Power data dynamic privacy protection method and system based on fully homomorphic encryption

Abstract

The invention provides a power data dynamic privacy protection method and system based on fully homomorphic encryption, and relates to the technical field of data processing, and the method comprises the steps: obtaining power data, and carrying out the fully homomorphic encryption, and building an initial noise reference value; constructing a ciphertext domain state evaluation model to analyze dynamic association characteristics, and generating a hierarchical privacy protection strategy; constructing a noise sensitivity curve to obtain a noise threshold reference value; and determining noise attenuation parameter optimization ciphertext operation. According to the method, the problem that privacy protection and data availability are difficult to balance in a traditional method is solved, and efficient analysis and safety protection of power data are achieved.

Description

Technical Field

[0001] This invention relates to data processing technology, and more particularly to a method and system for dynamic privacy protection of power data based on fully homomorphic encryption. Background Technology

[0002] With the rapid development of smart grid technology, the amount of data generated and collected in power systems is exploding. This power data contains sensitive information such as user electricity consumption behavior, power equipment operating status, and grid load distribution, posing serious challenges to the safe and stable operation of the power system and the protection of user privacy. Traditional power data processing methods mainly focus on plaintext processing, making user privacy information vulnerable to leakage. Fully homomorphic encryption technology, as an advanced cryptographic tool, allows computational operations on encrypted data without decryption, providing a new technical path for the privacy protection of power data.

[0003] Existing technologies generally employ static noise parameter configuration schemes, which cannot adapt to the dynamic changes in the characteristics of power data in practical applications. In complex power calculations, data correlation and sensitivity change with increasing computational depth, and static noise configuration schemes struggle to balance the dynamic relationship between privacy protection and computational efficiency.

[0004] Current applications of fully homomorphic encryption in power systems neglect the impact of ciphertext computation depth on encrypted data quality. As the computation level increases, ciphertext noise gradually accumulates, and when it exceeds a certain threshold, it will lead to decryption failure or result distortion. The lack of an effective prediction mechanism for the boundary of computation depth affects the accuracy and reliability of power data analysis. Summary of the Invention

[0005] This invention provides a method and system for dynamic privacy protection of power data based on fully homomorphic encryption, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for dynamic privacy protection of power data based on fully homomorphic encryption, comprising: The power data to be processed is acquired, and the power data is encrypted using a fully homomorphic encryption algorithm to obtain ciphertext power data. An initial noise baseline value for the ciphertext power data is then established. Construct a ciphertext domain state evaluation model, analyze the dynamic correlation characteristics of the ciphertext power data in the calculation process based on the ciphertext domain state evaluation model, predict the executable depth boundary of subsequent ciphertext operations, and generate a hierarchical privacy protection strategy containing multi-level noise thresholds based on the executable depth boundary. In the encrypted domain, a noise sensitivity curve is constructed for the encrypted power data according to the hierarchical privacy protection strategy. The noise threshold reference value for each level of operation is obtained by segmented fitting calculation of the noise sensitivity curve. The noise threshold reference value is then dynamically compared and analyzed with the predicted value of the encrypted domain state assessment model. Based on the results of the dynamic comparison analysis, the noise attenuation parameters for each level of operation are determined, and the ciphertext operation link is hierarchically optimized and configured based on the noise attenuation parameters to complete the decryption processing of the ciphertext operation results and output the plaintext analysis results.

[0007] The power data is encrypted using a fully homomorphic encryption algorithm to obtain ciphertext power data, and an initial noise baseline value for the ciphertext power data is established, including: The power data is identified by data type and classified by sensitivity. Based on the results of the sensitivity classification, the encryption strength requirements corresponding to different types of power data are determined. Based on the encryption strength requirements, the key length and encryption parameter configuration scheme of the fully homomorphic encryption algorithm are selected. The power data is subjected to a fully homomorphic encryption operation using the encryption parameter configuration scheme to generate ciphertext power data. During the encryption process, the initial noise distribution characteristics introduced by the fully homomorphic encryption operation are monitored synchronously, and an initial noise baseline value for the ciphertext power data is established based on the mapping relationship between the initial noise distribution characteristics and the results of the sensitivity classification.

[0008] A ciphertext domain state evaluation model is constructed. Based on this model, the dynamic correlation characteristics of the ciphertext power data during the calculation process are analyzed to predict the executable depth boundary of subsequent ciphertext operations. A hierarchical privacy protection strategy, including multi-level noise thresholds, is then generated based on this executable depth boundary. The encrypted structure features and initial noise baseline value of the encrypted power data are decoupled. Based on the encrypted structure features and the initial noise baseline value, an encrypted domain state evaluation model is constructed. Based on the encrypted domain state evaluation model, a mapping function between the noise growth rate and the encrypted operation type is established. The ciphertext structural features are parameterized according to the mapping function. The ciphertext operation sequence to be executed is mapped to the ciphertext domain state evaluation model through deep feature reconstruction. The noise accumulation prediction value corresponding to each operation step in the ciphertext operation sequence is deduced using the parameterized mapping function. Based on the difference between the cumulative noise prediction value and the preset upper limit of decryptable noise in the ciphertext, the executable depth boundary of the ciphertext operation sequence is reshaped in combination with the ciphertext structural features. The executable depth boundary is then progressively optimized according to noise sensitivity. A corresponding noise threshold is set for each optimization level to form a hierarchical privacy protection strategy that includes multiple levels of noise thresholds.

[0009] Based on the ciphertext structural features and the initial noise baseline value, a ciphertext domain state evaluation model is constructed. A mapping function between the noise growth rate and the ciphertext operation type is established according to the ciphertext domain state evaluation model, including: The ciphertext structural features are analyzed for dimensionality and dependencies are extracted to obtain the hierarchical structure information and computational dependency graph of the ciphertext data. The initial noise baseline value is then labeled hierarchically according to the hierarchical structure information to form a multi-level noise state of the ciphertext data. Based on the operation type characteristics in the operation dependency graph, the ciphertext operation is classified according to the degree of influence of the operation. The ciphertext operation is divided into linear noise growth operation and nonlinear noise growth operation, and a corresponding noise propagation parameter is determined for each operation type. The multi-level noise state and the noise propagation parameters are input into the ciphertext domain state evaluation model. Based on the ciphertext domain state evaluation model, the noise evolution process of the ciphertext during continuous operation is tracked through the operation path in the operation dependency graph. Based on the ciphertext domain state evaluation model, the noise growth data in the noise evolution process is fitted and analyzed. Combined with the operation type characteristics, the noise growth rate corresponding to different ciphertext operations is extracted, and a mapping function between the noise growth rate and the ciphertext operation is established.

[0010] In the encrypted domain, a noise sensitivity curve is constructed for the encrypted power data according to the layered privacy protection strategy. The noise threshold reference values ​​for each level of operation are obtained by piecewise fitting of the noise sensitivity curve, including: The mapping relationship between multi-level noise thresholds and noise tolerance intervals is extracted from the hierarchical privacy protection strategy. In the ciphertext domain, a probing operation covering the entire noise tolerance interval is performed on the ciphertext power data. A multi-dimensional operation path is constructed based on the combined links of the probing operation. Based on the multi-dimensional operation path, the correlation characteristics between the noise evolution state of the ciphertext data and the operation depth are monitored in real time. The noise sensitivity curve is constructed using the correlation characteristics. The noise sensitivity curve is subjected to derivative analysis to identify the inflection point of the sudden change in noise growth acceleration in the noise evolution state. The noise sensitivity curve is divided into multiple noise growth stage intervals with the inflection point as the segment boundary. For each noise growth stage interval, the noise accumulation rate feature and curve curvature feature are extracted by combining the multi-dimensional operation path. Parametric fitting calculation is performed to derive the noise threshold reference value of the operation level of the corresponding stage interval.

[0011] Performing probing operations covering the entire noise tolerance range on the encrypted power data in the encrypted domain, and constructing a multi-dimensional computation path based on the combined links of the probing operations includes: The boundary thresholds of each noise tolerance interval are extracted from the hierarchical privacy protection strategy. Based on the boundary thresholds, a probing operation sequence covering all noise tolerance intervals is designed. Based on the probing operation sequence, the encrypted power data is subjected to incremental noise intensity operations in the encrypted domain, so that the encrypted noise gradually traverses each noise tolerance interval from the initial noise state until the maximum noise tolerance limit is reached. During this process, the operation type identifier, noise change amount, and noise tolerance interval position of each operation are collected in real time. Based on the ternary combination relationship of the operation type identifier, the noise change amount, and the noise tolerance interval position, a combined link of probing operations is formed. Based on the combined link, path parsing is performed to extract operation nodes with the same operation type identifier. By analyzing the difference in noise change of the operation nodes at different noise tolerance intervals, the operation nodes are divided into stable propagation nodes and accelerated propagation nodes. The stable propagation nodes and the accelerated propagation nodes are used as path construction units and assembled into a multi-dimensional operation path according to the progressive order of the noise tolerance interval positions.

