A network and information security encryption method based on data security

Abstract

This invention discloses a network and information security encryption method based on data security, relating to the fields of information security and data governance technology. The method includes the following steps: S1, outputting a binary security beacon; S2, generating an initial ciphertext data packet; S3, verifying the homomorphic noise budget value; S4, by improving the EfficientNet-B0 model, introducing a resonance frequency selection mechanism based on physical resonance, performing discrete recursive convolution, and implementing adaptive evolution and classification of the ciphertext adaptation feature flow based on dynamic collaborative thresholds, scaling factors, and distribution parameters according to feature statistical distribution; S5, generating the target transmission frame; S6, outputting the network state fingerprint and verifying the transmission path security; S7, verifying the correctness of the plaintext result. This invention effectively solves the problem of homomorphic operation noise interfering with ciphertext feature extraction, achieving high-precision classification for ciphertext adaptation.

Description

Technical Field

[0001] This invention relates to the fields of information security and data governance technology, and in particular to a network and information security encryption method based on data security. Background Technology

[0002] With the widespread deployment of edge computing and IoT technologies, network transmission environments exhibit complex characteristics of high dynamism, strong heterogeneity, and resource constraints. Traditional data encryption transmission methods mainly rely on pre-set static keys or fixed certificate systems, which have significant authentication flaws when dealing with physical layer attacks targeting edge nodes. While existing fully homomorphic encryption schemes support ciphertext-state computation, they face a bottleneck problem when handling complex neural network inference: ciphertext noise accumulates rapidly with computation depth, making it impossible to complete high-precision ciphertext classification tasks before the noise budget is exhausted. Furthermore, classic secure transmission channels often struggle to incorporate the physical fluctuation characteristics of the underlying data link for dynamic path integrity verification. This results in the inability to detect anomalies in a timely manner when subjected to man-in-the-middle attacks or routing spoofing, allowing maliciously tampered data packets to easily bypass traditional boundary protection mechanisms and enter the decryption stage, seriously threatening the data security and computational reliability of networks and information systems.

[0003] Therefore, how to provide a network and information security encryption method based on data security is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a network and information security encryption method based on data security. It utilizes a resonance frequency selection mechanism based on physical resonance and an improved EfficientNet-B0 model to perform feature extraction and classification on the initial ciphertext data packets. Specifically, it includes: mining semantic information through multidimensional convolution mapping and depthwise convolution operations, mapping each channel of the reciprocal residual feature tensor to a second-order damped harmonic oscillator system. The natural frequency and damping ratio of the harmonic oscillator are dynamically calculated based on the statistical distribution of channel energy, and mapped to coefficients of a second-order difference equation to construct a discrete recursive operator. The discrete recursive operator is used to perform recursive convolution operations on the reciprocal residual feature tensor within a sliding window to output the initial resonance response features. A frequency response benchmark is determined based on the natural frequency, and the response matching deviation between each local spatial feature in the initial resonance response features and the frequency response benchmark is calculated. A dynamic selection threshold is calculated based on the statistical distribution characteristics of the response matching deviation, and local spatial features with response matching deviations less than the dynamic selection threshold are selected as effective signal components. An adaptive gain factor is generated based on the response matching deviation, and the effective signal components are weighted and amplified to output a signal-clean resonance enhancement feature tensor. By integrating multi-scale contextual information through cascaded fusion operations, a composite scaling coefficient is calculated based on the statistical feature distribution of the resonant enhanced feature tensor to adaptively coordinate the number of channels and spatial size, outputting a global semantic high-dimensional feature map. Online statistics are performed based on the statistical feature distribution of the feature descriptor to construct dynamic distribution parameters, which are adaptively projected onto the real-time class boundary, outputting the intermediate ciphertext of the classification probability density. This mechanism introduces a resonance frequency selection mechanism based on physical resonance, achieving adaptive evolution of the feature flow for fully homomorphic encrypted ciphertext data without data decryption. This effectively solves the problem of homomorphic operation noise interfering with ciphertext feature extraction, achieving high-precision classification with ciphertext adaptation.

[0005] A network and information security encryption method based on data security according to an embodiment of the present invention specifically includes: S1. Collect hardware fingerprint data of edge nodes and perform non-linear hash mapping and scrambling operations to generate a master key. Use the master key to initialize the homomorphic encryption parameter set and convert it into a binary security beacon for external broadcast. S2. Obtain the public key based on the binary security beacon, perform homomorphic encryption on the original sensitive data to generate the initial ciphertext data packet, and encapsulate the homomorphic operation auxiliary parameters into the ciphertext header structure; S3. Receive the initial ciphertext data packet and verify its integrity. Parse the ciphertext header to obtain the homomorphic noise budget value and operation type identifier, so as to determine whether the homomorphic noise budget value meets the executable requirements of homomorphic operation. S4. If the homomorphic noise budget value meets the executability requirements for performing homomorphic operations, the initial ciphertext data packet is input into the improved EfficientNet-B0 model without data decryption. A resonance frequency selection mechanism based on physical resonance is introduced to perform discrete recursive convolution. Based on the feature statistical distribution, dynamic collaborative threshold, scaling factor and distribution parameters are used to implement ciphertext adaptation feature flow adaptive evolution and classification, and output intermediate result ciphertext. S5. Obtain the current network link status parameters and encode them into a network status fingerprint. Encapsulate the intermediate result ciphertext with the network status fingerprint to generate the target transmission frame. S6. Transmit the target transmission frame through an encrypted tunnel, obtain the pre-shared session key and establish a decryption channel, parse the network state fingerprint in the target transmission frame and verify the security of the transmission path. S7. After verifying the security of the transmission path, obtain the private key and decrypt the ciphertext of the intermediate result verified by the path to obtain the plaintext of the operation result. Verify the correctness of the plaintext of the operation result based on the encapsulated homomorphic operation auxiliary parameters, and generate the final audit data.

[0006] Optionally, S1 specifically includes: S11. Collect edge node hardware fingerprint data, perform non-linear hash mapping and scrambling operations on the edge node hardware fingerprint data, and generate a master key; S12. Perform homomorphic encryption parameter set initialization calculation based on the master key to obtain the homomorphic encryption parameter set; S13. Convert the homomorphic encryption parameter set to binary format to generate a binary security beacon.

[0007] Optionally, S2 specifically includes: S21. Parse the binary security beacon to extract the homomorphic encryption parameter set, and calculate and extract the public key based on the homomorphic encryption parameter set; S22. Obtain the original sensitive data, and use the public key to perform homomorphic encryption on the original sensitive data to generate the initial ciphertext data; S23. Extract homomorphic operation auxiliary parameters based on the homomorphic encryption parameter set, perform formatted encoding on the homomorphic operation auxiliary parameters, and generate a ciphertext header structure; S24. Using the ciphertext header structure as a preamble and the initial ciphertext data as the payload, the ciphertext header structure and the initial ciphertext data are sequentially concatenated to generate the initial ciphertext data packet.

