomorphic encryption, the square of each vector element in the difference vector is calculated;

[0025] By utilizing the addition property of homomorphic encryption, the summation of all squared terms in each difference vector is performed to obtain the distribution offset value of each encrypted sample subset relative to the standard reference data.

[0026] Optionally, the standard reference data includes the first-order and second-order raw moments corresponding to each feature dimension in the encrypted sample set and the ciphertext label;

[0027] The process involves constructing a block index for the encrypted sample set based on the time information of each encrypted sample, resulting in multiple encrypted sample subsets. The standard reference data is then transformed into a standard reference vector that matches the feature dimensions of the encrypted sample subsets, including:

[0028] The time span of the encrypted sample set is divided into multiple consecutive time intervals based on a preset time window length;

[0029] The time interval to which each ciphertext sample belongs is determined based on the time information of each ciphertext sample, and a corresponding block index identifier is assigned to each ciphertext sample based on the time interval.

[0030] By combining ciphertext samples with the same block index identifier, multiple encrypted sample subsets are obtained;

[0031] According to the preset feature arrangement order, the first-order and second-order origin moments of each feature dimension in the standard reference data are concatenated to obtain a standard reference vector that matches the feature dimensions of the encrypted sample subset.

[0032] Optionally, in the encrypted state, performing a polynomial approximation calculation on each encrypted sample subset to obtain the ciphertext feature information of each encrypted sample subset includes:

[0033] In the encrypted state, each ciphertext sample and its corresponding ciphertext label in each encrypted sample subset are combined into a ciphertext vector to be processed, and the square value of each element in the ciphertext vector to be processed is calculated.

[0034] The original values ​​of all ciphertext vectors to be processed within each encrypted sample subset are summed in ciphertext to obtain the first summed value. The squared values ​​of all ciphertext vectors to be processed within each encrypted sample subset are summed in ciphertext to obtain the second summed value.

[0035] The first accumulated value and the second accumulated value are combined to obtain the ciphertext feature information of each encrypted sample subset.

[0036] Optionally, the step of dimensionally aligning each ciphertext feature information using the standard reference vector to map the position of the encrypted sample subset in the statistical distribution space yields multiple ciphertext feature descriptors, including:

[0037] According to the arrangement rules of the standard reference vector, the first and second accumulated values ​​in each ciphertext feature information are rearranged to obtain the rearranged ciphertext feature information;

[0038] Obtain the number of samples in each encrypted sample subset, and generate the reciprocal factor of each encrypted sample subset based on the number of samples;

[0039] Based on the reciprocal factor, the multiplication property of homomorphic encryption is used to perform mean processing on the sorted ciphertext feature information to obtain multiple statistical distribution coordinates. The statistical distribution coordinates are the positions of the corresponding encrypted sample subsets mapped to the statistical distribution space.

[0040] Each statistical distribution coordinate is used as the ciphertext feature descriptor for the corresponding encrypted sample subset.

[0041] Secondly, this application provides an automated data cleaning and labeling quality assessment system for AI training, including:

[0042] The acquisition module is used to acquire an encrypted sample set for AI training from an untrusted domain and standard reference data from a trusted domain. The encrypted sample set includes multiple ciphertext samples and multiple ciphertext labels, and the standard reference data includes statistical summary data of benign data corresponding to the encrypted sample set and the ciphertext labels.

[0043] The construction module is used to construct a block index for the encrypted sample set based on the time information of each encrypted sample, to obtain multiple encrypted sample subsets, and to transform the standard reference data into a standard reference vector that matches the feature dimensions of the encrypted sample subsets.

[0044] The calculation module is used to perform polynomial approximation calculation on each encrypted sample subset in the encrypted state to obtain the ciphertext feature information of each encrypted sample subset, and to perform dimensional alignment on each ciphertext feature information using the standard reference vector as the standard to map the position of the encrypted sample subset in the statistical distribution space to obtain multiple ciphertext feature descriptors.

[0045] The calculation module is also used to calculate the distribution offset value between each ciphertext feature descriptor and the standard reference vector;

[0046] The elimination module is used to determine that the corresponding encrypted sample subset has a distribution anomaly and to eliminate it from the encrypted sample set when the distribution offset value is greater than a preset security threshold, so as to achieve quality assessment of data cleaning and labeling.

[0047] Thirdly, this application provides an electronic device, comprising:

[0048] Memory, used to store computer programs;

[0049] A processor, used to execute the computer program to implement the steps of the data cleaning and labeling quality assessment method for AI training as described in the first aspect above.

[0050] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the data cleaning and labeling quality assessment method for AI training as described in the first aspect above.

[0051] The data cleaning and labeling quality assessment method for AI training provided in this application addresses the privacy leakage risks and compliance challenges arising from plaintext data transmission and aggregation in cross-institutional collaboration scenarios by obtaining encrypted samples and labels at the source. It resolves the data structure heterogeneity issue between encrypted environments and statistical summaries. It enables the reconstruction and extraction of encrypted feature descriptors reflecting data distribution characteristics without decryption. It ensures accurate assessment of the deviation of the current batch of data from a benign distribution even when the data is not visible. While guaranteeing data privacy and security, it also solves the problem of performance degradation in downstream AI models caused by the infiltration of low-quality data.