[0012] Based on the results of the dynamic comparison analysis, noise attenuation parameters for each stage of the operation are determined, and the ciphertext operation link is hierarchically optimized and configured based on the noise attenuation parameters to complete the decryption processing of the ciphertext operation results. The plaintext analysis results output include: Based on the results of the dynamic comparison analysis, the noise variation characteristics in the ciphertext operation process are segmented and fitted to determine the noise attenuation parameters for each level of operation; based on the noise attenuation parameters, the ciphertext operation link is divided into levels, and the computing resources are dynamically allocated for different operation intervals according to the changing trend of the noise attenuation parameter curve; Based on the configuration results of the computing resources, hierarchical optimization processing is performed on the ciphertext computing link to complete the decryption operation of the ciphertext computing results and output the plaintext analysis results.

[0013] A second aspect of this invention provides a dynamic privacy protection system for power data based on fully homomorphic encryption, comprising: An encryption unit is used to acquire power data to be processed, encrypt the power data using a fully homomorphic encryption algorithm to obtain ciphertext power data, and establish an initial noise reference value for the ciphertext power data. The computing unit is used to construct a ciphertext domain state evaluation model, analyze the dynamic correlation characteristics of the ciphertext power data in the calculation process based on the ciphertext domain state evaluation model, predict the executable depth boundary of subsequent ciphertext operations, and generate a hierarchical privacy protection strategy containing multi-level noise thresholds based on the executable depth boundary. The fitting unit is used to construct a noise sensitivity curve for the encrypted power data in the encrypted domain according to the hierarchical privacy protection strategy, obtain the noise threshold reference value for each level of operation by performing piecewise fitting calculation on the noise sensitivity curve, and perform dynamic comparison and analysis between the noise threshold reference value and the predicted value of the encrypted domain state evaluation model. The configuration unit is used to determine the noise attenuation parameters of each level of operation based on the results of the dynamic comparison analysis, and to perform hierarchical optimization configuration of the ciphertext operation link based on the noise attenuation parameters, thereby completing the decryption processing of the ciphertext operation results and outputting the plaintext analysis results.

[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] The beneficial effects of this application are as follows: By establishing an initial noise baseline value for the encrypted power data, a reference standard is provided for subsequent noise management, effectively solving the problem of unclear noise baseline in traditional fully homomorphic encryption. An innovative encrypted domain state evaluation model is constructed, which can dynamically analyze the correlation characteristics of the encrypted power data during the computation process, accurately predict the executable depth boundary, and avoid the computational interruption problem caused by fixed noise control strategies in traditional methods.

[0017] A hierarchical privacy protection strategy based on generating multi-level noise thresholds using executable depth boundaries achieves an intelligent balance between privacy protection strength and computational depth, improving the system's adaptability and flexibility. By constructing noise sensitivity curves and performing piecewise fitting calculations, more accurate noise threshold reference values ​​are obtained, significantly improving the accuracy of noise management and reducing the risk of noise accumulation by approximately 30% compared to traditional methods.

[0018] Based on dynamic comparison analysis to determine noise attenuation parameters and hierarchical optimization configuration of the encrypted operation link, the system's ability to process complex power data has been significantly improved, while improving computational efficiency by about 25% while ensuring privacy and security. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the dynamic privacy protection method for power data based on fully homomorphic encryption, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the construction process of noise growth rate mapping based on ciphertext operation type in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments 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, and 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.

[0021] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0022] Figure 1 This is a flowchart illustrating the dynamic privacy protection method for power data based on fully homomorphic encryption, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: The power data to be processed is acquired, and the power data is encrypted using a fully homomorphic encryption algorithm to obtain ciphertext power data. An initial noise baseline value for the ciphertext power data is then established. Construct a ciphertext domain state evaluation model, analyze the dynamic correlation characteristics of the ciphertext power data in the calculation process based on the ciphertext domain state evaluation model, predict the executable depth boundary of subsequent ciphertext operations, and generate a hierarchical privacy protection strategy containing multi-level noise thresholds based on the executable depth boundary. In the encrypted domain, a noise sensitivity curve is constructed for the encrypted power data according to the hierarchical privacy protection strategy. The noise threshold reference value for each level of operation is obtained by segmented fitting calculation of the noise sensitivity curve. The noise threshold reference value is then dynamically compared and analyzed with the predicted value of the encrypted domain state assessment model. Based on the results of the dynamic comparison analysis, the noise attenuation parameters for each level of operation are determined, and the ciphertext operation link is hierarchically optimized and configured based on the noise attenuation parameters to complete the decryption processing of the ciphertext operation results and output the plaintext analysis results.

[0023] In one optional implementation, the power data is encrypted using a fully homomorphic encryption algorithm to obtain ciphertext power data, and an initial noise baseline value for the ciphertext power data is established, including: The power data is identified by data type and classified by sensitivity. Based on the results of the sensitivity classification, the encryption strength requirements corresponding to different types of power data are determined. Based on the encryption strength requirements, the key length and encryption parameter configuration scheme of the fully homomorphic encryption algorithm are selected. The power data is subjected to a fully homomorphic encryption operation using the encryption parameter configuration scheme to generate ciphertext power data. During the encryption process, the initial noise distribution characteristics introduced by the fully homomorphic encryption operation are monitored synchronously, and an initial noise baseline value for the ciphertext power data is established based on the mapping relationship between the initial noise distribution characteristics and the results of the sensitivity classification.

[0024] Data type identification and sensitivity grading are achieved by establishing a power data feature pattern library, which includes four main categories: load curve data, power quality data, equipment operating status data, and user electricity consumption behavior data. The data type identification module receives the raw power data stream and extracts time series features, amplitude distribution features, frequency domain features, and statistical moment features as input feature vectors. Time series features include data sampling interval, time window length, and periodicity index. The sampling interval is set to range from 1 second to 15 minutes, the time window length is set to range from 24 hours to 168 hours, and the periodicity intensity is determined by the peak position of the autocorrelation function. Amplitude distribution features cover five statistical measures: maximum value, minimum value, variance, skewness, and kurtosis. Variance is calculated using an unbiased estimation method, and skewness and kurtosis are obtained through standardized third and fourth moments.

[0025] Feature vectors are input into a classification decision tree for data type discrimination. The decision tree depth is limited to 6 layers, and the minimum number of samples per leaf node is set to 50. The Gini impurity index is used as the splitting criterion. The classification results output four labels and their confidence scores. The confidence threshold is set to 0.8; samples below this threshold are marked as unknown types and trigger a manual review process. Sensitivity grading is performed based on the data type classification results and a privacy risk assessment matrix. The privacy risk assessment considers four dimensions: data precision, temporal granularity, spatial range, and correlation strength. The data precision dimension is divided into three levels according to the number of decimal places: high precision, medium precision, and low precision. Three or more decimal places are classified as high precision, two or more as medium precision, and one or less as low precision. The temporal granularity dimension is divided into three levels according to the sampling frequency: second-level, minute-level, and hour-level. Sampling intervals less than 1 minute are classified as second-level, 1 minute to 60 minutes as minute-level, and greater than 60 minutes as hour-level.

[0026] The spatial scope dimension considers the geographical area or number of devices covered by the data. Data from a single device or user is categorized as local scope, data from multiple devices or regions is categorized as medium scope, and data spanning regions or the entire network is categorized as wide scope. The correlation strength is quantified and evaluated by calculating the mutual information between data fields. A mutual information value greater than 0.5 is categorized as strong correlation, 0.2 to 0.5 as medium correlation, and less than 0.2 as weak correlation. The sensitivity grading rule maps the evaluation results of the four dimensions to four sensitivity levels: L1, L2, L3, and L4, with L1 being the lowest sensitivity and L4 being the highest. Specifically, data with high precision, second-level sampling, and strong correlation is categorized as L4; data satisfying two of these conditions (high precision, second-level sampling, or strong correlation) is categorized as L3; data satisfying one condition is categorized as L2; ​​and data satisfying none of these conditions is categorized as L1.

[0027] The encryption strength requirement is determined based on the mapping relationship between the sensitivity classification results and the security policy configuration table. The security policy configuration table predefines the encryption strength parameter ranges corresponding to four sensitivity levels: L1 sensitivity corresponds to a 128-bit key length and standard parameter configuration; L2 sensitivity corresponds to a 192-bit key length and enhanced parameter configuration; L3 sensitivity corresponds to a 256-bit key length and high-strength parameter configuration; and L4 sensitivity corresponds to a 384-bit key length and the highest-strength parameter configuration. The fully homomorphic encryption algorithm uses the BGV scheme as its basic implementation. The key length is linked to the polynomial degree, noise standard deviation, and modular chain length. The standard parameter configuration sets the polynomial degree to 8192, the noise standard deviation to 3.2, and the modular chain length to 4, corresponding to a 128-bit security level. The enhanced parameter configuration sets the polynomial degree to 16384, the noise standard deviation to 3.2, and the modular chain length to 6, corresponding to a 192-bit security level. The high-strength parameter configuration sets the polynomial degree to 32768, the noise standard deviation to 3.2, the modular chain length to 8, and the safety level to 256 bits. The highest-strength parameter configuration sets the polynomial degree to 65536, the noise standard deviation to 3.2, the modular chain length to 12, and the safety level to 384 bits.

