CPE anti-recognition privacy protection system based on lightweight homomorphic encryption security large model

Abstract

The invention discloses a CPE anti-recognition privacy protection system based on a lightweight homomorphic encryption security large model. Relates to the cross technical field of network security and privacy calculation. Comprising a lightweight homomorphic encryption module, a scattering invariance neural network identification module and a quantitative perception training module. The lightweight homomorphic encryption module processes the obtained original CPE data into a ciphertext, and inputs the ciphertext into the scattering invariance neural network identification module; the scattering invariance neural network identification module extracts and identifies the ciphertext and inputs the ciphertext as feature data to the quantitative perception training module; and the quantitative perception training module performs quantitative processing on the feature data, performs analog quantization and training optimization according to the ciphertext in the lightweight homomorphic encryption module, and adjusts parameters and a quantization strategy of the model. According to the method, organic unification of privacy security, anti-attack capability and operation efficiency can be realized, and a systematic solution is provided for security deployment of CPE identification.

Description

Technical Field

[0001] The present invention relates to the cross - technical field of network security and privacy computing, and more specifically, to a CPE anti - recognition privacy protection system based on a lightweight homomorphic encryption secure large model. Background Art

[0002] The Common Platform Enumeration (CPE) technology is an international general software asset identification standard for software supply chain security management. It can accurately locate software components and their version information, provide support for vulnerability correlation analysis, etc., and improve the vulnerability response and threat warning capabilities. However, the existing CPE technology deployment faces three constraints: data privacy, malicious attack interference, and model complexity adaptation.

[0003] In recent years, deep learning technology has brought new opportunities to the software security field. Based on the automated feature learning mechanism of deep learning, it mines the correlation rules of software metadata and improves the accuracy of software version identification. The intelligent recognition system based on large models also shows multi - modal processing advantages. However, the CPE intelligent recognition technology faces severe security challenges. Attackers construct malicious interference data, resulting in inaccurate recognition and delayed warning. The existing defense solutions have limited protection effects and will also increase the model complexity and computational overhead, making it difficult to meet the requirements of privacy protection, anti - interference, and edge device adaptation at the same time.

[0004] The adversarial defense technology in the field of artificial intelligence provides a direction for solving the privacy protection problem of the CPE recognition system. Technologies represented by adversarial training and data purification enhance the survival ability of the model under malicious attacks and have adaptation advantages in the CPE recognition scenario. However, when the existing defense solutions are engineered and implemented, there are problems of privacy leakage and difficulty in meeting the lightweight requirements of edge nodes.

[0005] Therefore, it is an urgent problem for those skilled in the art to propose a privacy protection defense system for the CPE recognition scenario and achieve the collaborative optimization of data security, anti - interference ability, and computational efficiency through three - stage technological innovation. Summary of the Invention

[0006] In view of this, the present invention provides a CPE anti - recognition privacy protection system based on a lightweight homomorphic encryption secure large model. This system overcomes the compatibility problem between encrypted computing and intelligent recognition, and realizes the organic unity of privacy security, anti - attack ability, and operation efficiency in scenarios such as industrial Internet of Things, providing a systematic solution for the secure deployment of CPE recognition.

[0007] To achieve the above object, the present invention adopts the following technical solutions: A CPE anti - recognition privacy protection system based on a lightweight homomorphic encryption secure large model, comprising: a lightweight homomorphic encryption module, a scattering - invariant neural network recognition module, and a quantization - aware training module;

[0008] After processing the obtained original CPE data into ciphertext, the lightweight homomorphic encryption module inputs it into the scattering invariance neural network recognition module;

[0009] After extracting and recognizing the ciphertext, the scattering invariance neural network recognition module inputs the result as feature data into the quantization-aware training module;

[0010] The quantization-aware training module performs quantization processing on the feature data, and performs simulated quantization and training optimization according to the ciphertext in the lightweight homomorphic encryption module to adjust the parameters and quantization strategy of the model.

[0011] Preferably, the lightweight homomorphic encryption module includes a key generation layer, a key encryption layer, a homomorphic operation layer, a homomorphic decryption layer, and a lightweight optimization layer connected in sequence;

[0012] The key generation layer is used to generate the public key, private key, and evaluation key required for homomorphic encryption, and dynamically optimize the key parameter configuration;

[0013] The key encryption layer is used to introduce a block processing strategy, perform block coding on the unstructured original CPE data, and encrypt it in parallel;

[0014] The homomorphic operation layer is used to perform addition and multiplication homomorphic operations in the ciphertext state;

[0015] The homomorphic decryption layer is used to decrypt the ciphertext calculation result into plaintext output, verify the data integrity, and decrypt the original CPE data;

[0016] The lightweight optimization layer constructs a sparse polynomial ring based on the Ring-LWE problem and protects the privacy of the sparse structure through random masking technology.