[0008] Optionally, S3 specifically includes: S31. Receive the initial ciphertext data packet, parse the header structure of the initial ciphertext data packet to extract the integrity check code, and use the integrity check code to perform integrity check calculation on the initial ciphertext data packet. S32. If the integrity verification calculation passes, parse the header structure of the initial encrypted data packet to extract the homomorphic noise budget value and operation type identifier, and query the homomorphic encryption parameter set based on the operation type identifier to obtain the corresponding type of noise consumption estimate value. S33. Calculate the difference between the homomorphic noise budget value and the noise consumption estimate value as the remaining noise margin. If the remaining noise margin is greater than zero, it is determined that the homomorphic noise budget value meets the executability requirements of homomorphic operation.

[0009] Optionally, the improved EfficientNet-B0 model includes an initial feature extraction layer, an inverse residual feature enhancement layer, a resonance frequency selection layer, a multi-scale feature fusion layer, and an adaptive classification output layer. The initial feature extraction layer is used to receive the initial encrypted data packet, extract the basic spatial geometric features through multidimensional convolutional mapping and feature normalization, and output the initial feature map. The inverse residual feature enhancement layer is used to receive the initial feature map, mine semantics and expand the dimension using deep convolution operation, construct a nonlinear operator for ciphertext adaptation to perform approximation transformation, fuse nonlinear features with linear channel features, and output the inverse residual feature tensor. The resonant frequency selection layer is used to introduce a resonant frequency selection mechanism based on physical resonance, and the specific execution process includes: Each channel of the inverse residual characteristic tensor is mapped to a second-order damped harmonic oscillator system. The natural frequency and damping ratio of the harmonic oscillator are dynamically calculated based on the statistical distribution of the channel energy and mapped to the coefficients of the second-order difference equation to construct a discrete recursive operator. The discrete recursive operator is used to perform recursive convolution operation on the inverse residual feature tensor in a sliding window to output the initial resonance response features; the frequency response benchmark is determined based on the natural frequency, and the response matching deviation between each local spatial feature in the initial resonance response features and the frequency response benchmark is calculated. The dynamic selection threshold is calculated based on the statistical distribution characteristics of the response matching deviation. Local spatial features with response matching deviations less than the dynamic selection threshold are selected as effective signal components. An adaptive gain factor is generated based on the response matching deviation to weight and amplify the effective signal components, and a pure resonance enhancement feature tensor is output. The multi-scale feature fusion layer is used to receive the resonant enhancement feature tensor, integrate multi-scale context information using cascaded fusion operations, calculate composite scaling coefficients based on the statistical feature distribution of the resonant enhancement feature tensor to adaptively coordinate the number of channels and spatial size, and output a global semantic high-dimensional feature map. The adaptive classification output layer is used to reduce the dimensionality of the high-dimensional feature map to feature descriptors, perform online statistics based on the statistical feature distribution of the feature descriptors to construct dynamic distribution parameters, adaptively project them to the real-time class boundary, and output the intermediate result ciphertext of the classification probability density.

[0010] Optionally, S5 specifically includes: S51. Collect the status parameters of the current network link, analyze the statistical distribution characteristics of the status parameters, perform hash encoding operation based on the statistical distribution characteristics, and generate network status fingerprint; S52. Obtain the intermediate result ciphertext, map the network state fingerprint to the frame header identifier, and map the intermediate result ciphertext to the frame body payload. S53. Sequentially concatenate and encapsulate the frame header identifier and frame body payload to generate the target transmission frame.

[0011] Optionally, S6 specifically includes: S61. Receive the target transmission frame, retrieve the pre-shared session key stored locally and construct the decryption context, perform frame structure parsing on the target transmission frame, and output the frame header identifier and frame body payload. S62. Parse the frame header identifier to extract the reference link statistical feature vector recorded by the sender; S63. Collect the current network link status parameters of the receiving end, including round-trip time jitter sequence, data packet arrival interval distribution and instantaneous link throughput; S64. Construct a sliding time window, perform statistical analysis on network link state parameters, and extract the mean, variance, and kurtosis index; map the mean, variance, and kurtosis index to a vector space to construct a local high-dimensional statistical feature vector; S65. Calculate information entropy based on high-dimensional statistical feature vectors as network fluctuation features, and cascade and fuse network fluctuation features with high-dimensional statistical feature vectors to generate feature data to be verified. S66. Encapsulate the feature data to be verified into a feature vector to be verified, and use it as a local network state fingerprint. S67. Track the historical statistical distribution characteristics of the local network status fingerprint within the sliding time window, and dynamically construct an adaptive confidence interval based on the dispersion of the historical statistical distribution characteristics; calculate the feature deviation between the local network status fingerprint and the baseline link statistical feature vector. If the feature deviation falls within the adaptive confidence interval, the transmission path security verification is deemed successful, and the intermediate result ciphertext verified by the path is output.

[0012] Optionally, the S7 specifically includes: S71. Retrieve the local private key and construct a decryption context, obtain the path-verified intermediate result ciphertext, and use the decryption context to perform a homomorphic decryption operation on the path-verified intermediate result ciphertext to generate the operation result plaintext. S72. Parse the encapsulation structure of the intermediate result ciphertext verified by the path to extract the auxiliary parameters of the homomorphic operation, including the initial noise distribution characteristics, the polynomial ring order and the check hash value. S73. Based on the initial noise distribution characteristics and the order of the polynomial ring, construct an adaptive noise tolerance boundary and define the adaptive noise tolerance boundary as a correctness verification constraint. S74. Calculate the actual statistical distribution characteristics of the plaintext of the operation result and the real-time hash value, substitute the actual statistical distribution characteristics into the correctness verification constraints for matching, and compare the real-time hash value with the verification hash value; S75. If the actual statistical distribution characteristics are within the adaptive noise tolerance boundary and the real-time hash value is consistent with the verification hash value, output the verification result identifier that the verification has passed; otherwise, output the verification result identifier that the verification has failed. S76. Filter the plaintext of the operation result according to the verification result identifier, associate and encapsulate the plaintext of the verified operation result with the corresponding homomorphic operation auxiliary parameters, and generate the final audit data.

[0013] The beneficial effects of this invention are: (1) This invention realizes key generation and initialization from the physical device layer to the homomorphic encryption parameter set by constructing a nonlinear hash mapping and scrambling operation system based on edge node hardware fingerprints. Edge node hardware fingerprint data is collected and nonlinear hash mapping and scrambling operations are performed to generate a master key. The master key is used to initialize the homomorphic encryption parameter set and convert it into a binary security beacon. The binary security beacon is parsed to extract the homomorphic encryption parameter set. Based on the homomorphic encryption parameter set, a public key is calculated and homomorphic encryption operations are performed on the original sensitive data. The homomorphic operation auxiliary parameters are formatted and encapsulated into a ciphertext header structure. Integrity check codes are used to verify data integrity, and the executability of the homomorphic operation is determined based on the homomorphic noise budget value and the noise consumption estimate value. This system utilizes the uniqueness and physical non-cloning properties of hardware fingerprints, combined with the ciphertext state computation capability of homomorphic encryption, to ensure strong randomness in key generation and confidentiality in data transmission. Simultaneously, the pre-verification mechanism of the noise budget ensures the availability of subsequent ciphertext operations.