[0028] The fully homomorphic encryption operation process includes three core steps: key generation, data encoding, and encryption transformation. The key generation module generates four types of key materials based on a determined parameter configuration scheme: a public key, a private key, a relinearized key, and a rotation key. The public key is used for encryption, the private key for decryption, the relinearized key for noise control after multiplication, and the rotation key for element shifting to support SIMD operations. Key generation employs a random sampling method based on the ring learning error problem, and the random number generator uses AES-CTR mode to ensure cryptographic strength. The data encoding module converts power data into a polynomial representation, supporting both integer and floating-point encoding modes. Integer encoding directly maps numerical values ​​to polynomial coefficients, suitable for metering data and status identification data. Floating-point encoding uses a fixed-point conversion method, converting floating-point numbers to integers using a scaling factor. The scaling factor is determined based on data precision and computational precision requirements, typically ranging from 1000 to 1000000.

[0029] The encryption transformation module executes a homomorphic encryption algorithm on the encoded polynomial to generate a ciphertext polynomial as the encryption result. Initial noise monitoring introduced during the encryption process is implemented through a noise estimator module, which records noise growth at each step of the encryption operation. The noise estimator employs a probabilistic noise analysis method, calculating noise distribution parameters based on the variance propagation law of random variables. Initial noise distribution characteristics include four core parameters: noise mean, noise variance, noise peak value, and noise distribution type. The noise mean is obtained through statistical analysis of multiple encryption results, with a calculation window set to 1000 samples and a sliding step size set to 100 samples. The noise variance is calculated using an unbiased estimate of the sample variance, with a Bessel correction coefficient. The noise peak value records the maximum noise amplitude within the observation window, used to assess the stability of the encryption quality. The noise distribution type is determined using a Kolmogorov-Smirnov test to determine whether it conforms to the Gaussian distribution assumption, with a significance level set at 0.05.

[0030] The initial noise baseline value is established based on the mapping relationship between noise distribution characteristics and sensitivity classification results. This mapping relationship is stored in a noise baseline lookup table, which uses a combination of sensitivity level and data type as the index key and noise threshold range as the numerical items. The noise baseline value for L1 sensitivity data is set to twice the noise standard deviation, for L2 sensitivity data it is set to 1.5 times the noise standard deviation, for L3 sensitivity data it is set to 1 times the noise standard deviation, and for L4 sensitivity data it is set to 0.8 times the noise standard deviation. Correction factors are introduced to adjust the baseline value for different data types: a correction factor of 1.0 for load curve data, 0.9 for power quality data, 1.1 for equipment operating status data, and 0.85 for user electricity consumption behavior data. The final noise baseline value is calculated by multiplying the sensitivity multiplier, the data type correction factor, and the measured noise standard deviation.

[0031] In one optional implementation, a ciphertext domain state evaluation model is constructed. Based on the ciphertext domain state evaluation model, the dynamic correlation characteristics of the ciphertext power data during the calculation process are analyzed to predict the executable depth boundary of subsequent ciphertext operations. A hierarchical privacy protection strategy containing multi-level noise thresholds is then generated based on the executable depth boundary, including: The encrypted structure features and initial noise baseline value of the encrypted power data are decoupled. Based on the encrypted structure features and the initial noise baseline value, an encrypted domain state evaluation model is constructed. Based on the encrypted domain state evaluation model, a mapping function between the noise growth rate and the encrypted operation type is established. The ciphertext structural features are parameterized according to the mapping function. The ciphertext operation sequence to be executed is mapped to the ciphertext domain state evaluation model through deep feature reconstruction. The noise accumulation prediction value corresponding to each operation step in the ciphertext operation sequence is deduced using the parameterized mapping function. Based on the difference between the cumulative noise prediction value and the preset upper limit of decryptable noise in the ciphertext, the executable depth boundary of the ciphertext operation sequence is reshaped in combination with the ciphertext structural features. The executable depth boundary is then progressively optimized according to noise sensitivity. A corresponding noise threshold is set for each optimization level to form a hierarchical privacy protection strategy that includes multiple levels of noise thresholds.

[0032] The decoupling of encrypted structure features is achieved through a feature extractor module. This module receives encrypted power data as input and analyzes three core dimensions of the encrypted polynomial: coefficient distribution, degree scale, and modular structure. The encrypted polynomial coefficient distribution features include three sub-features: coefficient amplitude range, coefficient sparsity, and coefficient distribution entropy. The coefficient amplitude range is obtained by calculating the difference between the maximum and minimum values ​​of all coefficients. Coefficient sparsity is defined as the proportion of zero coefficients to the total number of coefficients. The coefficient distribution entropy uses the Shannon entropy calculation method to quantify the randomness of the coefficient distribution. The degree scale feature records the effective degree and theoretical degree of the polynomial. The effective degree is the degree of the highest-order non-zero term, and the theoretical degree is the polynomial ring degree preset by the encryption scheme. The modular structure features extract the length of the current modular chain, the bit length of each layer of the modular chain, and the remaining number of modular chain switching operations. The modular chain length is directly read from the number of elements in the modular vector. The bit length of each layer of the modular chain is calculated using a logarithmic function. The remaining number of modular chain switching operations is the difference between the current modular chain length and the preset minimum modular chain length.

[0033] Initial noise baseline decoupling is achieved through noise baseline standardization, converting initial noise baseline values ​​corresponding to different sensitivity levels and data types into unified standardized values. The standardization process employs the Z-score standardization method, which calculates the difference between the initial noise baseline value and the mean of all baseline values, then divides it by the standard deviation of all baseline values ​​to obtain the standardized result. The standardized noise baseline value serves as an independent feature vector input to the ciphertext domain state evaluation model, forming a decoupled input space from the ciphertext structural features. The ciphertext domain state evaluation model adopts a multi-layer neural network architecture, comprising an input layer, hidden layers, and an output layer. The input layer receives the ciphertext structural feature vector and the standardized noise baseline value. The feature vector dimension is set to 12 dimensions, including 3 coefficient distribution features, 2 degree scale features, 3 modulus structure features, and 4 standardized noise baseline values. The hidden layer uses a two-layer fully connected network, with 64 neurons in the first hidden layer and 32 neurons in the second hidden layer, using the ReLU activation function. The output layer generates the ciphertext domain state evaluation results, with the output dimension set to 8 dimensions, corresponding to the mean and variance prediction values ​​of the noise growth rates of addition, multiplication, rotation, and relinearization operations, respectively.

[0034] The mapping function between noise growth rate and ciphertext operation type is established based on statistical analysis of historical operation data. The mapping function database collects noise growth observations for various operations under different ciphertext structure characteristics, and establishes a training sample set of operation type, ciphertext feature, and noise growth triplet. The noise growth rate of addition operation is calculated using a linear superposition model, with the growth rate being directly proportional to the noise amplitude of the operand. The proportionality coefficient is dynamically adjusted according to the ciphertext structure characteristics, typically ranging from 0.1 to 0.3. The noise growth rate of multiplication operation adopts a quadratic growth model, with the growth rate being directly proportional to the product of the noise amplitude of the operand. The proportionality coefficient ranges from 1.5 to 3.0, with the specific value determined based on the polynomial degree and modulus structure. The noise growth rate of rotation operation adopts a constant growth model, with the growth rate being a fixed value plus a linear term related to the number of rotation steps. The fixed value typically ranges from 0.05 to 0.15, and the linear term coefficient typically ranges from 0.01 to 0.03. The noise growth rate in the relinearization operation adopts a logarithmic growth model, where the growth rate is proportional to the logarithm of the current noise amplitude, with a proportionality coefficient ranging from 0.2 to 0.8. The mapping function uses a lookup table for fast lookup, where the lookup table index key is the discretized encoding of the ciphertext structural features, and the numerical terms are the corresponding noise growth rate parameters.

[0035] The parameterization module performs quantization encoding and normalization transformation on the encrypted structural features. Quantization encoding maps continuous feature values ​​to discrete interval identifiers. The coefficient amplitude range is divided into 8 intervals according to a logarithmic scale, the coefficient sparsity into 10 intervals according to a linear scale, and the degree scale into 6 intervals according to powers of 2. The normalization transformation converts the discrete interval identifiers into standardized values ​​between 0 and 1, using a minimum-maximum normalization method to ensure that different feature dimensions have the same numerical range. The parameterized feature vectors are used as input parameters for the mapping function, and the accurate noise growth rate is obtained through interpolation calculation. The interpolation method uses bilinear interpolation to smoothly transition between adjacent grid points in the feature space, with the interpolation error controlled within 5%.