[0052] Furthermore, this application employs a dynamic feedback update mechanism to select a qualified subset of encrypted samples. Utilizing the cumulative property of homomorphic encryption, the ciphertext feature descriptors of these qualified subsets are aggregated to generate updated statistical information, and the standard reference vector is iteratively updated. This updated standard reference vector is then used in subsequent evaluation cycles. This addresses the problem of static standard reference data struggling to cope with data distribution drift common in financial or medical scenarios, ensuring that the evaluation benchmark can be adjusted to follow the actual evolution trend of healthy data. It also improves the usability of the cleaned data and the robustness of model training. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 A flowchart illustrating the quality assessment method for data cleaning and labeling for AI training provided in this application embodiment;

[0055] Figure 2 A flowchart illustrating a method for obtaining an updated standard reference vector, provided in an embodiment of this application;

[0056] Figure 3 A flowchart illustrating a method for obtaining a weighted ciphertext feature descriptor provided in an embodiment of this application;

[0057] Figure 4 A schematic diagram of the structure of an automated data cleaning and labeling quality assessment system for AI training provided in an embodiment of this application;

[0058] Figure 5 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0059] Given the contradiction between data privacy protection and the need for high-quality AI training in scenarios such as financial risk control, smart healthcare, and government integration, and considering the serious privacy leakage risks faced by existing centralized plaintext evaluation schemes, as well as the technical challenges of being unable to effectively evaluate data and label quality when data is not visible due to the fact that conventional encryption methods destroy the statistical characteristics of data.

[0060] This application addresses the issue of secure data source isolation by obtaining a set of encrypted samples from the untrusted domain and a benign statistical summary from the trusted domain. Furthermore, it utilizes block indexing and dimension alignment techniques to establish a unified computational benchmark, overcoming the heterogeneity barrier of data structures in encrypted environments. The core innovation lies in introducing polynomial approximation computation under homomorphic encryption to reconstruct data feature information and map its statistical distribution position within the ciphertext domain, thereby restoring the statistical characteristics masked by encryption. Finally, based on the distribution offset metric between ciphertext feature descriptors and standard reference vectors, it achieves anomaly detection and removal in a fully ciphertext environment. This ensures strict privacy compliance while accurately identifying and cleaning low-quality data, guaranteeing the training efficiency of downstream AI models.

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The core of this application is to provide a quality assessment method for data cleaning and labeling for AI training. A flowchart illustrating one specific implementation is shown below. Figure 1 As shown, the method includes:

[0063] In this embodiment, prior to data acquisition, the homomorphic encryption system needs to be initialized and configured. Specifically, the key management terminal of the trusted domain executes a key generation algorithm to generate a public key, a private key, and an evaluation key. The public key is distributed to the untrusted domain for encrypting the original training samples and labels; the private key is strictly stored in the decryption terminal of the trusted domain and is strictly prohibited from transmission over the network; the evaluation key, including a relinearization key and a rotation key, is distributed to the nodes performing the quality evaluation calculation. In particular, considering that ciphertext multiplication operations can lead to ciphertext dimensionality expansion, a relinearization key is required. After subsequent ciphertext multiplication operations, this relinearization key is used to remap the expanded ciphertext back to the original dimensional space to ensure continuous computation capability and control communication overhead.

[0064] Step 101: Obtain an encrypted sample set for AI training from the untrusted domain and standard reference data from the trusted domain. The encrypted sample set includes multiple ciphertext samples and multiple ciphertext labels. The standard reference data includes statistical summary data of benign data corresponding to the encrypted sample set and the ciphertext labels.

[0065] In this step, the untrusted domain refers to the data holder or data transmission environment whose original data cannot be directly shared in plaintext due to privacy regulations or data security restrictions. The encrypted sample set refers to the dataset processed using homomorphic encryption technology. The ciphertext sample refers to the encrypted feature data vector within the encrypted sample set. The ciphertext label refers to the encrypted classification label or regression value corresponding to the ciphertext sample within the encrypted sample set. The trusted domain refers to a computing environment or authoritative third-party institution with security certification and data auditing capabilities. The standard reference data refers to the dataset used as a quality assessment benchmark, including the first-order and second-order raw moments corresponding to each feature dimension in the encrypted sample set and the ciphertext label. Benign data refers to verified historical high-quality data that does not contain malicious noise and has a normal distribution. Statistical summary data refers to metadata obtained after statistical calculations on the benign data.

[0066] In this embodiment, a privacy-preserving gateway is first deployed at the untrusted domain to collect the original training samples and labels. Before encrypting the samples using a homomorphic encryption algorithm, the original training samples and labels need to be standardized and preprocessed and encoded. Since homomorphic encryption algorithms typically only support operations on integers or real polynomials, non-numerical data needs to be converted into numerical data. For example, for text-type labels, one-hot encoding or label encoding is used to map them into numerical vectors; for missing values, mean-filling or zero-filling strategies are used to complete them.

[0067] Subsequently, the processed real-number vector is encoded into a plaintext polynomial using an encoder in a homomorphic encryption scheme. The encoded plaintext polynomial is then encrypted using a homomorphic encryption algorithm to obtain ciphertext samples and ciphertext labels. Multiple ciphertext samples and labels are combined to generate an encrypted sample set. Simultaneously, benign data is acquired within the trusted domain. Statistical indices for each feature dimension and label in the benign data are calculated in plaintext to obtain statistical summary data, which is then encapsulated as standard reference data.

[0068] In the embodiments of this application, the homomorphic encryption algorithm may employ, but is not limited to, a fully homomorphic encryption scheme or some homomorphic encryption schemes that support additive and multiplicative homomorphism, such as the CKKS (Cheon-Kim-Kim-Song) scheme, the BGV (Brakerski-Gentry-Vaikuntanathan) scheme, or the BFV (Brakerski / Fan-Vercauteren) scheme. Among them, the CKKS scheme is particularly suitable for the statistical moment calculation and Euclidean distance metric scenarios involved in this application due to its efficient support for real number operations and approximate calculation characteristics.

[0069] For example, in a scenario where Institution A and Institution B conduct joint financial risk control modeling, Institution A, acting as the untrusted domain, encrypts the user transaction data it holds, generating an encrypted sample set S. This set includes ciphertext samples X and ciphertext labels Y. Ciphertext sample X includes encrypted values ​​of transaction amount, transaction frequency, and transaction time. Institution B, acting as the trusted domain, calculates the average transaction amount U and the second-order original moment M of the transaction amount based on historical fraud-free transaction data, using these as standard reference data.