[0014] (2) This invention establishes a feature flow adaptive evolution and classification system for ciphertext adaptation by adopting an improved EfficientNet-B0 model and a resonance frequency selection mechanism based on physical resonance. The initial ciphertext data packet is input into the improved EfficientNet-B0 model, and spatial geometric features and semantic information are extracted using multidimensional convolution mapping and depthwise convolution operations. Each channel of the inverse residual feature tensor is mapped to a second-order damped resonator system. The natural frequency and damping ratio of the resonator are dynamically calculated based on the statistical distribution of channel energy, and mapped to the coefficients of the second-order difference equation to construct a discrete recursive operator. The discrete recursive operator is used to perform recursive convolution operations on the inverse residual feature tensor in a sliding window to output the initial resonance response features. The frequency response benchmark is determined based on the natural frequency, the response matching deviation is calculated, and the dynamic selection threshold and adaptive gain factor are calculated using statistical distribution features to weight and amplify the effective signal components, outputting a signal-pure resonance enhancement feature tensor. The composite scaling coefficient and dynamic distribution parameters are calculated based on the statistical feature distribution to adaptively coordinate the number of channels and spatial size and project them onto the real-time class boundary. This system achieves high-precision feature extraction and noise suppression for fully homomorphic encrypted ciphertext data without data decryption by introducing an adaptive collaborative mechanism that combines the frequency selection characteristics and statistical distribution features of the physical resonance model. This ensures the accuracy and robustness of neural network inference in the ciphertext domain. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a network and information security encryption method based on data security proposed in this invention; Figure 2 This is a flowchart illustrating the working principle of the improved EfficientNet-B0 model, an improved network and information security encryption method based on data security proposed in this invention. Detailed Implementation

[0016] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figure 1 and Figure 2 A network and information security encryption method based on data security, specifically including: S1. Collect hardware fingerprint data of edge nodes and perform non-linear hash mapping and scrambling operations to generate a master key. Use the master key to initialize the homomorphic encryption parameter set and convert it into a binary security beacon for external broadcast. S2. Obtain the public key based on the binary security beacon, perform homomorphic encryption on the original sensitive data to generate the initial ciphertext data packet, and encapsulate the homomorphic operation auxiliary parameters into the ciphertext header structure; S3. Receive the initial ciphertext data packet and verify its integrity. Parse the ciphertext header to obtain the homomorphic noise budget value and operation type identifier, so as to determine whether the homomorphic noise budget value meets the executable requirements of homomorphic operation. S4. If the homomorphic noise budget value meets the executability requirements for performing homomorphic operations, the initial ciphertext data packet is input into the improved EfficientNet-B0 model without data decryption. A resonance frequency selection mechanism based on physical resonance is introduced to perform discrete recursive convolution. Based on the feature statistical distribution, dynamic collaborative threshold, scaling factor and distribution parameters are used to implement ciphertext adaptation feature flow adaptive evolution and classification, and output intermediate result ciphertext. S5. Obtain the current network link status parameters and encode them into a network status fingerprint. Encapsulate the intermediate result ciphertext with the network status fingerprint to generate the target transmission frame. S6. Transmit the target transmission frame through an encrypted tunnel, obtain the pre-shared session key and establish a decryption channel, parse the network state fingerprint in the target transmission frame and verify the security of the transmission path. S7. After verifying the security of the transmission path, obtain the private key and decrypt the ciphertext of the intermediate result verified by the path to obtain the plaintext of the operation result. Verify the correctness of the plaintext of the operation result based on the encapsulated homomorphic operation auxiliary parameters, and generate the final audit data.

[0018] In this embodiment, S1 specifically includes: S11. Read the CPU serial number, NIC physical address, and motherboard unique identifier of the edge node, and concatenate them into a 256-bit original hardware fingerprint string; set the hash function to SHA-256, input the original hardware fingerprint string into the hash function, and calculate and output a 256-bit initial hash digest; initialize a 256-bit scrambling index table, fill in the index values ​​from 0 to 255 in sequence, and for each position of the index value from 0 to 255, use the formula (current position index value plus the value of the corresponding byte of the current initial hash digest plus 256) modulo 256 to obtain a random index value, swap the values ​​at the current position index value and the random index value in the scrambling index table to generate the final scrambling index table; use the values ​​in the scrambling index table as position guides, adjust the byte order in the initial hash digest according to the guides, move the bytes in the initial hash digest to the new position specified by the scrambling index table, thereby generating a scrambled hash digest, and define the scrambled hash digest as the master key.

[0019] S12. Read the master key, set the modulus of the polynomial ring to 32768, the polynomial degree to 1024, and the coefficient range to -16384 to 16383; initialize a seed generator using the master key, set the seed generator algorithm to AES-128, input the master key as the seed into the seed generator, and generate a binary-distributed pseudo-random sequence; convert every 16 binary bits in the pseudo-random sequence into a decimal integer, take the remainder of the converted integer divided by 32768, and when the result is greater than or equal to 16384, subtract 32768 to map it to the interval between -16384 and 16383, generating the polynomial coefficient set; encapsulate the polynomial degree, modulus, and polynomial coefficient set to obtain the homomorphic encryption parameter set.

[0020] S13. Read the homomorphic encryption parameter set, convert the polynomial degree value 1024 into a 32-bit unsigned integer and append it to the data stream header, convert the modulus value 32768 into a 32-bit unsigned integer and append it to the data stream header; traverse the polynomial coefficient set in the homomorphic encryption parameter set, for each decimal integer coefficient, if the value is greater than or equal to 0, directly convert it into a 16-bit binary original code, if the value is less than 0, calculate the result after adding 32768 and convert the result into a 16-bit binary string to generate a coefficient data stream; concatenate the binary strings of all coefficients to generate a coefficient data stream of length 16384 bits; append header data consisting of degree and modulus to the front of the coefficient data stream, and append a 32-bit frame header to the very front, setting the hexadecimal value of the frame header to 0xAF51 to identify the start of the data stream; define the data stream with frame header, parameter header and coefficients as a binary security beacon, as input data for subsequent broadcast transmission.

[0021] In this embodiment, S2 specifically includes: S21. Read the binary security beacon, parse the beacon header to extract the modulus value 32768 and the polynomial degree value 1024; calculate the length of the data bits in the beacon after removing the frame header and parameter header, divide this length by 16 to obtain the number of coefficients, convert each 16 bits of binary data into a decimal integer, if the integer is greater than or equal to 16384, subtract 32768 to map it to the interval -16384 to 16383, generating a polynomial coefficient set, thereby obtaining the homomorphic encryption parameter set; initialize an empty polynomial as the secret polynomial, set the highest power of this polynomial to 1023, and use a discrete Gaussian distribution sampling algorithm with a standard deviation of 3.0 to sequentially... Generate 1024 integers as coefficients of the polynomial and fill them into the secret polynomial; initialize an empty polynomial as a uniformly random polynomial, and use a random number generator to generate 1024 integers between 0 and 32767, which are then filled into the uniformly random polynomial as coefficients; perform polynomial multiplication on the uniformly random polynomial and the secret polynomial, and take the remainder of each coefficient of the product modulo 32768; perform addition on the remainder result and an error polynomial generated by sampling to generate the first component of the public key; take the opposite of each coefficient of the secret polynomial to generate the second component of the public key; combine the first and second components to generate the public key.