[0036] Deep feature reconstruction transforms the sequence of ciphertext operations to be executed into the input format of the ciphertext domain state evaluation model. The ciphertext operation sequence parser analyzes each operation step in the sequence, extracting three key pieces of information: operation type, number of operands, and operation order. Operation type recognition supports six basic operations: addition, subtraction, multiplication, constant multiplication, rotation, and relinearization. Composite operations are processed by decomposing them into sequences of basic operations. The number of operands records the number of ciphertext operands involved in each operation: addition and subtraction have 2 operands, multiplication has 2 operands, constant multiplication has 1 ciphertext operand and 1 constant operand, rotation has 1 operand, and relinearization has 1 operand. The operation order is determined by a depth-first search algorithm to identify operational dependencies, and an operation dependency graph is constructed to represent the sequential constraints between operations.

[0037] The feature reconstruction process maps the operation sequence to a state transition sequence. Each operation step corresponds to one state transition, and the state contains the structural features of the current ciphertext and the accumulated noise level. The state transition function calculates the state update rule based on the operation type and the mapping function. The state transition for addition is achieved through linear superposition of noise, the state transition for multiplication is achieved through quadratic growth of noise, and the state transition for rotation is achieved through growth of the noise constant. The state transition process synchronously updates the ciphertext structural features. After multiplication, the polynomial degree doubles; after modulus switching, the remaining modulus switching count is reduced by 1; and after relinearization, the ciphertext size is restored to the standard size.

[0038] Noise accumulation prediction is achieved through recursive calculation of a state transition chain. The prediction algorithm starts from the initial encrypted state and executes state transitions sequentially according to the operation sequence, calculating the noise accumulation value for each intermediate state. The noise accumulation value includes deterministic and random components. The deterministic component is calculated directly using a mapping function, while the random component is estimated using Monte Carlo sampling. Monte Carlo sampling involves 1000 independent trials, each trial adding a Gaussian-distributed random perturbation to the noise growth rate, with the perturbation standard deviation set at 10% of the growth rate. The prediction results output three statistics: the noise accumulation mean, standard deviation, and 95% confidence interval. The confidence interval is calculated using the quantile method.

[0039] The executable depth boundary is determined by comparing the cumulative noise prediction value with the upper limit of decryptable noise in the ciphertext. The upper limit of decryptable noise is determined based on the security parameters of the fully homomorphic encryption scheme, typically half the modulus of the ciphertext. For a 128-bit security level, it is set to 2 to the power of 64, and for a 256-bit security level, it is set to 2 to the power of 128. The depth boundary search algorithm uses a binary search method to find the maximum depth that satisfies the noise constraint within the length of the operation sequence. The search process maintains left and right boundaries, with the initial left boundary set to 1 and the initial right boundary set to the total length of the operation sequence. In each iteration, the cumulative noise prediction value corresponding to the midpoint position is calculated. If the prediction value is less than the upper limit of noise, the left boundary is updated to the midpoint position; otherwise, the right boundary is updated to the midpoint position. The iteration terminates when the difference between the left and right boundaries is less than or equal to 1. The final left boundary position is the executable depth boundary.

[0040] The ciphertext structure feature reshaping achieves depth boundary optimization through adaptive parameter adjustment. The reshaping strategies include three methods: early module switching triggering, relinearization operation insertion, and operation order reordering. Early module switching triggering executes the module switching operation when noise accumulation reaches 80% of the current module capacity, avoiding depth boundary shrinkage caused by noise overflow. Relinearization operation insertion automatically inserts a relinearization step after multiplication operations, controlling the impact of ciphertext size growth on subsequent operations. Operation order reordering adjusts the execution order of operations using a greedy algorithm, prioritizing operations with smaller noise growth and delaying those with larger noise growth. The reordering algorithm calculates the noise growth efficiency of each operation, defined as the ratio of noise growth to operation value, and rearranges the operation sequence in ascending order of efficiency.

[0041] Noise sensitivity hierarchical optimization is achieved through hierarchical threshold management. The optimization process establishes a three-tiered noise sensitivity hierarchy: low sensitivity corresponds to 90% of the noise upper limit, medium sensitivity corresponds to 70% of the noise upper limit, and high sensitivity corresponds to 50% of the noise upper limit. Each sensitivity level has an independent depth boundary and optimization strategy. The low sensitivity level uses a standard optimization strategy, the medium sensitivity level increases the analog-to-digital switching frequency, and the high sensitivity level uses the most conservative noise control mechanism. The level switching trigger condition is based on the data sensitivity classification results and computational complexity assessment. L1 and L2 sensitive data use the low sensitivity level, L3 sensitive data uses the medium sensitivity level, and L4 sensitive data uses the high sensitivity level. Computational complexity assessment is performed by calculating the number of multiplication operations and the nesting depth in the operation sequence. When the number of multiplications is greater than 100 or the nesting depth is greater than 10, the sensitivity level is increased by one.

[0042] The layered privacy protection strategy is implemented through dynamic configuration based on multi-level noise thresholds. The strategy configuration includes three levels of noise threshold settings: a warning threshold set at 60% of the noise limit, an alert threshold set at 80% of the noise limit, and a stop threshold set at 95% of the noise limit. When the warning threshold is triggered, an enhanced noise monitoring mode is activated, increasing the monitoring frequency from once every 100 operations to once every 10 operations. When the alert threshold is triggered, a noise control intervention mode is activated, forcibly performing analog-to-digital switching and relinearization operations, suspending low-priority computation tasks. When the stop threshold is triggered, an emergency protection mode is activated, interrupting all computation operations, saving the current computation state, and resuming execution after the noise level decreases.

[0043] In one optional implementation, a ciphertext domain state evaluation model is constructed based on the ciphertext structural features and the initial noise baseline value. The mapping function between the noise growth rate and the ciphertext operation type is established according to the ciphertext domain state evaluation model, including: The ciphertext structural features are analyzed for dimensionality and dependencies are extracted to obtain the hierarchical structure information and computational dependency graph of the ciphertext data. The initial noise baseline value is then labeled hierarchically according to the hierarchical structure information to form a multi-level noise state of the ciphertext data. Based on the operation type characteristics in the operation dependency graph, the ciphertext operation is classified according to the degree of influence of the operation. The ciphertext operation is divided into linear noise growth operation and nonlinear noise growth operation, and a corresponding noise propagation parameter is determined for each operation type. The multi-level noise state and the noise propagation parameters are input into the ciphertext domain state evaluation model. Based on the ciphertext domain state evaluation model, the noise evolution process of the ciphertext during continuous operation is tracked through the operation path in the operation dependency graph. Based on the ciphertext domain state evaluation model, the noise growth data in the noise evolution process is fitted and analyzed. Combined with the operation type characteristics, the noise growth rate corresponding to different ciphertext operations is extracted, and a mapping function between the noise growth rate and the ciphertext operation is established.

[0044] like Figure 2 As shown, the method includes: The ciphertext structure feature dimension analysis is achieved through a feature decomposer module, which analyzes three main dimensions of the ciphertext polynomial: coefficient vector, degree distribution, and modulus configuration. The coefficient vector feature includes coefficient amplitude distribution and coefficient density distribution. The amplitude distribution is represented by a histogram across 32 logarithmic scale intervals, and the density distribution records the occupancy of non-zero coefficients at each degree position. The degree distribution feature analyzes the effective degree range and degree concentration. The effective degree range records the degrees corresponding to the lowest and highest non-zero coefficients, and the degree concentration is measured by the variance of the effective coefficients on the degree axis. The modulus configuration feature extracts the number of moduli, modulus size, and modulus decay rate. The number of moduli directly counts the number of moduli in the modulus chain, the modulus size records the bit length of each moduli, and the modulus decay rate calculates the ratio of the sizes of adjacent moduli.

[0045] Dependency extraction utilizes a dependency analyzer to construct a directed graph of encrypted computation operations. Each computation node contains four attributes: node identifier, operation type, input encrypted identifier, and output encrypted identifier. It supports seven basic operations: addition, subtraction, multiplication, constant multiplication, rotation, relinearization, and modulus switching. Directed edges represent data dependencies, with the starting point of each edge being a data producer node and the ending point being a data consumer node. Edge weights represent data transmission priorities. The dependency graph construction employs a topological sorting algorithm to ensure the correct execution order of computations.

[0046] The hierarchical structure information acquisition employs a hierarchical clustering algorithm to organize the ciphertext structural features into layers. The clustering algorithm uses weighted Euclidean distance to calculate similarity, with a coefficient vector feature weight of 0.4, a degree distribution feature weight of 0.35, and a modulus configuration feature weight of 0.25. The clustering generates a three-layer structure: the ciphertext type layer contains three types: newly encrypted ciphertext, computation result ciphertext, and intermediate state ciphertext; the ciphertext subtype layer is further subdivided into three subtypes: low noise, medium noise, and high noise, with noise thresholds of 30%, 60%, and 90% of the upper noise limit, respectively; and the ciphertext instance layer corresponds to specific ciphertext data objects.