[0070] Step 102: Based on the time information of each ciphertext sample, construct a block index for the encrypted sample set to obtain multiple encrypted sample subsets, and transform the standard reference data into a standard reference vector that matches the feature dimensions of the encrypted sample subsets.

[0071] In this step, time information refers to the timestamp information in the encrypted sample data used to record the time of data generation or the data collection batch. Feature dimension refers to the number of dimensions in the encrypted sample data vector or the number of dimensions in the encrypted labels. The standard reference vector refers to the vector data generated by arranging and combining the statistical indicators in the standard reference data in a specific order.

[0072] Step 201: Divide the time span of the encrypted sample set into multiple consecutive time intervals based on a preset time window length.

[0073] In this step, the preset time window length refers to the time granularity parameter used to segment the data stream, which can be a fixed number of minutes, hours, or days.

[0074] In this embodiment, a pre-set time window length is first obtained. This length can be dynamically configured based on the actual data throughput of the business and the performance requirements of privacy computing. Based on this time window length, the overall time span covered by the encrypted sample set is evenly divided to generate continuous time intervals.

[0075] For example, the time window length T is set to 30 minutes. The time span of the encrypted transaction data collected by Institution A within a day from 0:00 to 24:00 is divided into 48 consecutive time intervals, such as interval 1 from 00:00 to 00:30 and interval 2 from 00:30 to 01:00.

[0076] Step 202: Determine the time interval to which each ciphertext sample belongs based on the time information of each ciphertext sample, and assign a corresponding block index identifier to each ciphertext sample based on the time interval.

[0077] In this step, the block index identifier is a unique identifier assigned to each ciphertext sample, which indicates the time interval batch to which the sample belongs.

[0078] In this embodiment, each ciphertext sample in the encrypted sample set is traversed, its accompanying time information is read, and it is compared with the predefined time intervals to determine the specific time interval in which the sample falls. Based on the comparison results, ciphertext samples falling into the same time interval are assigned the same block index identifier, thereby achieving logical grouping.

[0079] For example, if the time information of a certain ciphertext sample X is read as 00:15, it is determined to belong to interval 1, and a block index identifier I1 is assigned to it. If the time information of another ciphertext sample X is read as 00:45, it is determined to belong to interval 2, and a block index identifier I2 is assigned to it.

[0080] Step 203: Combine ciphertext samples with the same block index identifier to obtain multiple encrypted sample subsets.

[0081] In this step, the encrypted sample subset refers to the set of ciphertext samples and ciphertext tags with the same index identifier obtained after constructing the block index.

[0082] In this embodiment of the application, all ciphertext samples carrying the same block index identifier and their corresponding ciphertext tags are physically or logically combined to construct multiple independent subsets of encrypted samples. For example, all samples identified as I1 are combined into an encrypted sample subset S1.

[0083] Step 204: According to the preset feature arrangement order, the first-order and second-order original moments of each feature dimension in the standard reference data are concatenated to obtain a standard reference vector that matches the feature dimensions of the encrypted sample subset.

[0084] In this step, the preset feature arrangement order refers to the predefined connection order of the various statistical indicators when constructing the vector. The first-order raw moment refers to the mean statistic of the data. The second-order raw moment refers to the squared mean statistic of the data.

[0085] In this embodiment, according to a preset feature arrangement order, the first-order and second-order raw moments corresponding to each feature dimension and label are extracted from the standard reference data. These extracted statistics are then concatenated sequentially to generate a one-dimensional numerical vector, i.e., the standard reference vector.

[0086] For example, for standard reference data, the first-order raw moment U1 and second-order raw moment M1 of transaction amount, and the first-order raw moment U2 and second-order raw moment M2 of transaction frequency are extracted. These are concatenated in the order of amount first, then frequency, and first-order first, then second-order, to obtain the standard reference vector V, represented as [U1, M1, U2, M2]. This vector V will serve as a benchmark for measuring the distribution quality of the encrypted sample subset S1.

[0087] Step 103: In the encrypted state, perform polynomial approximation calculation on each encrypted sample subset to obtain the ciphertext feature information of each encrypted sample subset, and use the standard reference vector as the standard to perform dimensional alignment on each ciphertext feature information to map the position of the encrypted sample subset in the statistical distribution space, thereby obtaining multiple ciphertext feature descriptors.

[0088] In this step, ciphertext feature information refers to the intermediate statistical aggregation result of the encrypted sample subset calculated under homomorphic encryption, before it has undergone mean-based processing. The ciphertext feature descriptor refers to the final ciphertext vector, after mean-based and structure-aligned processing, that accurately represents the coordinate position of the encrypted sample subset in the statistical distribution space.

[0089] In this application, the polynomial approximation calculation refers to using low-order polynomials, such as quadratic polynomials, to approximate the statistical distribution characteristics of the data. Specifically, since homomorphic encryption incurs huge computational costs when dealing with complex nonlinear functions, such as probability density functions, this embodiment constructs a first-order raw moment (mean) and a second-order raw moment (square mean) within the ciphertext domain by calculating the squared terms and accumulation terms of the ciphertext elements, thereby approximating the true distribution characteristics of the data in the form of polynomial operations.

[0090] Step 301: In the encrypted state, combine each ciphertext sample and its corresponding ciphertext label in each encrypted sample subset into a ciphertext vector to be processed, and calculate the square value of each element in the ciphertext vector to be processed.

[0091] In this step, the ciphertext vector to be processed refers to the unified encrypted vector formed by concatenating the feature data of the ciphertext sample with the corresponding ciphertext label in terms of dimension.

[0092] In this embodiment, for each piece of data in the encrypted sample subset, its ciphertext sample and ciphertext label are first obtained. The ciphertext label is then appended as a new feature dimension to the end of the ciphertext sample to construct the ciphertext vector to be processed. Next, utilizing the ciphertext multiplication property supported by the homomorphic encryption algorithm, each encrypted element in the ciphertext vector to be processed is multiplied by itself to obtain the square value corresponding to each dimension of the vector.