[0022] S22. Obtain the original sensitive data, set the plaintext scaling factor to 256, multiply each data point value in the original sensitive data by the scaling factor 256; take the remainder of the product result modulo 32768, and use the remainder value as the coefficient of the plaintext polynomial; initialize an empty ciphertext polynomial, generate a random mask polynomial for encryption using a random number generator, and set the coefficient of the random mask polynomial to be between 0 and 32767; perform polynomial multiplication operation on the random mask polynomial and the first component of the public key, take the remainder of the product result modulo 32768, and generate the first intermediate term; perform polynomial multiplication operation on the plaintext polynomial and the second component of the public key, take the remainder of the product result modulo 32768, and generate the second intermediate term; perform addition operation on the first intermediate term and the second intermediate term, add the operation result to a sampled noise polynomial, take the remainder of the addition result modulo 32768, and generate the initial ciphertext data.

[0023] S23. Extract homomorphic operation auxiliary parameters based on the homomorphic encryption parameter set, set the specific value of the version number field to 1, and convert the value 1 into 8-bit binary data; set the modulus parameter value to 32768, and convert the value 32768 into 16-bit binary data; set the scaling factor parameter value to 256, and convert the value 256 into 16-bit binary data; concatenate the 8-bit version number binary data, the 16-bit modulus binary data, and the 16-bit scaling factor binary data in sequence to generate a 40-bit binary basic stream; perform polynomial division on the 40-bit binary basic stream, set the divisor polynomial to the standard CRC-16 polynomial, and use the 16-bit remainder obtained from the operation as the check code; append the 16-bit check code to the end of the 40-bit binary basic stream to generate a ciphertext header structure with a total length of 56 bits.

[0024] S24. Read the initial ciphertext data, calculate the total number of polynomial coefficients contained in the initial ciphertext data, and convert each coefficient into a 16-bit binary string; concatenate the 16-bit binary strings corresponding to all coefficients in order to generate a ciphertext payload binary stream; place the ciphertext header structure at the beginning of the data packet as a preamble; append the ciphertext payload binary stream to the end of the ciphertext header structure; calculate the total length of the appended data packet, set the padding block length of the data packet to 128 bits, and if the total length is not an integer multiple of 128 bits, pad the end of the data packet with bytes of value 0 until the total length of the data packet is an integer multiple of 128 bits, and generate the initial ciphertext data packet.

[0025] In this embodiment, S3 specifically includes: S31. Read the initial encrypted data packet, locate the frame header identifier at the beginning of the data packet, and extract the 56-bit header structure after the frame header; read a 16-bit data segment starting from the 41st bit of the header and define it as the integrity check code; read the first 40 bits of binary data in the header, set the divisor polynomial to the standard CRC-16 polynomial, perform polynomial division on the first 40 bits of binary data, and compare the 16-bit remainder obtained with the integrity check code; if the two values ​​are completely equal, output the verification result that the integrity check has passed; read the binary data from the 9th to the 24th bit of the header structure, convert it into a decimal integer, and define it as the homomorphic noise budget value; read the binary data from the 25th to the 32nd bit of the header structure and convert it into an operation type identifier.

[0026] S32. Read the operation type identifier and set the query index table to include noise parameters for addition and multiplication operations; use the operation type identifier to match the corresponding noise consumption estimate in the index table, set the noise consumption estimate for addition to the value 30, and set the noise consumption estimate for multiplication to the value 200; define the matched noise consumption estimate as the noise consumption amount of this operation.

[0027] S33. Read the homomorphic noise budget value as the minuend, read the noise consumption of this operation as the subtrahend, perform the subtraction operation of minuend minus subtrahend, and define the result as the remaining noise margin; compare the remaining noise margin with the decision threshold 0; when the value of the remaining noise margin is greater than the decision threshold 0, generate a logic decision signal for the homomorphic operation to be executable, confirming that the homomorphic noise budget value meets the executable requirements of the homomorphic operation; when the value of the remaining noise margin is less than or equal to the decision threshold 0, generate a logic decision signal for the homomorphic operation to be unexecutable, and refuse to execute the homomorphic operation.

[0028] In this embodiment, the improved EfficientNet-B0 model includes an initial feature extraction layer, an inverse residual feature enhancement layer, a resonance frequency selection layer, a multi-scale feature fusion layer, and an adaptive classification output layer: The initial feature extraction layer reads the initial encrypted data packet, using the two-dimensional array of the initial encrypted data packet as the input matrix. The convolution kernel size is set to 3 rows and 3 columns, the number of convolution kernels is set to 32, the stride is set to 2 pixels, and the padding method is to add a ring of zeros around the edges. A sliding window convolution operation is performed on the input matrix using the convolution kernels, multiplying the kernel elements with the corresponding elements of the input matrix and summing the results to generate 32 feature maps. The mean and standard deviation of all pixel values ​​on each feature map are calculated. The mean value is subtracted from each pixel value and then divided by the standard deviation to generate normalized feature values. The activation function is set to a modified linear unit (MRU), which outputs 0 when the input value is less than 0 and the original value when the input value is greater than 0. Activation is then performed on the normalized feature values, outputting a spatial geometric fundamental feature map.

[0029] The inverse residual feature enhancement layer is used to read the spatial geometric fundamental feature map. The depthwise convolution kernel is set to a size of 3 rows and 3 columns, and the convolution stride is set to 1 pixel. The depthwise convolution kernel performs convolution operations on each channel of the spatial geometric fundamental feature map to extract deep semantic features. The dilation factor is set to 6 times, multiplying the number of channels of the deep semantic features by 6 to generate an expanded feature tensor. The activation function is set to the hyperbolic tangent function, calculating the hyperbolic tangent value of the input value and performing a pointwise nonlinear transformation on the expanded feature tensor to generate nonlinear features. A 1-row, 1-column convolution kernel restores the number of channels of the nonlinear features to their original values, generating linear channel features. Finally, a residual concatenation operation is performed between the nonlinear features and the linear channel features, adding elements at corresponding positions, to output the inverse residual feature tensor.

[0030] The resonant frequency selection layer is used to introduce a resonant frequency selection mechanism based on physical resonance. The specific execution process includes: Read the inverse residual feature tensor and map each channel value of the inverse residual feature tensor to a second-order damped harmonic oscillator system; calculate the sum of squares of all values ​​in each channel as the channel energy, and set the average channel energy as the energy reference; calculate the ratio of the channel energy to the energy reference, and set the ratio as the natural frequency coefficient, and set the specific value of the damping ratio to 0.1; use the natural frequency coefficient and the damping ratio to calculate the coefficients of the second-order difference equation, calculate the forward coefficient as 2 times the natural frequency coefficient, calculate the feedback coefficient as the negative of the damping ratio, and construct a discrete recursive operator; set the length of the sliding window to 3, and use the discrete recursive operator to perform recursive convolution operation on the inverse residual feature tensor to output the initial resonance response characteristics.