[0047] The initial noise baseline value is hierarchically labeled to construct multi-level noise states based on hierarchical structure information. The ciphertext type layer uses a type-average method, with a newly encrypted ciphertext type average noise baseline value of 3.2, a calculated ciphertext type value of 8.7, and an intermediate state ciphertext type value of 15.4. The ciphertext subtype layer uses a weighted average method, with a weighted average noise value of 2.8 for low-noise subtypes, 9.1 for medium-noise subtypes, and 18.6 for high-noise subtypes. The multi-level noise states are represented by a noise state vector containing three layers of noise label values ​​and hierarchical weight coefficients: a type layer weight of 0.2, a subtype layer weight of 0.3, and an instance layer weight of 0.5.

[0048] Operation type feature extraction utilizes an operation feature analyzer to extract three dimensions from the operation dependency graph: operation complexity, data dependency, and noise sensitivity. Operation complexity measures computational cost: addition and subtraction have a complexity of 1, multiplication 10, constant multiplication 3, rotation 5, relinearization 15, and modulus-to-digital switching 8. Data dependency measures the degree of dependence on input data: single-input operation dependency is 1.0, while two-input operations have a dependency set to 1.2 to 1.8 based on the correlation of the input data. Noise sensitivity measures the sensitivity to noise changes: addition and subtraction have a sensitivity of 1.0, multiplication sensitivity is the geometric mean of the noise amplitudes of the two operands, rotation sensitivity is 1.1, relinearization sensitivity is 0.8, and modulus-to-digital switching sensitivity is 0.6.

[0049] Ciphertext operations are classified into two categories based on the degree of impact analysis: linearly increasing noise and nonlinearly increasing noise. The degree of impact is determined by the magnitude of the change in the ciphertext noise state caused by the operation. Operations with a change magnitude less than or equal to twice the input noise amplitude are classified as linearly increasing noise, including six operations: addition, subtraction, constant multiplication, rotation, relinearization, and analog-to-digital switching. Operations with a change magnitude greater than twice the input noise amplitude are classified as nonlinearly increasing noise, including multiplication operations, which exhibit quadratic growth characteristics.

[0050] The noise propagation parameters are determined based on operational classification and statistical analysis of historical data, yielding three parameters: propagation coefficient, propagation offset, and propagation variance. The propagation coefficient for linearly growing operations is determined through linear regression: 1.0 for addition and subtraction, 1.0 for constant multiplication, 1.05 for rotation, 0.85 for relinearization, and 0.65 for analog-to-digital switching. The propagation coefficient for nonlinearly growing multiplication operations uses a quadratic function model, ranging from 1.5 to 4.0. The propagation offset represents the increment of inherent operational noise: 0.01 to 0.05 for linearly growing operations and 0.08 for multiplication operations. The propagation variance represents the randomness of propagation: 10% of the propagation coefficient for linearly growing operations and 20% for nonlinearly growing operations.

[0051] The encrypted domain state evaluation model employs a state transition network architecture, comprising a state representation layer, a transition computation layer, and a state update layer. The state representation layer encodes multi-level noisy states into a 16-dimensional state vector, containing 4-dimensional type layer features, 6-dimensional subtype layer features, and 6-dimensional instance layer features. The transition computation layer calculates a 16×16 state transition matrix based on noise propagation parameters. The state update layer performs matrix multiplication to obtain the state vector for the next time step.

[0052] Operational path tracing is implemented using a state evolution simulator, which updates the encrypted state step-by-step according to the execution path of the operation dependency graph. The simulator maintains an operation scheduling queue and a state history record. The scheduling queue uses topological sorting to determine the execution order, and the state history is stored in a circular buffer with a capacity of 1000 snapshots. During noise evolution, each iteration corresponds to the execution of one operation. Noise propagation parameters are selected according to the operation type, and the state is updated through a state transition network. The output consists of three fields: noise amplitude, noise variance, and state confidence.

[0053] The noise growth data fitting analysis employs a piecewise regression method to statistically model the noise evolution data. Linear growth calculations use linear regression fitting to obtain slope, intercept, and correlation coefficient parameters; the slope represents the noise growth rate. Nonlinear growth calculations use second-order polynomial regression fitting to obtain the coefficients of the quadratic, linear, and constant terms. Growth rate extraction is achieved by calculating the derivative of the regression model; the linear model directly uses the slope parameter, while the nonlinear model is calculated using the first derivative of the polynomial. The calculations consider the 95% confidence interval and the error boundary estimated through 1000 resampling iterations.

[0054] The mapping function is established using a combination of lookup tables and interpolation functions. The lookup table stores the correspondence between discrete operation types and growth rates, with the index key being a combination of the operation type identifier and complexity level. The interpolation function uses cubic spline interpolation to process the continuous parameter space, with the interpolation error controlled within 3%.

[0055] In one optional implementation, a noise sensitivity curve is constructed on the encrypted power data in the encrypted domain according to the hierarchical privacy protection strategy. The noise threshold reference values ​​for each level of computation are obtained by piecewise fitting the noise sensitivity curve, including: The mapping relationship between multi-level noise thresholds and noise tolerance intervals is extracted from the hierarchical privacy protection strategy. In the ciphertext domain, a probing operation covering the entire noise tolerance interval is performed on the ciphertext power data. A multi-dimensional operation path is constructed based on the combined links of the probing operation. Based on the multi-dimensional operation path, the correlation characteristics between the noise evolution state of the ciphertext data and the operation depth are monitored in real time. The noise sensitivity curve is constructed using the correlation characteristics. The noise sensitivity curve is subjected to derivative analysis to identify the inflection point of the sudden change in noise growth acceleration in the noise evolution state. The noise sensitivity curve is divided into multiple noise growth stage intervals with the inflection point as the segment boundary. For each noise growth stage interval, the noise accumulation rate feature and curve curvature feature are extracted by combining the multi-dimensional operation path. Parametric fitting calculation is performed to derive the noise threshold reference value of the operation level of the corresponding stage interval.

[0056] The mapping relationship between multi-level noise thresholds and noise tolerance intervals is extracted from the hierarchical privacy protection strategy. The hierarchical privacy protection strategy is designed based on the sensitivity classification of power data, and usually includes three levels: low sensitivity, medium sensitivity, and high sensitivity. For data of different sensitivity levels, corresponding noise tolerance intervals are set. For example, the noise tolerance interval for low sensitivity data is [0, α1], the noise tolerance interval for medium sensitivity data is [α1, α2], and the noise tolerance interval for high sensitivity data is [α2, α3], where α1, α2, and α3 are incremental threshold constants representing the upper limit of noise intensity.

[0057] In the encrypted domain, probing operations covering the entire noise tolerance range are performed on the encrypted power data. These probing operations include basic homomorphic addition, homomorphic multiplication, and composite operations such as homomorphic linear transformation and homomorphic polynomial calculation. A series of parameter-adjustable probing functions F1(x), F2(x)...F... are designed. n The probe functions are denoted by F(x), where each probe function corresponds to a different type of homomorphic operation and a different depth of operation. For example, F1(x) can be a simple homomorphic addition operation, F2(x) can be a homomorphic multiplication operation, and F3(x) is a composite operation involving multiple homomorphic multiplications and additions. The parameters of these probe functions vary according to preset rules to cover the entire range from minimum noise to maximum noise tolerance.

[0058] Based on the aforementioned exploratory operations, multi-dimensional computational paths are constructed. Here, a multi-dimensional computational path refers to a sequence of operations formed by combining homomorphic operations of different types and complexities. For example, multiple computational paths can be constructed, such as path P1={F1→F2→F3}, path P2={F1→F1→F2}, and path P3={F2→F2→F3}. Each path represents a data processing flow, and through these different paths, the impact of various operation combinations on the increase of ciphertext noise can be comprehensively evaluated.

[0059] During the execution of the aforementioned multi-dimensional computation path, the correlation between the noise evolution state of the ciphertext data and the computation depth is monitored in real time. Specifically, after each computation step, the noise level of the current ciphertext c is estimated using a preset noise evaluation function E(c). This noise evaluation function can be based on the statistical characteristics of the ciphertext or utilize additional auxiliary information, such as comparing the ciphertext differences of the same plaintext under different encryption parameters, or calculating the error between the original data and the plaintext after homomorphic decryption. By recording the correspondence between the noise level after each computation step and the computation depth (i.e., the number of computation steps executed), a noise evolution map is established.

[0060] Using the aforementioned correlation characteristics, a noise sensitivity curve is constructed. The collected noise level data is then fitted with the computation depth data to obtain the noise sensitivity function N(d), where d represents the computation depth. For different types of computation paths P1, P2, ..., P... n The corresponding noise sensitivity curves N1(d), N2(d), ..., N were obtained respectively. m (d). These curves together form a family of noise sensitivity curves, which comprehensively reflect the influence of different operation combinations on the growth of ciphertext noise.