[0093] For example, based on a data point in the encrypted sample subset S1, its ciphertext sample X is [E10] and its ciphertext label Y is [E1]. The two are concatenated to obtain the ciphertext vector V to be processed, denoted as [E10, E1]. A homomorphic square operation is performed on the ciphertext vector V to obtain the squared value vector V2, denoted as [E100, E1].

[0094] Step 302: Accumulate the original values ​​of all ciphertext vectors to be processed within each encrypted sample subset to obtain the first accumulated value, and accumulate the square values ​​of all ciphertext vectors to be processed within each encrypted sample subset to obtain the second accumulated value.

[0095] In this step, the original value refers to the original encrypted value of the ciphertext vector to be processed without squaring. The first accumulated value is the ciphertext sum of the first moments of all data in the encrypted sample subset. The second accumulated value is the ciphertext sum of the second moments of all data in the encrypted sample subset.

[0096] In this embodiment, all data in the encrypted sample subset is traversed. Utilizing the addition property of homomorphic encryption, the original value of each ciphertext vector to be processed is sequentially accumulated into the first accumulator, yielding the first accumulated value. Simultaneously, the square value of each ciphertext vector to be processed is sequentially accumulated into the second accumulator, yielding the second accumulated value.

[0097] For example, suppose the encrypted sample subset S1 includes two data items: data 1 has a vector Va of [E10, E1] and a squared value Vb of [E100, E1]; data 2 has a vector Vc of [E20, E0] and a squared value Vd of [E400, E0]. After ciphertext accumulation, the first accumulated value A1 is [E30, E1], and the second accumulated value A2 is [E500, E1].

[0098] Step 303: Combine the first accumulated value with the second accumulated value to obtain the ciphertext feature information of each encrypted sample subset.

[0099] In this step, ciphertext feature information refers to encrypted aggregated data that includes first-order and second-order statistical properties.

[0100] In this embodiment, the calculated first and second accumulated values ​​are concatenated along their vector dimensions. Typically, the first accumulated value is placed before the second accumulated value, forming an extended ciphertext vector, which represents the ciphertext feature information of the encrypted sample subset. This information fully preserves the distributional statistical characteristics of the subset and remains completely undecrypted.

[0101] For example, the first accumulated value A1 and the second accumulated value A2 are concatenated to obtain the ciphertext feature information I1, which has the form [E30, E1, E500, E1]. It is worth noting that a noise management mechanism is implemented based on the noise generated by homomorphic multiplication operations. After the squaring operation or subsequent multiplication operations, the noise level of the ciphertext is checked. If the noise is close to a threshold, an analog-to-digital conversion operation, i.e., a rescaling operation, is performed. This refreshes the ciphertext noise by discarding some low-order data, controlling the accumulated noise within an acceptable range. This ensures that after multiple rounds of polynomial approximation calculations, the obtained ciphertext feature information still maintains high computational accuracy and decryptability.

[0102] Step 311: According to the arrangement rules of the standard reference vector, rearrange the first and second accumulated values ​​in each ciphertext feature information to obtain the rearranged ciphertext feature information.

[0103] In this embodiment, the metadata definition of the standard reference vector is read to obtain the arrangement order of its feature dimensions. Using this order as a template, the encrypted feature information is reorganized. Specifically, the first and second accumulated value elements in the encrypted feature information are moved and swapped according to the template order to ensure that the accumulated value of each feature dimension is aligned to the position of the corresponding statistic in the standard reference vector.

[0104] For example, suppose the standard reference vector V is in the order of [first moment of transaction amount, second moment of transaction amount, first moment of label, second moment of label]. The encrypted feature information I1 is [first accumulated value of transaction amount, first accumulated value of label, second accumulated value of transaction amount, second accumulated value of label]. According to the arrangement rules, I1 is rearranged into I2, in the form of [first accumulated value of transaction amount, second accumulated value of transaction amount, first accumulated value of label, second accumulated value of label].

[0105] Step 312: Obtain the number of samples in each encrypted sample subset, and generate the reciprocal factor of each encrypted sample subset based on the number of samples.

[0106] In this step, the number of samples refers to the total number of ciphertext samples in the currently processed subset of encrypted samples. The reciprocal factor is the reciprocal of the number of samples.

[0107] In this embodiment, the number of samples N in the current encrypted sample subset is counted. The reciprocal 1 / N of this value N is calculated to obtain the reciprocal factor. This factor will be used in subsequent steps to convert the sum to the mean. Here, the number of samples N is a public parameter or a plaintext value that can be calculated by the trusted domain during the preprocessing stage, thus enabling the pre-calculation of its reciprocal factor 1 / N as the operand for subsequent homomorphic scalar multiplication, avoiding the technical difficulty of directly performing ciphertext division under homomorphic encryption. For example, assuming that the encrypted sample subset S1 includes 2 data items, i.e., the number of samples N is 2, the reciprocal factor K is calculated to be 0.5.

[0108] Step 313: Based on the reciprocal factor, using the multiplication property of homomorphic encryption, the sorted ciphertext feature information is averaged to obtain multiple statistical distribution coordinates, where the statistical distribution coordinates are the positions of the corresponding encrypted sample subsets mapped to the statistical distribution space.

[0109] In this step, the statistical distribution coordinates refer to the ciphertext vector after mean normalization, whose values ​​represent the central location and dispersion of the data within the ciphertext domain.

[0110] In this embodiment, the scalar multiplication property supported by the homomorphic encryption algorithm is utilized to multiply each ciphertext element in the arranged ciphertext feature information vector by its reciprocal factor. This operation is equivalent to a division operation in the ciphertext state, thereby converting the accumulated sum into the mean and the second-order raw moment, completing the mapping from aggregate total to distribution characteristics.