[0031] Read the initial resonance response characteristics and determine the frequency response benchmark based on the inherent frequency coefficients; calculate the difference between the value of each local spatial feature in the initial resonance response characteristics and the value of the frequency response benchmark, and define the absolute value of the difference as the response matching deviation; calculate the mean and standard deviation of the response matching deviation, and set the mean plus twice the standard deviation as the dynamic selection threshold; compare the response matching deviation with the dynamic selection threshold, and select the local spatial features with response matching deviation less than the dynamic selection threshold as effective signal components; calculate the gain factor as the reciprocal of the response matching deviation, calculate the product of the value 1 and the gain factor and add a smoothing parameter of 0.01 to prevent numerical overflow, and perform weighted amplification on the effective signal components to output the resonance enhancement feature tensor.

[0032] A multi-scale feature fusion layer is used to read the resonant enhancement feature tensor. The convolutional kernel sizes of different scales are set to 1 row 1 column, 3 row 3 columns, and 5 row 5 columns, respectively. Features are extracted using convolutional kernels of different scales. The extracted features are concatenated end to end in the channel dimension to generate a cascaded feature map. The channel variance of the cascaded feature map is calculated, and the mean of the channel variance is set as the distribution statistic. The scaling factor is set as the square root of the distribution statistic. The scaling factor is used to perform an adaptive coordination transformation of the number of channels and spatial size on the cascaded feature map, and output a global semantic high-dimensional feature map.

[0033] The adaptive classification output layer reads the global semantic high-dimensional feature map, performs global average pooling, calculates the sum of the values ​​of each spatial location in the global semantic high-dimensional feature map across all channels, and divides it by the total number of spatial pixels, compressing the spatial dimension of the feature map to 1 row and 1 column to generate a one-dimensional channel feature vector. It then reads the length of the one-dimensional channel feature vector, sets the weight matrix of the fully connected layer to 128 rows and N columns, where N is the length of the one-dimensional channel feature vector, performs multiplication and addition operations on the weight matrix of the fully connected layer and the one-dimensional channel feature vector, generating a 128-dimensional feature descriptor. It calculates the maximum and minimum values ​​of the feature descriptor, setting the difference between the maximum and minimum values ​​as the dynamic distribution parameter. It sets the projection boundary as the midpoint of the dynamic distribution parameter, subtracts the projection boundary from each value in the feature descriptor, divides the result by the dynamic distribution parameter, and generates a normalized projection value. Finally, it sets the activation function to the Softmax function, calculates the exponential function value of the normalized projection value, divides the exponential function value by the sum of all exponential function values, and outputs the intermediate encrypted result of the classification probability density.

[0034] The improved EfficientNet-B0 model proposed in this step is similar to the traditional EfficientNet-B0 model in that it is based on the multi-level feature abstraction and hierarchical semantic mapping theory of deep convolutional neural networks. That is, it extracts the low-level spatial geometric features through multi-dimensional convolution operations, uses the inverse residual structure for deep feature mining and dimensional transformation, and projects the high-dimensional feature map onto the class probability space through the classification output layer to output the final classification result.

[0035] The difference lies in that this invention overcomes the limitations of traditional convolutional neural networks in processing fully homomorphic encrypted ciphertext data, which are susceptible to noise interference and static feature extraction. It adds a resonant frequency selection layer after the inverse residual enhancement, introducing a resonant frequency selection mechanism based on physical resonance. This mechanism maps the feature channels to a second-order damped resonator system, uses discrete recursive operators to perform sliding window recursive convolution operations, and calculates a dynamically selected threshold and adaptive gain factor based on response matching deviation to weighted amplify the effective signal components, rather than relying solely on fixed linear convolution kernels and static activation functions for feature transfer.

[0036] The beneficial effects of the improvements are that, by introducing the frequency selection characteristics and gain control of the physical resonance model, this invention embeds the resonance filtering mechanism from signal processing into the forward propagation of the neural network, breaking the limitations of traditional methods in the ciphertext domain where the feature texture becomes blurred and the classification accuracy decreases due to the superposition of homomorphic noise. It achieves a precise conversion from static geometric feature extraction to dynamic resonance signal enhancement. This design significantly enhances the model's ability to perceive weak features in fully homomorphic encrypted ciphertext data, can adaptively filter out noise interference based on statistical distribution, and outputs a pure resonance enhancement feature tensor, effectively improving the feature flow evolution accuracy of ciphertext adaptation and the classification robustness in untrusted environments.

[0037] In this embodiment, S5 specifically includes: S51. Collect the round-trip time delay, packet loss rate, and throughput of the current network link. Set the sampling time window to 1 second and collect 1000 data points within 1 second. Calculate the average round-trip time delay, the sum of packet loss rates, and the variance of throughput. Construct a hash encoding function with an input length of 4 and an output length of 32. Set the hash function to the MD5 algorithm. Concatenate the average round-trip time delay, the sum of packet loss rates, the variance of throughput, and the sampling time window values ​​into an input string. Input this input string into the MD5 hash function for calculation and output a 128-bit binary hash value. Extract the first 32 bits of the 128-bit binary hash value and convert these 32 bits of binary data into an unsigned integer, which is defined as the network state fingerprint.

[0038] S52. Read the intermediate result ciphertext of the classification probability density output in step S46, encapsulate all the binary bits of the intermediate result ciphertext, and define it as the frame body payload; read the network state fingerprint, set the value range of the network state fingerprint to 0 to 4294967295, convert the fingerprint value into 32-bit binary data, and use the 32-bit binary data as the start identifier of the data frame, defining it as the frame header identifier.

[0039] S53. Read the frame header identifier and frame body payload, set the frame header identifier to the beginning of the data packet, set the frame body payload to the end of the frame header identifier, perform a head-to-tail concatenation operation on the two to generate a binary data stream with a total length of 32 bits plus the frame body payload bits, and define the binary data stream as the target transmission frame.

[0040] In this embodiment, S6 specifically includes: S61. Read the pre-shared session key stored locally, construct a decryption context, use the context to perform frame header parsing operation on the input target transmission frame, locate the first 32 bits of binary data at the beginning of the data packet, extract it and define it as the frame header identifier; read the remaining binary data stream immediately following the frame header identifier, extract it and define it as the frame body payload.

[0041] S62. Read the frame header identifier, use the pre-shared session key as the key input, set the initialization vector to all zeros, execute the AES-128 decryption algorithm, and output a 128-bit decrypted binary bit stream; set the floating-point conversion precision to 32 bits, group the decrypted binary bit stream into 32-bit groups, and obtain a total of 4 groups of data. Convert each group of binary data into IEEE754 standard decimal floating-point numbers, and concatenate them in order to generate a reference link statistical feature vector containing 4 elements recorded by the sender.

[0042] S63. Collect real-time status data of the current network link at the receiving end, set the sampling frequency to 100 Hz, the collection time to 1 second, and acquire a total of 100 data points; record the round-trip delay jitter value, data packet arrival interval value and instantaneous link throughput value of each data point, and generate a network link status parameter sequence.