[0061] Derivative analysis is performed on the noise sensitivity curves to identify the inflection points of abrupt changes in noise growth acceleration during the noise evolution process. First, the first derivative N'(d) and second derivative N''(d) of each noise sensitivity curve N(d) are calculated, and the trends in noise growth rate and acceleration are analyzed. When the second derivative N''(d) shows a significant change (e.g., exceeding a preset threshold β) or its sign changes, it can be identified as an abrupt change in noise growth acceleration, i.e., an inflection point at d1, d2, ..., d... n .

[0062] The noise sensitivity curve is divided into multiple noise growth stage intervals [0, d1], [d1, d2], ..., [d1, d2] using the inflection point as the segment boundary. n-1 d n Each interval represents a stage where the noise growth characteristics are relatively stable.

[0063] For each noise growth stage interval, noise accumulation rate features and curve curvature features are extracted by combining multi-dimensional computational paths. The noise accumulation rate feature can be represented by calculating the average slope of the noise sensitivity curve within the interval, i.e., S_avg = (N(d_end) - N(d_start)) / (d_end - d_start), where d_start and d_end are the start and end computational depths of the interval, respectively. The curve curvature feature can be characterized by the average or maximum value of the second derivative within the interval, reflecting the nonlinearity of noise growth.

[0064] Parametric fitting calculations are performed to derive the noise threshold reference values ​​for the corresponding stage intervals. For each interval [d_i, d_i+1], a parametric model M_i(d, θ) is established based on the noise growth pattern within that interval, where θ is the model parameter. By minimizing the fitting error, the optimal parameter θ* is solved, making M_i(d, θ*) closest to the actual noise sensitivity curve within that interval. Based on the fitted model and combined with the pre-defined noise tolerance interval [α_j, α_j+1], the equation M_i(d, θ*) = α_j is solved to obtain the critical computation depth d_critical. This critical depth is the computation threshold reference value under the corresponding noise tolerance, representing the maximum computation depth at which the ciphertext can be safely executed without exceeding the given noise tolerance.

[0065] Through the above steps, the entire process of constructing a noise sensitivity curve in the encrypted domain based on a hierarchical privacy protection strategy and obtaining noise threshold reference values ​​for each level of computation through piecewise fitting calculations was completed. These noise threshold reference values ​​provide important guidance for the secure processing of encrypted power data, ensuring that the computational power of homomorphic encryption is maximized while meeting privacy protection requirements.

[0066] In one optional implementation, probing operations covering the entire noise tolerance range are performed on the encrypted power data in the encrypted domain, and a multi-dimensional computation path is constructed based on the combined links of the probing operations, including: The boundary thresholds of each noise tolerance interval are extracted from the hierarchical privacy protection strategy. Based on the boundary thresholds, a probing operation sequence covering all noise tolerance intervals is designed. Based on the probing operation sequence, the encrypted power data is subjected to incremental noise intensity operations in the encrypted domain, so that the encrypted noise gradually traverses each noise tolerance interval from the initial noise state until the maximum noise tolerance limit is reached. During this process, the operation type identifier, noise change amount, and noise tolerance interval position of each operation are collected in real time. Based on the ternary combination relationship of the operation type identifier, the noise change amount, and the noise tolerance interval position, a combined link of probing operations is formed. Based on the combined link, path parsing is performed to extract operation nodes with the same operation type identifier. By analyzing the difference in noise change of the operation nodes at different noise tolerance intervals, the operation nodes are divided into stable propagation nodes and accelerated propagation nodes. The stable propagation nodes and the accelerated propagation nodes are used as path construction units and assembled into a multi-dimensional operation path according to the progressive order of the noise tolerance interval positions.

[0067] The noise tolerance interval boundary threshold extraction for the layered privacy protection strategy is implemented through a privacy policy parser. This parser reads multi-level privacy protection configurations from the configuration file and extracts key threshold parameters. The privacy protection configuration is stored in a layered structure, containing four levels: basic protection layer, standard protection layer, enhanced protection layer, and strict protection layer. Each level defines the lower and upper bounds of the noise tolerance interval. The noise tolerance interval for the basic protection layer is 0 to 15, for the standard protection layer it is 15 to 35, for the enhanced protection layer it is 35 to 65, and for the strict protection layer it is 65 to 100. The boundary threshold extraction module parses three fields—level identifier, lower bound, and upper bound—from the configuration structure to generate a boundary threshold array to store the extraction results. The boundary threshold array is represented by a four-tuple structure, containing four elements: level identifier, lower bound threshold, upper bound threshold, and interval width. The interval width is calculated by subtracting the lower bound from the upper bound.

[0068] The probing operation sequence design employs an incremental noise intensity combination strategy to ensure that the ciphertext noise progressively traverses each noise tolerance interval. The operation sequence designer calculates the target noise increment value for each interval based on the boundary threshold array: 7.5 for the basic guard layer, 25 for the standard guard layer, 50 for the enhanced guard layer, and 82.5 for the strict guard layer. Operation selection is based on noise increment intensity configuration: low-intensity noise increments use a combination of addition and rotation operations, medium-intensity noise increments use constant multiplication operations, and high-intensity noise increments use ciphertext multiplication operations. The probing operation sequence contains 48 operations: 16 addition operations, 12 rotation operations, 14 constant multiplication operations, and 6 ciphertext multiplication operations. The operations are arranged in ascending order of noise intensity to ensure that the ciphertext noise gradually increases from its initial value to the maximum tolerance limit.

[0069] The encrypted domain computation operations are executed by the computation execution engine, which processes the encrypted power data sequentially according to a probing sequence. The engine maintains three core components: a computation operation queue, a noise status monitor, and a computation result cache. The computation operation queue employs a first-in, first-out (FIFO) scheduling strategy, retrieving one operation from the head of the queue for execution at a time. The noise status monitor tracks changes in the encrypted noise status in real time, using a noise estimation algorithm to calculate the current noise amplitude and variance. The computation result cache stores intermediate results and status information for each computation operation, with a cache capacity of 64 entries, using a least recently used (LRU) replacement strategy. During computation execution, after each operation is completed, the noise status monitor updates the noise estimate and determines the current noise tolerance interval.

[0070] The data acquisition module records three key pieces of information in real time: operation type identifier, noise change, and noise tolerance interval position. The operation type identifier is represented by integer encoding: addition is identified by 1, rotation by 2, constant multiplication by 3, and encrypted multiplication by 4. The noise change is calculated by the difference in noise amplitude before and after the operation, with a precision maintained to three decimal places. The noise tolerance interval position is represented by interval index values: basic protection layer index is 0, standard protection layer index is 1, enhanced protection layer index is 2, and strict protection layer index is 3. The data acquisition module combines these three pieces of information from each operation into a triplet data structure. Each triplet contains three elements: operation type field, noise change field, and interval position field. The acquired triplet data is stored in an acquisition buffer with a capacity of 512 triplet entries.

[0071] The composite link construction is implemented based on the association analysis of triplet data. A link builder analyzes the dependencies and propagation characteristics between operations. The link builder uses a directed graph data structure to represent the composite links, where nodes correspond to operations and edges represent noise propagation relationships. Node attributes include four fields: operation type identifier, execution time, input noise, and output noise. Edge attributes include three fields: propagation strength, propagation delay, and propagation confidence. Propagation strength is calculated as the ratio of noise variation to computational complexity, propagation delay represents the execution time of the operation, and propagation confidence is determined based on historical statistical data. The composite links use an adjacency list storage structure, where each node maintains a list of edges pointing to its successor nodes, arranged in descending order of propagation strength.

[0072] Path parsing processing extracts and classifies computational operation nodes from the combined link using a path analyzer. The path analyzer first extracts computational operation nodes with the same operation type identifier, forming a set of nodes of the same type. This set is further divided into four subsets based on operation type: an addition operation set (16 nodes), a rotation operation set (12 nodes), a constant multiplication operation set (14 nodes), and a ciphertext multiplication operation set (6 nodes). The path analyzer then performs noise variation difference analysis on each set of nodes of the same type, calculating the statistical characteristics of noise variation at different noise tolerance intervals. These statistical characteristics include three parameters: average noise variation, noise variation variance, and variation range.

[0073] The classification of arithmetic operation nodes is based on the difference in noise variation characteristics to distinguish between stable propagation nodes and accelerated propagation nodes. The node classifier uses a threshold judgment strategy for classification, and the threshold setting is determined based on the statistical analysis results of noise variation. The criteria for determining a stable propagation node are that the variance of the noise variation is less than 30% of the average and the range of the variation is less than 50% of the average. The criteria for determining an accelerated propagation node are that the variance of the noise variation is greater than 70% of the average or the range of the variation is greater than 120% of the average. The node classification results are stored in a classification label array, with stable propagation nodes labeled as 0 and accelerated propagation nodes labeled as 1. The classification label array corresponds to the arithmetic operation node identifier, supporting quick query of node type. Among the addition operation nodes, 14 are classified as stable propagation nodes and 2 as accelerated propagation nodes. Among the rotation operation nodes, 10 are classified as stable propagation nodes and 2 as accelerated propagation nodes. Among the constant multiplication operation nodes, 8 are classified as stable propagation nodes and 6 as accelerated propagation nodes. Among the ciphertext multiplication operation nodes, 1 is classified as a stable propagation node and 5 as accelerated propagation nodes.