[0111] The core principle is to ensure that the encrypted sample subset to be evaluated and the standard reference data are in the same coordinate system in the statistical distribution space. By unifying the feature arrangement order and numerical dimensions of the two, the subsequent ciphertext subtraction operation can be performed accurately on a statistically consistent dimension with the same physical meaning, thereby ensuring the mathematical validity of the distribution offset calculation.

[0112] For example, each element in the rearranged ciphertext feature information I2 is multiplied by the reciprocal factor 0.5. For instance, multiplying [E30,E500,E1,E1] by 0.5 yields the statistical distribution coordinates C1 [E15,E250,E0.5,E0.5].

[0113] Step 314: Determine each statistical distribution coordinate as the ciphertext feature descriptor of the corresponding encrypted sample subset.

[0114] In this embodiment, the statistical distribution coordinates are directly labeled as the ciphertext feature descriptor of the encrypted sample subset. This descriptor corresponds completely to the standard reference vector in both dimensional structure and physical meaning, and can be directly input into the subsequent offset calculation module. For example, the statistical distribution coordinates C1 are determined as the ciphertext feature descriptor D1 of the encrypted sample subset S1. D1 represents the position of S1 in the statistical space.

[0115] Step 104: Calculate the distribution offset value between each ciphertext feature descriptor and the standard reference vector.

[0116] In this step, the distribution offset value refers to a quantitative indicator calculated within the ciphertext domain. It is used to measure the statistical distance difference between the encrypted sample subset to be evaluated and the benign standard reference data. The larger the value, the more significant the distribution difference between the two.

[0117] Figure 3 This is a flowchart illustrating a method for obtaining a weighted ciphertext feature descriptor, provided in an embodiment of this application.

[0118] Step 10411: Obtain the feature importance vector derived from the trusted domain, the feature importance vector including weight values ​​corresponding to the feature dimensions of each ciphertext feature descriptor.

[0119] In this step, the feature importance vector refers to a list of values ​​predefined by trusted domain experts to quantify the importance of each feature dimension in the quality assessment.

[0120] In this embodiment, a pre-configured feature importance vector is first obtained from the trusted domain. The dimension of this vector is exactly the same as that of the encrypted feature descriptor and the standard reference vector. The weight values ​​are typically set based on expert experience or by training a feature selection model on a benign dataset. The obtained feature importance vector can be in plaintext or encrypted form, depending on the security strategy.

[0121] For example, the obtained feature importance vector W is [0.8, 0.8, 0.2, 0.2], which represents the transaction amount related features, i.e., the weight of the first two dimensions is 0.8, and the transaction frequency related features, i.e., the weight of the last two dimensions is 0.2.

[0122] Step 10412: Using a homomorphic encryption algorithm, perform ciphertext multiplication on the standard reference vector and the feature importance vector to obtain a weighted standard reference vector.

[0123] In this embodiment, the scalar multiplication or ciphertext multiplication property supported by homomorphic encryption algorithms is utilized to multiply each ciphertext element in the standard reference vector by the weight value at the corresponding position in the feature importance vector. This operation incorporates the importance of the features into the baseline data. After calculation, a new ciphertext vector, namely the weighted standard reference vector, is generated.

[0124] For example, for the standard reference vector V=[U1,M1,U2,M2], multiplying it element-wise with W yields the weighted standard reference vector Va=[0.8U1,0.8M1,0.2U2,0.2M2].

[0125] Step 10413: Perform ciphertext multiplication on each feature element in the ciphertext feature descriptor and the corresponding weight value in the feature importance vector to obtain multiple weighted ciphertext feature descriptors.

[0126] In this embodiment, each encrypted feature descriptor is traversed, and for each feature dimension element, a homomorphic multiplication operation is performed between it and the weight value of the corresponding dimension in the feature importance vector. This process ensures that the data to be evaluated and the benchmark data undergo the exact same weighting transformation. After processing all feature dimensions, the weighted encrypted feature descriptors are obtained.

[0127] For example, for the ciphertext feature descriptor D1=[Eu1,Em1,Eu2,Em2] of the encrypted sample subset S1, multiplying it by W yields the weighted ciphertext feature descriptor Da=[0.8Eu1,0.8Em1,0.2Eu2,0.2Em2].

[0128] The step of calculating the distribution offset value between each ciphertext feature descriptor and the standard reference vector includes:

[0129] Calculate the distribution offset value between each weighted ciphertext feature descriptor and the weighted standard reference vector.

[0130] Step 401: Utilize the subtraction property of homomorphic encryption to calculate the difference vector between each weighted ciphertext feature descriptor and the weighted standard reference vector in the corresponding dimension.

[0131] In this step, the subtraction property of homomorphic encryption refers to the property of homomorphic encryption algorithms that allow operations on two ciphertexts to be performed, with the decryption result being equal to the subtraction of the corresponding plaintexts.

[0132] In this embodiment, each weighted ciphertext feature descriptor is traversed, and the ciphertext subtraction property supported by the homomorphic encryption algorithm is used to perform element-wise subtraction with the weighted standard reference vector in each corresponding feature dimension. This operation calculates the numerical difference between the data feature to be evaluated and the benign benchmark feature without decryption, generating a difference vector in ciphertext form consistent with the original feature dimensions.

[0133] For example, assume the weighted standard reference vector Va is [E10, E198, E0.1, E0.1]. The weighted ciphertext feature descriptor Da calculated in the previous steps is [E12, E200, E0.1, E0.1], which is the product of D1 and W mentioned earlier. First, calculate the difference between Da and Va to obtain the difference vector Vd = [E2, E2, E0, E0].

[0134] Step 402: Utilize the multiplication property of homomorphic encryption to calculate the square of each vector element in the difference vector.

[0135] In this step, the multiplication property of homomorphic encryption refers to the property of homomorphic encryption algorithms that allow operations on two ciphertexts to be performed, with the decryption result being equal to the product of the corresponding plaintexts.