[0043] S64. Construct a sliding time window with a length of 10 data points, sliding it backward from the starting position on the network link state parameter sequence, with each sliding step being 1 data point; for the data points within the sliding time window, calculate the arithmetic mean of the round-trip delay jitter, the variance of the packet arrival interval, and the kurtosis of the instantaneous link throughput; concatenate the arithmetic mean, variance, and kurtosis in order to construct a local high-dimensional statistical feature vector containing 3 elements.

[0044] S65. Read the local high-dimensional statistical feature vector, calculate the probability density of each element in the feature vector, and set the probability density value to the sum of the element values ​​divided by the sum of all element values; calculate the product of each probability density value and the logarithm of the probability density value to the base 2, add all the products and take the opposite number to obtain the information entropy value; concatenate the information entropy value with the local high-dimensional statistical feature vector to generate the feature data to be verified.

[0045] S66. Read the feature data to be verified, convert it into a column vector, set the dimension of the column vector to 4 rows and 1 column, and define the column vector as the local network state fingerprint of this round.

[0046] S67. Read the local network status fingerprint, setting the historical statistical length to 20, i.e., retaining the local network status fingerprints generated in the most recent 20 sliding time windows; for each dimension of data, calculate the arithmetic mean of the 20 fingerprint values ​​as the baseline mean for that dimension, and calculate the standard deviation as the fluctuation measure for that dimension; set the confidence coefficient to a value of 2, calculate the baseline mean of each dimension minus twice the standard deviation to obtain the lower confidence limit for that dimension, and calculate the baseline mean plus twice the standard deviation to obtain the upper confidence limit for that dimension, generating an adaptive confidence interval set consisting of 4 intervals; read the base data recorded by the sender output in step S62. The quasi-link statistical feature vector is used to calculate the Euclidean distance between the local network state fingerprint and the baseline link statistical feature vector, resulting in a distance value. The modulus of the local network state fingerprint is also calculated, resulting in a modulus value. The distance value is divided by the modulus value to obtain the relative feature deviation. Each dimension value of the local network state fingerprint is compared with the corresponding confidence interval in the adaptive confidence interval set. If the relative feature deviation value is less than 0.5 and all dimension values ​​fall within the corresponding confidence interval, the transmission path security verification is considered successful. The frame payload is then read and defined as the intermediate result ciphertext verified by the path.

[0047] The network status fingerprint dynamic verification process proposed in this step is similar to the traditional network data integrity verification process in that both are based on the theory of network communication link status monitoring and data frame structure parsing. That is, control information is obtained by extracting the frame header of the transmitted data packet, the payload data is processed by constructing a decryption context using the locally stored key, and a specific verification mechanism is used to logically determine the legality of the data transmission.

[0048] The difference lies in that this invention breaks through the limitations of traditional cryptographic authentication that ignores the dynamic changes in the physical environment of transmission. It adds a network physical feature acquisition step to extract link state parameters and replaces traditional hash comparison with high-dimensional statistical feature mapping. Round-trip time jitter, arrival interval distribution and instantaneous throughput are mapped to mean, variance and kurtosis indices to perform information entropy calculation. Finally, an adaptive confidence interval is constructed based on historical statistical distribution characteristics and the feature deviation is determined, rather than a single integrity check code matching.

[0049] The beneficial effects of this improvement are that, by using multidimensional statistical features and adaptive confidence intervals, the physical link fluctuation characteristics are rigidly embedded into the security verification process. This breaks the security blind spot caused by the cryptographic authentication being successful but the physical characteristics being abnormal when facing route hijacking or man-in-the-middle attacks, and achieves a precise transformation from static cryptographic verification to dynamic physical behavior perception. This design significantly enhances the ability to perceive transmission path anomalies, and can accurately identify illegal forwarding paths in a high-dimensional statistical feature space. Combined with adaptive confidence interval judgment, it effectively improves the anti-attack capability and security of data transmission in untrusted network environments.

[0050] In this embodiment, S7 specifically includes: S71. Read the locally stored secret polynomial as the local private key and construct a homomorphic decryption context; read the path-verified intermediate result ciphertext output from step S67 and set the modulus value to 32768; perform polynomial multiplication on the path-verified intermediate result ciphertext using the secret polynomial, and take the remainder of each coefficient in the product result with respect to the modulus 32768; set the noise removal coefficient to a scaling factor of 256, divide the remainder polynomial coefficients by the noise removal coefficient, round the calculation result to the nearest integer, and generate a plaintext result containing 1024 integer coefficients.

[0051] S72. Read the frame header identifier extracted in step S61, parse the high 16 bits of binary data in the frame header identifier, convert it into an unsigned integer and define it as the initial noise distribution feature; parse the low 16 bits of binary data in the frame header identifier, convert it into an unsigned integer and define it as the polynomial ring order; read the intermediate result ciphertext verified by the path, locate the 16 bits of binary data at the end of the data stream and define it as the check hash value; encapsulate the initial noise distribution feature, the polynomial ring order and the check hash value into homomorphic operation auxiliary parameters.

[0052] S73. Read the initial noise distribution characteristics in the homomorphic operation auxiliary parameters, set the standard deviation of the characteristic to be a discrete Gaussian distribution, and read the order of the polynomial ring as the power value; set the noise tolerance factor to 3, calculate the product of the standard deviation and the noise tolerance factor 3, and then multiply it by the square root of the power value to obtain the basic noise threshold; set the safety boundary margin to 5, calculate the sum of the basic noise threshold and the safety boundary margin 5, define the calculation result as the adaptive noise tolerance boundary, and define the adaptive noise tolerance boundary as the correctness verification constraint.

[0053] S74. Read the 1024 integer coefficients from the plaintext of the operation result, calculate the absolute value of all coefficient values, sum the absolute values ​​to obtain the total noise value, and define the total noise value as the actual statistical distribution characteristic; construct a hash function with an input length equal to the plaintext bitstream of the operation result and an output length of 16 bits, set the specific form of the hash function as the CRC-16 algorithm, input the plaintext of the operation result into the CRC-16 algorithm for calculation, and generate a real-time hash value; read the adaptive noise tolerance boundary in the correctness verification constraint, compare the total noise value in the actual statistical distribution characteristic with the adaptive noise tolerance boundary; read the verification hash value, and compare the real-time hash value with the verification hash value bit by bit.

[0054] S75. Set the verification judgment logic. If the total noise value in the actual statistical distribution characteristics is less than the adaptive noise tolerance boundary, and all binary bits of the real-time hash value and the verification hash value are exactly the same, then generate a string with the character content "Success" and define it as the verification result identifier. If the total noise value is greater than or equal to the adaptive noise tolerance boundary, or if any bit of the real-time hash value and the verification hash value are different, then generate a string with the character content "Fail" and define it as the verification result identifier.

[0055] S76. Read the verification result identifier and check if the identifier content is "Success". If the check result is yes, read the plaintext of the operation result and the homomorphic operation auxiliary parameters, construct a key-value pair data structure, set the key name to "Parameters", set the key value to the homomorphic operation auxiliary parameters, set the key name to "Plaintext", set the key value to the plaintext of the operation result, encapsulate the key-value pair data structure into a JSON format string, and define it as the final audit data. If the check result is no, generate an empty string as the final audit data.