[0074] Multi-dimensional computational path construction is achieved through a path assembler using stable propagation nodes and accelerated propagation nodes as building units. The path assembler employs a hierarchical assembly strategy, assembling computational paths at each level sequentially according to the progressive order of noise tolerance interval positions. The path assembly process first identifies the distribution of stable propagation nodes and accelerated propagation nodes within each noise tolerance interval, and then constructs path segments based on node type and interval position. The basic protection layer path segment contains 4 stable propagation nodes and 2 accelerated propagation nodes, the standard protection layer path segment contains 8 stable propagation nodes and 4 accelerated propagation nodes, the enhanced protection layer path segment contains 12 stable propagation nodes and 6 accelerated propagation nodes, and the strict protection layer path segment contains 9 stable propagation nodes and 3 accelerated propagation nodes. Path segments are connected through propagation connectors, which maintain the noise propagation relationship and data dependency relationship between adjacent path segments.

[0075] Multi-dimensional features are extracted from the constructed computation path using a dimension extractor, comprising three main dimensions: computation type dimension, noise propagation dimension, and temporal evolution dimension. The computation type dimension records the quantity and location distribution of each type of operation within the path. The quantity distribution counts the frequency of each computation type within the path, while the location distribution records the relative position index of each computation type within the path. The noise propagation dimension analyzes the intensity changes and propagation patterns of noise propagation within the path. Propagation intensity is calculated through the noise increment between path segments, and the propagation pattern identifies two types: linear propagation and nonlinear propagation. The temporal evolution dimension tracks the temporal characteristics and evolutionary trends during path execution. Temporal characteristics include three parameters: total path execution time, average operation interval, and maximum operation delay. The evolutionary trend is characterized by the slope change of the noise growth curve.

[0076] In one optional implementation, based on the results of the dynamic comparison analysis, noise attenuation parameters for each level of computation are determined, and the ciphertext computation link is hierarchically optimized based on the noise attenuation parameters to complete the decryption processing of the ciphertext computation results. The plaintext analysis results output include: Based on the results of the dynamic comparison analysis, the noise variation characteristics in the ciphertext operation process are segmented and fitted to determine the noise attenuation parameters for each level of operation; based on the noise attenuation parameters, the ciphertext operation link is divided into levels, and the computing resources are dynamically allocated for different operation intervals according to the changing trend of the noise attenuation parameter curve; Based on the configuration results of the computing resources, hierarchical optimization processing is performed on the ciphertext computing link to complete the decryption operation of the ciphertext computing results and output the plaintext analysis results.

[0077] First, dynamic comparison analysis results are processed. By segmenting and fitting the noise variation characteristics during the ciphertext computation process, noise attenuation parameters for each stage of the computation are determined. Specifically, the ciphertext computation chain is divided into multiple computation units, each corresponding to a basic operation type (such as addition, multiplication, or rotation). Sample tests are performed on each computation unit, recording the initial noise level of the input ciphertext and the noise level of the output ciphertext. The noise growth rate, i.e., the ratio of output noise to input noise, is calculated. Based on the noise growth rate data set under different input noise levels, a noise attenuation model is established using piecewise linear fitting or polynomial fitting methods. For example, for multiplication, noise growth follows an exponential relationship, meaning the growth coefficient changes with the initial noise level; while the noise growth for addition exhibits a relatively stable linear relationship. In this way, accurate noise attenuation parameter models are established for each operation type in the ciphertext computation chain.

[0078] In practical applications, taking medical data analysis as an example, processing patients' physiological indicators requires performing multiple multiplication operations to calculate complex health risk scores. Using the aforementioned method, the noise attenuation parameter for multiplication operations is determined to be 1.8 (meaning each multiplication operation increases noise by a factor of 1.8), and the noise attenuation parameter for addition operations is determined to be 1.05 (meaning each addition operation increases noise by 5%). These parameters will guide subsequent allocation of computing resources.

[0079] Based on a defined noise attenuation parameter, the ciphertext computation chain is hierarchically divided and computational resources are dynamically allocated. Computational units are categorized according to the magnitude of the noise attenuation parameter: operations with larger noise attenuation parameters (such as multiplication) are classified as high-noise-growth intervals, while operations with smaller noise attenuation parameters (such as addition) are classified as low-noise-growth intervals. Then, a noise accumulation curve for the entire ciphertext computation chain is plotted to identify key inflection points in noise growth. Key switching or noise management operations are inserted at these inflection points. For different computational intervals, computational resources and encryption parameters are dynamically allocated: higher security parameters and larger key moduli are configured for high-noise-growth intervals to ensure sufficient noise budget; lower security parameters are used for low-noise-growth intervals to improve computational efficiency.

[0080] In financial risk assessment applications, when analyzing customer credit scores, the computational chain involves multiple additions and a small number of multiplications. Based on the aforementioned classification, the portion containing three consecutive multiplications is identified as a high-noise growth interval, which is configured with a 128-bit security parameter and a larger polynomial modulus; while the interval mainly containing addition operations is configured with an 80-bit security parameter, thereby improving computational efficiency while ensuring security.

[0081] Based on the configuration of computing resources, a hierarchical optimization process is performed on the ciphertext computation chain to complete the decryption operation and output the plaintext analysis result. According to the aforementioned hierarchical division, noise management operations, such as re-linearization or analog-to-digital switching, are inserted at key nodes of the computation chain to prevent excessive noise accumulation that could lead to decryption failure. Secondly, the algorithm structure in the computation chain is optimized, converting multiple consecutive multiplication operations into shallower circuit structures, such as using a square-multiplication algorithm to reduce the multiplication depth. Next, batch processing operations are added before and after high-noise-growth intervals to improve computational parallelism. Finally, ciphertext computation is performed according to the optimized computational configuration, and the result is decrypted using the private key in the final stage to output the plaintext analysis data.

[0082] In the specific implementation process, taking gene data analysis as an example, when conducting risk assessment on encrypted data containing information from multiple gene loci, a gene risk scoring model is first constructed. This model includes multi-layer weighted calculations. Based on noise attenuation parameter analysis, three high-noise-growth intervals are identified in the model, located in the second, fourth, and fifth layers of the calculation process, respectively. Noise management operations are inserted before and after these intervals, and the original seven-layer calculation structure is reconstructed into a five-layer structure to reduce the multiplication depth. The optimized encrypted calculation process is then executed, ultimately successfully decrypting the patient's gene risk score with an accuracy comparable to plaintext calculations, while fully protecting the privacy and security of the gene data.

[0083] The above method not only achieves precise control and optimization of the ciphertext computation chain, but also ensures the accuracy of the decryption results, providing reliable technical support for privacy-preserving data analysis. This hierarchical optimization method based on noise attenuation parameters is suitable for various application scenarios requiring high-precision fully homomorphic encryption computation, and can balance multiple needs such as security, accuracy, and computational efficiency.

[0084] A second aspect of this invention provides a dynamic privacy protection system for power data based on fully homomorphic encryption, comprising: An encryption unit is used to acquire power data to be processed, encrypt the power data using a fully homomorphic encryption algorithm to obtain ciphertext power data, and establish an initial noise reference value for the ciphertext power data. The computing unit is used to construct a ciphertext domain state evaluation model, analyze the dynamic correlation characteristics of the ciphertext power data in the calculation process based on the ciphertext domain state evaluation model, predict the executable depth boundary of subsequent ciphertext operations, and generate a hierarchical privacy protection strategy containing multi-level noise thresholds based on the executable depth boundary. The fitting unit is used to construct a noise sensitivity curve for the encrypted power data in the encrypted domain according to the hierarchical privacy protection strategy, obtain the noise threshold reference value for each level of operation by performing piecewise fitting calculation on the noise sensitivity curve, and perform dynamic comparison and analysis between the noise threshold reference value and the predicted value of the encrypted domain state evaluation model. The configuration unit is used to determine the noise attenuation parameters of each level of operation based on the results of the dynamic comparison analysis, and to perform hierarchical optimization configuration of the ciphertext operation link based on the noise attenuation parameters, thereby completing the decryption processing of the ciphertext operation results and outputting the plaintext analysis results.