[0136] In this embodiment, based on the difference vector, the ciphertext multiplication property supported by the homomorphic encryption algorithm is utilized to multiply each ciphertext element in the vector by itself. Through this operation, the feature difference values ​​of each dimension are transformed into non-negative squared forms. After the calculation is completed, an intermediate vector including the squared difference values ​​of all feature dimensions is obtained. For example, the squares of the elements in Vd are then calculated to obtain the squared term vector Vs=[E4,E4,E0,E0].

[0137] Step 403: Utilize the addition property of homomorphic encryption to accumulate all squared terms in each difference vector to obtain the distribution offset value of each encrypted sample subset relative to the standard reference data.

[0138] In this step, the addition property of homomorphic encryption refers to the property of homomorphic encryption algorithms that allow operations on two ciphertexts to be performed, and the decryption result is equal to the sum of the corresponding plaintexts.

[0139] In this embodiment, the ciphertext addition property supported by the homomorphic encryption algorithm is utilized to accumulate the elements in the vector that includes the squared terms of all feature dimensions. The accumulated result is output as a single ciphertext value, which is the distribution offset value of the encrypted sample subset.

[0140] By ciphertext-accumulating the squared terms of the differences in features across each dimension, the resulting distribution offset is physically equivalent to calculating the square of the weighted Euclidean distance between the encrypted sample subset and the standard reference vector in the statistical feature space. This metric can sensitively capture minute changes in the sample set's mean shift or dispersion anomalies, thus providing a quantitative basis for anomaly detection.

[0141] For example, the elements in Vs are summed to obtain the distribution offset value S1 as E8. This value E8 represents the distribution offset of the encrypted sample subset S1 relative to the standard reference data.

[0142] Step 105: When the distribution offset value is greater than the preset security threshold, it is determined that the corresponding encrypted sample subset has a distribution anomaly and is removed from the encrypted sample set to achieve quality assessment of data cleaning and labeling.

[0143] In this step, the preset security threshold refers to the upper limit of the encrypted value calculated by business experts or based on the distribution of historical benign data, which is used to determine whether the data quality is acceptable.

[0144] In this embodiment, a secure comparison algorithm within a secure multi-party computation protocol can be used to compare the distribution offset and a preset security threshold without decrypting them. If the comparison shows that the distribution offset is greater than the preset security threshold, a judgment signal is generated to confirm that the encrypted sample subset has a distribution anomaly, including feature distribution drift or label quality issues. A cleaning mechanism is then triggered, adding the block index identifier of the anomaly subset to a removal list. During subsequent model training data loading, all ciphertext samples corresponding to this identifier are automatically filtered out. Conversely, if the offset is less than or equal to the threshold, the data is retained. Finally, a cleaned, high-quality encrypted training dataset is output.

[0145] In this embodiment, the determination of the preset security threshold can be performed offline based on statistical principles. Specifically, during the system initialization phase, a subset of known benign encrypted data is selected as a verification set. Steps 103 and 104 are used to calculate the distribution offset value of each subset within this verification set, thereby constructing a probability distribution model of the benign distribution offset value. Based on this probability distribution model, the 99th percentile or three times the standard deviation boundary can be selected as the preset security threshold. This setting method ensures that the vast majority of normal benign data fluctuations are deemed acceptable, while also effectively capturing anomalous data subsets that significantly deviate from the normal range.

[0146] For example, the preset security threshold is set to E5. For the encrypted sample subset S1, its distribution offset value is E8. A security comparison protocol reveals that E8 is greater than E5, therefore S1 is determined to have a distribution anomaly, is marked as unusable, and removed from the training set. For another subset S2, if its distribution offset value is E3, which is less than E5, it is considered qualified and retained for AI model training.

[0147] Figure 2 This is a flowchart illustrating a method for obtaining an updated standard reference vector, provided in an embodiment of this application.

[0148] Step 10511: Select the subset of encrypted samples whose distribution offset value is less than or equal to the preset security threshold as the qualified subset of encrypted samples.

[0149] In this step, the qualified encrypted sample subset refers to encrypted data blocks that have been determined by quality assessment to be normally distributed, free of abnormal noise, and with reliable labels. These are allowed to be used for subsequent model training and standard updates.

[0150] In this embodiment, based on the comparison results, a subset of encrypted samples whose distribution offset values ​​do not exceed a preset security threshold are selected. The index identifiers of these subsets are recorded in a qualified list, and the corresponding ciphertext feature descriptors are extracted as the basis data for subsequent incremental updates. For example, for subset S2, its offset value E3 is less than the threshold E5, therefore S2 is marked as a qualified encrypted sample subset.

[0151] Step 10512: Perform ciphertext summation on the ciphertext feature descriptors corresponding to all qualified encrypted sample subsets to obtain updated statistical information.

[0152] In this step, updating statistics refers to the encrypted incremental data obtained by aggregating the characteristics of all qualified data in the current batch, which reflects the latest positive data distribution trend.

[0153] In this embodiment, ciphertext feature descriptors corresponding to all qualified encrypted sample subsets are obtained. Utilizing the additive property of homomorphic encryption, these descriptor vectors are accumulated dimension-by-dimensionally with their ciphertext. If the number of qualified subsets is large, their ciphertext mean can be further calculated, i.e., the accumulated value multiplied by the reciprocal of the number of qualified subsets. The resulting aggregate vector is the update statistics, representing the statistical distribution of newly incoming high-quality data within the current time window.

[0154] For example, suppose there are two qualified subsets S2 and S3 in the current batch, with descriptors D2=[E20,E400] and D3=[E22,E410], respectively. The sum of these two subsets and the average value are used to obtain the updated statistics Un=[E21,E405].

[0155] Step 10513: Based on the updated statistical information, update the standard reference vector using a homomorphic encryption algorithm to obtain the updated standard reference vector.

[0156] In this step, the updated standard reference vector refers to a new benchmark vector that incorporates historical benchmarks and the latest qualified data features, which is used for quality assessment in the next cycle.