[0056] Example 1: To verify the feasibility of this invention in financial data privacy computation and joint risk control, the method of this invention was applied to the interbank joint risk control system of a provincial commercial bank (hereinafter referred to as "Bank A"). In traditional financial risk control data sharing systems, centralized databases based on plaintext exchange or data transmission based on simple RSA encryption are typically used. These methods not only make it difficult to achieve deep feature sharing while ensuring the data sovereignty of each participating party, but also cannot perform complex neural network inference in encrypted form, resulting in the risk of data leakage and the inability to effectively utilize high-value customer profile features. To solve the above problems, Bank A decided to adopt a network and information security encryption method based on data security proposed in this invention.

[0057] During implementation, Bank A first collected hardware fingerprint data from edge nodes. Using sensors deployed at participating parties' gateways, it acquired CPU serial numbers, MAC addresses, and hardware configuration information. A master key was generated through non-linear hash mapping and scrambling operations. This master key was then used to initialize a homomorphic encryption parameter set, which was converted into a binary security beacon and broadcast externally. Simultaneously, Bank A's data security team performed rigorous privacy grading and feature standardization on the collected raw sensitive credit data, serving as the benchmark for model training and encryption operations.

[0058] Bank A extracts the homomorphic encryption parameter set by receiving a binary security beacon, calculates the public key based on the parameter set, and performs homomorphic encryption on the original sensitive data to generate an initial ciphertext data packet. Next, it performs integrity verification on the initial ciphertext data packet using an integrity check code, parses the ciphertext header structure to obtain the homomorphic noise budget value and operation type identifier, and calculates the remaining noise margin to determine the executability of the homomorphic operation, effectively filtering out ciphertext that cannot be computed due to insufficient noise budget. Subsequently, without decrypting the data, the initial ciphertext data packet is input into the improved EfficientNet-B0 model. Semantic information is mined using a reciprocal residual feature enhancement layer, and each channel is mapped to a second-order damped harmonic oscillator system. Based on the statistical distribution of channel energy, the natural frequency and damping ratio of the harmonic oscillator are dynamically calculated.

[0059] In the core encrypted inference stage, this invention introduces a resonance frequency selection mechanism based on physical resonance. It utilizes a discrete recursive operator to perform recursive convolution operations on the inverse residual feature tensor within a sliding window, calculating the response matching bias. Based on statistical distribution characteristics, it calculates a dynamically selected threshold and an adaptive gain factor to weight and amplify the effective signal components, outputting a pure resonance-enhanced feature tensor. A multi-scale feature fusion layer calculates composite scaling coefficients to integrate multi-scale contextual information, and an adaptive classification output layer constructs dynamic distribution parameters, adaptively projecting them onto the real-time class boundary to output the intermediate encrypted result of the classification probability density. Subsequently, it collects the current network link state parameters and encodes them as a network state fingerprint. The intermediate encrypted result and the network state fingerprint are then encapsulated to generate the target transmission frame.

[0060] During implementation, Bank A's technical team discovered that, compared to traditional plaintext transmission or ordinary encryption methods, the method of this invention significantly improves the security and inference accuracy of processing sensitive financial data in untrusted network environments. Traditional methods cannot retain the statistical characteristics of data in ciphertext for deep learning and are difficult to detect the risk of physical tampering in the transmission link. In contrast, the method of this invention effectively achieves the binding of data "usable but invisible" with the physical security of the transmission path through homomorphic encryption and ciphertext adaptive evolution based on physical resonance characteristics.

[0061] To further verify the actual performance of the method of the present invention, Bank A conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1: Table 1. Comparison of Data Security Processing Performance of Bank A's Interbank Joint Risk Control System

[0062] As shown in Table 1, the performance of the cross-bank joint risk control system was comprehensively improved after applying the method of this invention. The accuracy of encrypted inference increased from 85.2% using traditional methods to 96.5%, and the noise tolerance of homomorphic operations significantly improved from 20.5% to 65.8%, significantly increasing the feature utilization rate of encrypted data and providing a reliable basis for the joint risk control model. The transmission path tampering detection rate increased from 45.0% to 99.2%, effectively avoiding the data tampering risk caused by man-in-the-middle attacks. The network fluctuation adaptation latency was significantly reduced from 200 milliseconds to 25 milliseconds, significantly enhancing the system's timeliness. Furthermore, the system throughput increased from 500 TPS to 1200 TPS, and the compliance audit pass rate increased from 92.0% to 100.0%, significantly reducing manual operation and maintenance costs. The number of interventions by operation and maintenance personnel was also significantly reduced from 15 times / day to 1 time / day.

[0063] Through the method of this invention, Bank A successfully achieved privacy protection and encrypted intelligent computation in the process of cross-institutional data sharing, effectively eliminating the risks of data silos and privacy leaks, ensuring the secure flow of financial data, significantly improving the intelligence and automation level of the joint risk control system, significantly reducing the workload of compliance auditors, enhancing the stability and robustness of the risk control system, and providing strong technical support for secure data interaction under the digital transformation of finance.

[0064] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A network and information security encryption method based on data security, characterized in that, Includes the following steps: S1. Collect hardware fingerprint data of edge nodes and perform non-linear hash mapping and scrambling operations to generate a master key. Use the master key to initialize the homomorphic encryption parameter set and convert it into a binary security beacon for external broadcast. S2. Obtain the public key based on the binary security beacon, perform homomorphic encryption on the original sensitive data to generate the initial ciphertext data packet, and encapsulate the homomorphic operation auxiliary parameters into the ciphertext header structure; S3. Receive the initial ciphertext data packet and verify its integrity. Parse the ciphertext header to obtain the homomorphic noise budget value and operation type identifier, so as to determine whether the homomorphic noise budget value meets the executable requirements of homomorphic operation. S4. If the homomorphic noise budget value meets the executability requirements for performing homomorphic operations, the initial ciphertext data packet is input into the improved EfficientNet-B0 model without data decryption. A resonance frequency selection mechanism based on physical resonance is introduced to perform discrete recursive convolution. Based on the feature statistical distribution, dynamic collaborative threshold, scaling factor and distribution parameters are used to implement ciphertext adaptation feature flow adaptive evolution and classification, and output intermediate result ciphertext. S5. Obtain the current network link status parameters and encode them into a network status fingerprint. Encapsulate the intermediate result ciphertext with the network status fingerprint to generate the target transmission frame. S6. Transmit the target transmission frame through an encrypted tunnel, obtain the pre-shared session key and establish a decryption channel, parse the network state fingerprint in the target transmission frame and verify the security of the transmission path. S7. After verifying the security of the transmission path, obtain the private key and decrypt the ciphertext of the intermediate result verified by the path to obtain the plaintext of the operation result. Verify the correctness of the plaintext of the operation result based on the encapsulated homomorphic operation auxiliary parameters, and generate the final audit data.

2. The network and information security encryption method based on data security according to claim 1, characterized in that, S1 specifically includes: S11. Collect edge node hardware fingerprint data, perform non-linear hash mapping and scrambling operations on the edge node hardware fingerprint data, and generate a master key; S12. Perform homomorphic encryption parameter set initialization calculation based on the master key to obtain the homomorphic encryption parameter set; S13. Convert the homomorphic encryption parameter set to binary format to generate a binary security beacon.