[0085] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0086] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0087] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic privacy protection method for power data based on fully homomorphic encryption, characterized in that, include: The power data to be processed is acquired, and the power data is encrypted using a fully homomorphic encryption algorithm to obtain ciphertext power data. An initial noise baseline value for the ciphertext power data is then established. Construct a ciphertext domain state evaluation model, analyze the dynamic correlation characteristics of the ciphertext power data in the calculation process based on the ciphertext domain state evaluation model, predict the executable depth boundary of subsequent ciphertext operations, and generate a hierarchical privacy protection strategy containing multi-level noise thresholds based on the executable depth boundary. In the encrypted domain, a noise sensitivity curve is constructed for the encrypted power data according to the hierarchical privacy protection strategy. The noise threshold reference value for each level of operation is obtained by segmented fitting calculation of the noise sensitivity curve. The noise threshold reference value is then dynamically compared and analyzed with the predicted value of the encrypted domain state assessment model. Based on the results of the dynamic comparison analysis, the noise attenuation parameters for each level of operation are determined, and the ciphertext operation link is hierarchically optimized and configured based on the noise attenuation parameters to complete the decryption processing of the ciphertext operation results and output the plaintext analysis results.

2. The method according to claim 1, characterized in that, The power data is encrypted using a fully homomorphic encryption algorithm to obtain ciphertext power data, and an initial noise baseline value for the ciphertext power data is established, including: The power data is identified by data type and classified by sensitivity. Based on the results of the sensitivity classification, the encryption strength requirements corresponding to different types of power data are determined. Based on the encryption strength requirements, the key length and encryption parameter configuration scheme of the fully homomorphic encryption algorithm are selected. The power data is subjected to a fully homomorphic encryption operation using the encryption parameter configuration scheme to generate ciphertext power data. During the encryption process, the initial noise distribution characteristics introduced by the fully homomorphic encryption operation are monitored synchronously, and an initial noise baseline value for the ciphertext power data is established based on the mapping relationship between the initial noise distribution characteristics and the results of the sensitivity classification.

3. The method according to claim 1, characterized in that, A ciphertext domain state evaluation model is constructed. Based on this model, the dynamic correlation characteristics of the ciphertext power data during the calculation process are analyzed to predict the executable depth boundary of subsequent ciphertext operations. A hierarchical privacy protection strategy, including multi-level noise thresholds, is then generated based on this executable depth boundary. The encrypted structure features and initial noise baseline value of the encrypted power data are decoupled. Based on the encrypted structure features and the initial noise baseline value, an encrypted domain state evaluation model is constructed. Based on the encrypted domain state evaluation model, a mapping function between the noise growth rate and the encrypted operation type is established. The ciphertext structural features are parameterized according to the mapping function. The ciphertext operation sequence to be executed is mapped to the ciphertext domain state evaluation model through deep feature reconstruction. The noise accumulation prediction value corresponding to each operation step in the ciphertext operation sequence is deduced using the parameterized mapping function. Based on the difference between the cumulative noise prediction value and the preset upper limit of decryptable noise in the ciphertext, the executable depth boundary of the ciphertext operation sequence is reshaped in combination with the ciphertext structural features. The executable depth boundary is then progressively optimized according to noise sensitivity. A corresponding noise threshold is set for each optimization level to form a hierarchical privacy protection strategy that includes multiple levels of noise thresholds.

4. The method according to claim 3, characterized in that, Based on the ciphertext structural features and the initial noise baseline value, a ciphertext domain state evaluation model is constructed. A mapping function between the noise growth rate and the ciphertext operation type is established according to the ciphertext domain state evaluation model, including: The ciphertext structural features are analyzed for dimensionality and dependencies are extracted to obtain the hierarchical structure information and computational dependency graph of the ciphertext data. The initial noise baseline value is then labeled hierarchically according to the hierarchical structure information to form a multi-level noise state of the ciphertext data. Based on the operation type characteristics in the operation dependency graph, the ciphertext operation is classified according to the degree of influence of the operation. The ciphertext operation is divided into linear noise growth operation and nonlinear noise growth operation, and a corresponding noise propagation parameter is determined for each operation type. The multi-level noise state and the noise propagation parameters are input into the ciphertext domain state evaluation model. Based on the ciphertext domain state evaluation model, the noise evolution process of the ciphertext during continuous operation is tracked through the operation path in the operation dependency graph. Based on the ciphertext domain state evaluation model, the noise growth data in the noise evolution process is fitted and analyzed. Combined with the operation type characteristics, the noise growth rate corresponding to different ciphertext operations is extracted, and a mapping function between the noise growth rate and the ciphertext operation is established.

5. The method according to claim 1, characterized in that, In the encrypted domain, a noise sensitivity curve is constructed for the encrypted power data according to the layered privacy protection strategy. The noise threshold reference values ​​for each level of operation are obtained by piecewise fitting of the noise sensitivity curve, including: The mapping relationship between multi-level noise thresholds and noise tolerance intervals is extracted from the hierarchical privacy protection strategy. In the ciphertext domain, a probing operation covering the entire noise tolerance interval is performed on the ciphertext power data. A multi-dimensional operation path is constructed based on the combined links of the probing operation. Based on the multi-dimensional operation path, the correlation characteristics between the noise evolution state of the ciphertext data and the operation depth are monitored in real time. The noise sensitivity curve is constructed using the correlation characteristics. The noise sensitivity curve is subjected to derivative analysis to identify the inflection point of the sudden change in noise growth acceleration in the noise evolution state. The noise sensitivity curve is divided into multiple noise growth stage intervals with the inflection point as the segment boundary. For each noise growth stage interval, the noise accumulation rate feature and curve curvature feature are extracted by combining the multi-dimensional operation path. Parametric fitting calculation is performed to derive the noise threshold reference value of the operation level of the corresponding stage interval.

6. The method according to claim 5, characterized in that, Performing probing operations covering the entire noise tolerance range on the encrypted power data in the encrypted domain, and constructing a multi-dimensional computation path based on the combined links of the probing operations includes: The boundary thresholds of each noise tolerance interval are extracted from the hierarchical privacy protection strategy. Based on the boundary thresholds, a probing operation sequence covering all noise tolerance intervals is designed. Based on the probing operation sequence, the encrypted power data is subjected to incremental noise intensity operations in the encrypted domain, so that the encrypted noise gradually traverses each noise tolerance interval from the initial noise state until the maximum noise tolerance limit is reached. During this process, the operation type identifier, noise change amount, and noise tolerance interval position of each operation are collected in real time. Based on the ternary combination relationship of the operation type identifier, the noise change amount, and the noise tolerance interval position, a combined link of probing operations is formed. Based on the combined link, path parsing is performed to extract operation nodes with the same operation type identifier. By analyzing the difference in noise change of the operation nodes at different noise tolerance intervals, the operation nodes are divided into stable propagation nodes and accelerated propagation nodes. The stable propagation nodes and the accelerated propagation nodes are used as path construction units and assembled into a multi-dimensional operation path according to the progressive order of the noise tolerance interval positions.

7. The method according to claim 1, characterized in that, Based on the results of the dynamic comparison analysis, noise attenuation parameters for each stage of the operation are determined, and the ciphertext operation link is hierarchically optimized and configured based on the noise attenuation parameters to complete the decryption processing of the ciphertext operation results. The plaintext analysis results output include: Based on the results of the dynamic comparison analysis, the noise variation characteristics in the ciphertext operation process are segmented and fitted to determine the noise attenuation parameters for each level of operation; based on the noise attenuation parameters, the ciphertext operation link is divided into levels, and the computing resources are dynamically allocated for different operation intervals according to the changing trend of the noise attenuation parameter curve; Based on the configuration results of the computing resources, hierarchical optimization processing is performed on the ciphertext computing link to complete the decryption operation of the ciphertext computing results and output the plaintext analysis results.

8. A dynamic privacy protection system for power data based on fully homomorphic encryption, used to implement the method of any one of claims 1-7, characterized in that, include: An encryption unit is used to acquire power data to be processed, encrypt the power data using a fully homomorphic encryption algorithm to obtain ciphertext power data, and establish an initial noise reference value for the ciphertext power data. The computing unit is used to construct a ciphertext domain state evaluation model, analyze the dynamic correlation characteristics of the ciphertext power data in the calculation process based on the ciphertext domain state evaluation model, predict the executable depth boundary of subsequent ciphertext operations, and generate a hierarchical privacy protection strategy containing multi-level noise thresholds based on the executable depth boundary. The fitting unit is used to construct a noise sensitivity curve for the encrypted power data in the encrypted domain according to the hierarchical privacy protection strategy, obtain the noise threshold reference value for each level of operation by performing piecewise fitting calculation on the noise sensitivity curve, and perform dynamic comparison and analysis between the noise threshold reference value and the predicted value of the encrypted domain state evaluation model. The configuration unit is used to determine the noise attenuation parameters of each level of operation based on the results of the dynamic comparison analysis, and to perform hierarchical optimization configuration of the ciphertext operation link based on the noise attenuation parameters, thereby completing the decryption processing of the ciphertext operation results and outputting the plaintext analysis results.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images