[0157] In this embodiment, an exponentially weighted moving average or a simple weighted fusion strategy is used to update the benchmark. Specifically, a preset update weight factor k is obtained, and the product of the update statistics and k, as well as the product of the current standard reference vector and 1-k, are calculated using homomorphic scalar multiplication. Then, the two results are added together using homomorphic addition to generate the updated standard reference vector.

[0158] The update process can be: Vn = (1-k)×V + k×Un. Where V is the standard reference vector for the current period, Un is the update statistics for the current period, and k is the preset learning rate. Since homomorphic encryption algorithms inherently support scalar multiplication between ciphertext and plaintext, as well as addition between ciphertexts, the above iterative update process can be efficiently completed in the fully ciphertext state, ensuring that the evaluation benchmark can continuously follow the evolution trend of healthy data.

[0159] For example, suppose the current standard reference vector is V = [E20, E400], and the updated weight k is 0.1. Calculate E20.1 by multiplying E21 by 0.1 and E400 by 0.9, and E400.5 by multiplying E405 by 0.1 and E400 by 0.9. The final updated standard reference vector is Vn = [E20.1, E400.5]. Vn will then be used to evaluate the next batch of data.

[0160] This application addresses the privacy risks and compliance challenges arising from plaintext data transmission and aggregation in cross-institutional collaboration scenarios by obtaining encrypted samples and tags at the source. It resolves the data structure heterogeneity issue between encrypted environments and statistical digests. It enables the reconstruction and extraction of encrypted feature descriptors reflecting data distribution characteristics without decryption. It ensures accurate assessment of the deviation of the current batch of data from a benign distribution even when the data is not visible. While guaranteeing data privacy and security, it also solves the problem of performance degradation in downstream AI models caused by the infiltration of low-quality data.

[0161] Figure 4 This is a schematic diagram of a specific implementation of the automated data cleaning and labeling quality assessment system for AI training provided in this application, with reference to... Figure 4 The system may include:

[0162] The acquisition module 21 is used to acquire an encrypted sample set for AI training from the untrusted domain and standard reference data from the trusted domain. The encrypted sample set includes multiple ciphertext samples and multiple ciphertext labels. The standard reference data includes statistical summary data of benign data corresponding to the encrypted sample set and the ciphertext labels.

[0163] The construction module 22 is used to construct a block index for the encrypted sample set based on the time information of each encrypted sample, to obtain multiple encrypted sample subsets, and to transform the standard reference data into a standard reference vector that matches the feature dimensions of the encrypted sample subsets.

[0164] The calculation module 23 is used to perform polynomial approximation calculation on each encrypted sample subset in the encrypted state to obtain the ciphertext feature information of each encrypted sample subset, and to perform dimension alignment on each ciphertext feature information using the standard reference vector as the standard to map the position of the encrypted sample subset in the statistical distribution space to obtain multiple ciphertext feature descriptors.

[0165] The calculation module 23 is also used to calculate the distribution offset value between each ciphertext feature descriptor and the standard reference vector;

[0166] The elimination module 24 is used to determine that the corresponding encrypted sample subset has a distribution anomaly and to eliminate it from the encrypted sample set when the distribution offset value is greater than a preset security threshold, so as to achieve quality assessment of data cleaning and labeling.

[0167] The automated data cleaning and labeling quality assessment system for AI training in this application is used to implement the aforementioned data cleaning and labeling quality assessment method for AI training. Therefore, the specific implementation of the automated data cleaning and labeling quality assessment system for AI training can be found in the embodiment section of the data cleaning and labeling quality assessment method for AI training above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0168] Figure 5 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0169] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described data cleaning and labeling quality assessment methods for AI training.

[0170] The electronic device may include a processor 510 and a memory 520 storing computer program instructions.

[0171] Specifically, the processor 510 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0172] Memory 520 may include mass storage for data or instructions. For example, and not limitingly, memory 520 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 520 may include removable or non-removable (or fixed) media. Where appropriate, memory 520 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 520 is non-volatile solid-state memory.

[0173] Memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to the first aspect of this disclosure.

[0174] The processor 510 reads and executes computer program instructions stored in the memory 520 to implement any data cleaning and labeling quality assessment method for AI training in the above embodiments.

[0175] In one example, the electronic device may also include a communication interface 530 and a bus 540. Wherein, such as Figure 5 As shown, the processor 510, memory 520, and communication interface 530 are connected through bus 540 and complete communication with each other.

[0176] The communication interface 530 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0177] Bus 540 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 540 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0178] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described data cleaning and labeling quality assessment methods for AI training.

[0179] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0180] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in the above embodiments of the data cleaning and labeling quality assessment method for AI training.

[0181] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0182] The above provides a detailed description of the data cleaning and labeling quality assessment method and system for AI training provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A quality assessment method for data cleaning and labeling for AI training, characterized in that, Applicable to privacy-preserving AI training sample evaluation, including: Acquire an encrypted sample set for AI training from an untrusted domain and standard reference data from a trusted domain. The encrypted sample set includes multiple ciphertext samples and multiple ciphertext labels. The standard reference data includes statistical summary data of benign data corresponding to the encrypted sample set and the ciphertext labels. Based on the time information of each ciphertext sample, the encrypted sample set is constructed by block indexing to obtain multiple encrypted sample subsets, and the standard reference data is transformed into a standard reference vector that matches the feature dimensions of the encrypted sample subsets. In the encrypted state, a polynomial approximation calculation is performed on each encrypted sample subset to obtain the ciphertext feature information of each encrypted sample subset. Then, using the standard reference vector as the standard, the dimension alignment of each ciphertext feature information is performed to map the position of the encrypted sample subset in the statistical distribution space, resulting in multiple ciphertext feature descriptors. Calculate the distribution offset value between each ciphertext feature descriptor and the standard reference vector; When the distribution offset value is greater than the preset security threshold, it is determined that the corresponding encrypted sample subset has a distribution anomaly and is removed from the encrypted sample set to achieve quality assessment of data cleaning and labeling.