3. The network and information security encryption method based on data security according to claim 1, characterized in that, S2 specifically includes: S21. Parse the binary security beacon to extract the homomorphic encryption parameter set, and calculate and extract the public key based on the homomorphic encryption parameter set; S22. Obtain the original sensitive data, and use the public key to perform homomorphic encryption on the original sensitive data to generate the initial ciphertext data; S23. Extract homomorphic operation auxiliary parameters based on the homomorphic encryption parameter set, perform formatted encoding on the homomorphic operation auxiliary parameters, and generate a ciphertext header structure; S24. Using the ciphertext header structure as a preamble and the initial ciphertext data as the payload, the ciphertext header structure and the initial ciphertext data are sequentially concatenated to generate the initial ciphertext data packet.

4. The network and information security encryption method based on data security according to claim 1, characterized in that, S3 specifically includes: S31. Receive the initial ciphertext data packet, parse the header structure of the initial ciphertext data packet to extract the integrity check code, and use the integrity check code to perform integrity check calculation on the initial ciphertext data packet. S32. If the integrity verification calculation passes, parse the header structure of the initial encrypted data packet to extract the homomorphic noise budget value and operation type identifier, and query the homomorphic encryption parameter set based on the operation type identifier to obtain the corresponding type of noise consumption estimate value. S33. Calculate the difference between the homomorphic noise budget value and the noise consumption estimate value as the remaining noise margin. If the remaining noise margin is greater than zero, it is determined that the homomorphic noise budget value meets the executability requirements of homomorphic operation.

5. A network and information security encryption method based on data security according to claim 1, characterized in that, The improved EfficientNet-B0 model includes an initial feature extraction layer, an inverse residual feature enhancement layer, a resonance frequency selection layer, a multi-scale feature fusion layer, and an adaptive classification output layer. The initial feature extraction layer is used to receive the initial encrypted data packet, extract the basic spatial geometric features through multidimensional convolutional mapping and feature normalization, and output the initial feature map. The inverse residual feature enhancement layer is used to receive the initial feature map, mine semantics and expand the dimension using deep convolution operation, construct a nonlinear operator for ciphertext adaptation to perform approximation transformation, fuse nonlinear features with linear channel features, and output the inverse residual feature tensor. The resonant frequency selection layer is used to introduce a resonant frequency selection mechanism based on physical resonance, and the specific execution process includes: Each channel of the inverse residual characteristic tensor is mapped to a second-order damped harmonic oscillator system. The natural frequency and damping ratio of the harmonic oscillator are dynamically calculated based on the statistical distribution of the channel energy and mapped to the coefficients of the second-order difference equation to construct a discrete recursive operator. The discrete recursive operator is used to perform recursive convolution operation on the inverse residual feature tensor in a sliding window to output the initial resonance response features; the frequency response benchmark is determined based on the natural frequency, and the response matching deviation between each local spatial feature in the initial resonance response features and the frequency response benchmark is calculated. The dynamic selection threshold is calculated based on the statistical distribution characteristics of the response matching deviation. Local spatial features with response matching deviations less than the dynamic selection threshold are selected as effective signal components. An adaptive gain factor is generated based on the response matching deviation to weight and amplify the effective signal components, and a pure resonance enhancement feature tensor is output. The multi-scale feature fusion layer is used to receive the resonant enhancement feature tensor, integrate multi-scale context information using cascaded fusion operations, calculate composite scaling coefficients based on the statistical feature distribution of the resonant enhancement feature tensor to adaptively coordinate the number of channels and spatial size, and output a global semantic high-dimensional feature map. The adaptive classification output layer is used to reduce the dimensionality of the high-dimensional feature map to feature descriptors, perform online statistics based on the statistical feature distribution of the feature descriptors to construct dynamic distribution parameters, adaptively project them to the real-time class boundary, and output the intermediate result ciphertext of the classification probability density.

6. A network and information security encryption method based on data security according to claim 1, characterized in that, S5 specifically includes: S51. Collect the status parameters of the current network link, analyze the statistical distribution characteristics of the status parameters, perform hash encoding operation based on the statistical distribution characteristics, and generate network status fingerprint; S52. Obtain the intermediate result ciphertext, map the network state fingerprint to the frame header identifier, and map the intermediate result ciphertext to the frame body payload. S53. Sequentially concatenate and encapsulate the frame header identifier and frame body payload to generate the target transmission frame.

7. A network and information security encryption method based on data security according to claim 1, characterized in that, S6 specifically includes: S61. Receive the target transmission frame, retrieve the pre-shared session key stored locally and construct the decryption context, perform frame structure parsing on the target transmission frame, and output the frame header identifier and frame body payload. S62. Parse the frame header identifier to extract the reference link statistical feature vector recorded by the sending end; S63. Collect the current network link status parameters of the receiving end, including round-trip time jitter sequence, data packet arrival interval distribution and instantaneous link throughput; S64. Construct a sliding time window, perform statistical analysis on network link state parameters, and extract the mean, variance, and kurtosis index; map the mean, variance, and kurtosis index to a vector space to construct a local high-dimensional statistical feature vector; S65. Calculate information entropy based on high-dimensional statistical feature vectors as network fluctuation features, and cascade and fuse network fluctuation features with high-dimensional statistical feature vectors to generate feature data to be verified. S66. Encapsulate the feature data to be verified into a feature vector to be verified, and use it as a local network state fingerprint. S67. Track the historical statistical distribution characteristics of the local network status fingerprint within the sliding time window, and dynamically construct an adaptive confidence interval based on the dispersion of the historical statistical distribution characteristics; calculate the feature deviation between the local network status fingerprint and the baseline link statistical feature vector. If the feature deviation falls within the adaptive confidence interval, the transmission path security verification is deemed successful, and the intermediate result ciphertext verified by the path is output.

8. A network and information security encryption method based on data security according to claim 1, characterized in that, Specifically, S7 includes: S71. Retrieve the local private key and construct a decryption context, obtain the path-verified intermediate result ciphertext, and use the decryption context to perform a homomorphic decryption operation on the path-verified intermediate result ciphertext to generate the operation result plaintext. S72. Parse the encapsulation structure of the intermediate result ciphertext verified by the path to extract the auxiliary parameters of the homomorphic operation, including the initial noise distribution characteristics, the polynomial ring order and the check hash value. S73. Based on the initial noise distribution characteristics and the order of the polynomial ring, construct an adaptive noise tolerance boundary and define the adaptive noise tolerance boundary as a correctness verification constraint. S74. Calculate the actual statistical distribution characteristics of the plaintext of the operation result and the real-time hash value, substitute the actual statistical distribution characteristics into the correctness verification constraints for matching, and compare the real-time hash value with the verification hash value; S75. If the actual statistical distribution characteristics are within the adaptive noise tolerance boundary and the real-time hash value is consistent with the verification hash value, output the verification result identifier that the verification has passed; otherwise, output the verification result identifier that the verification has failed. S76. Filter the plaintext of the operation result according to the verification result identifier, associate and encapsulate the plaintext of the verified operation result with the corresponding homomorphic operation auxiliary parameters, and generate the final audit data.

Images