2. The method according to claim 1, characterized in that, The method further includes: The subset of encrypted samples whose distribution offset value is less than or equal to a preset security threshold is considered a qualified subset of encrypted samples. Perform ciphertext summation on the ciphertext feature descriptors corresponding to all qualified encrypted sample subsets to obtain updated statistical information; Based on the updated statistical information, the standard reference vector is updated using a homomorphic encryption algorithm to obtain the updated standard reference vector.

3. The method according to claim 1, characterized in that, The method further includes: Obtain a feature importance vector derived from the trusted domain, the feature importance vector including weight values ​​corresponding to the feature dimensions of each ciphertext feature descriptor; Using a homomorphic encryption algorithm, the standard reference vector and the feature importance vector are multiplied in ciphertext to obtain a weighted standard reference vector; Each feature element in the ciphertext feature descriptor is multiplied by the corresponding weight value in the feature importance vector to obtain multiple weighted ciphertext feature descriptors. The step of calculating the distribution offset value between each ciphertext feature descriptor and the standard reference vector includes: Calculate the distribution offset value between each weighted ciphertext feature descriptor and the weighted standard reference vector.

4. The method according to claim 3, characterized in that, The step of calculating the distribution offset between each weighted ciphertext feature descriptor and the weighted standard reference vector includes: Using the subtraction property of homomorphic encryption, the difference vector between each weighted ciphertext feature descriptor and the weighted standard reference vector in the corresponding dimension is calculated. Utilizing the multiplication property of homomorphic encryption, the square of each vector element in the difference vector is calculated; By utilizing the addition property of homomorphic encryption, the summation of all squared terms in each difference vector is performed to obtain the distribution offset value of each encrypted sample subset relative to the standard reference data.

5. The method according to claim 1, characterized in that, The standard reference data includes the first-order and second-order raw moments corresponding to each feature dimension in the encrypted sample set and the ciphertext label; The process involves constructing a block index for the encrypted sample set based on the time information of each encrypted sample, resulting in multiple encrypted sample subsets. The standard reference data is then transformed into a standard reference vector that matches the feature dimensions of the encrypted sample subsets, including: The time span of the encrypted sample set is divided into multiple consecutive time intervals based on a preset time window length; The time interval to which each ciphertext sample belongs is determined based on the time information of each ciphertext sample, and a corresponding block index identifier is assigned to each ciphertext sample based on the time interval. Ciphertext samples with the same block index identifier are combined to obtain multiple encrypted sample subsets; According to the preset feature arrangement order, the first-order and second-order origin moments of each feature dimension in the standard reference data are concatenated to obtain a standard reference vector that matches the feature dimensions of the encrypted sample subset.

6. The method according to claim 1, characterized in that, In the encrypted state, a polynomial approximation calculation is performed on each encrypted sample subset to obtain the ciphertext feature information of each encrypted sample subset, including: In the encrypted state, each ciphertext sample and its corresponding ciphertext label in each encrypted sample subset are combined into a ciphertext vector to be processed, and the square value of each element in the ciphertext vector to be processed is calculated. The original values ​​of all ciphertext vectors to be processed within each encrypted sample subset are summed in ciphertext to obtain the first summed value. The squared values ​​of all ciphertext vectors to be processed within each encrypted sample subset are summed in ciphertext to obtain the second summed value. The first accumulated value and the second accumulated value are combined to obtain the ciphertext feature information of each encrypted sample subset.

7. The method according to claim 1, characterized in that, The step involves using the standard reference vector as a standard to perform dimensional alignment on each ciphertext feature information to map the position of the encrypted sample subset in the statistical distribution space, resulting in multiple ciphertext feature descriptors, including: According to the arrangement rules of the standard reference vector, the first and second accumulated values ​​in each ciphertext feature information are rearranged to obtain the rearranged ciphertext feature information; Obtain the number of samples in each encrypted sample subset, and generate the reciprocal factor of each encrypted sample subset based on the number of samples; Based on the reciprocal factor, the multiplication property of homomorphic encryption is used to perform mean processing on the sorted ciphertext feature information to obtain multiple statistical distribution coordinates. The statistical distribution coordinates are the positions of the corresponding encrypted sample subsets mapped to the statistical distribution space. Each statistical distribution coordinate is used as the ciphertext feature descriptor for the corresponding encrypted sample subset.

8. An automated data cleaning and labeling quality assessment system for AI training, characterized in that, include: The acquisition module is used to acquire an encrypted sample set for AI training from an untrusted domain and standard reference data from a trusted domain. The encrypted sample set includes multiple ciphertext samples and multiple ciphertext labels, and the standard reference data includes statistical summary data of benign data corresponding to the encrypted sample set and the ciphertext labels. The construction module is used to construct a block index for the encrypted sample set based on the time information of each encrypted sample, to obtain multiple encrypted sample subsets, and to transform the standard reference data into a standard reference vector that matches the feature dimensions of the encrypted sample subsets. The calculation module is used to perform polynomial approximation calculation on each encrypted sample subset in the encrypted state to obtain the ciphertext feature information of each encrypted sample subset, and to perform dimensional alignment on each ciphertext feature information using the standard reference vector as the standard to map the position of the encrypted sample subset in the statistical distribution space to obtain multiple ciphertext feature descriptors. The calculation module is also used to calculate the distribution offset value between each ciphertext feature descriptor and the standard reference vector; The elimination module is used to determine that the corresponding encrypted sample subset has a distribution anomaly and to eliminate it from the encrypted sample set when the distribution offset value is greater than a preset security threshold, so as to achieve quality assessment of data cleaning and labeling.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the data cleaning and labeling quality assessment method for AI training as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the data cleaning and labeling quality assessment method for AI training as described in any one of claims 1 to 7.

Images