[0017] Preferably, the key generation layer generates the public key, private key, and evaluation key based on the Paillier encryption method.

[0018] Preferably, the scattering invariance neural network recognition module includes a multi-scale scattering transform layer, an invariance feature learning layer, and a feature fusion enhancement layer connected in sequence;

[0019] The multi-scale scattering transform layer performs multi-scale decomposition on the input ciphertext, constructs a feature basis with translational and rotational invariance, and obtains multi-scale scattering coefficients, which provide a basic feature representation for the invariance feature learning layer;

[0020] The invariance feature learning layer learns and optimizes the feature data according to the multi-scale scattering coefficients, and transmits the optimized feature data to the feature fusion enhancement layer;

[0021] The feature fusion enhancement layer fuses the optimized feature data at different levels based on the multi-scale scattering coefficient.

[0022] Preferably, the multi-scale scattering transformation layer uses a dynamic wavelet basis set to perform multi-scale decomposition on the input ciphertext to extract low-frequency features and high-frequency details with geometric invariance; and constructs a feature expression base with translation and rotation invariance through hierarchical scattering coefficient calculation;

[0023] The invariant feature learning layer introduces a geometric transformation data augmentation strategy to improve the model's robustness to input perturbations through adversarial training; a feature stability loss function is deployed to constrain the geometric invariance of the network output;

[0024] The feature fusion enhancement layer performs cross-scale fusion on the multi-scale scattering coefficients, strengthens the key feature channels through the attention weighting mechanism, and implements feature space orthogonalization processing to eliminate redundant features and improve feature discrimination.

[0025] Preferably, the quantization-aware training module includes a quantization strategy control layer and a quantization training optimization layer. The quantization strategy control layer is used to dynamically analyze the parameter distribution characteristics of each layer of the model, analyze the impact of quantization error on model accuracy based on gradient propagation, give priority to protecting the accuracy of key layers, and realize the compression of the progressive CPE identification model with controllable accuracy loss;

[0026] The quantization training optimization layer simulates quantization noise during the training phase, introduces dynamic quantization noise in back propagation, and enhances the model's adaptability to low-bitwidth calculations through pseudo-quantization operations; implements gradient rescaling and calibration technology, and maintains training stability through gradient amplitude normalization and direction correction.

[0027] Through the above technical solutions, it can be seen that compared with the existing technology, the present invention discloses a CPE anti-identification privacy protection system based on a lightweight homomorphic encryption security large model. The advantages are: the present invention innovatively constructs a technical system for collaborative optimization of "encryption-defense-efficiency", and realizes the full-process confidential processing of data through lightweight homomorphic encryption, breaking through the bottleneck that traditional encryption technology cannot support ciphertext model training; integrating the geometric invariance characteristics of the scattering invariant network and the dynamic defense mechanism, significantly improving the model's anti-interference ability and adaptability to complex environments; and creating an original model lightweight compression optimization strategy based on a quantization-aware training framework to achieve a coordinated improvement in model volume compression and inference efficiency. This solution systematically solves the compatibility problem between encrypted computing and intelligent identification, achieves an organic unity of privacy protection strength, anti-attack capability and operational efficiency in edge computing scenarios, and provides feasible technical support for building a zero-trust security architecture. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on the provided drawings.

[0029] Figure 1 Schematic diagram of the CPE anti-identification privacy protection system based on the lightweight homomorphic encryption secure large model provided by the present invention;

[0030] Figure 2 Schematic diagram of the structure of the lightweight homomorphic encryption module provided by the present invention;

[0031] Figure 3 Schematic diagram of the structure of the scattering invariance neural network recognition module provided by the present invention;

[0032] Figure 4 Schematic diagram of the structure of the quantization-aware training module provided by the present invention. Detailed implementation manners

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0034] As Figure 1 shown, the embodiments of the present invention disclose a CPE anti-identification privacy protection system based on a lightweight homomorphic encryption secure large model, including: a lightweight homomorphic encryption module, a scattering invariance neural network recognition module, and a quantization-aware training module;

[0035] The lightweight homomorphic encryption module processes the obtained original CPE data into ciphertext and then inputs it into the scattering invariance neural network recognition module;

[0036] The scattering invariance neural network recognition module extracts and recognizes the ciphertext and then inputs it as feature data into the quantization-aware training module;

[0037] The quantization-aware training module performs quantization processing on the feature data, and performs simulated quantization and training optimization according to the ciphertext in the lightweight homomorphic encryption module to adjust the parameters and quantization strategies of the model.

[0038] As Figure 2As shown, specifically, the lightweight homomorphic encryption module includes a key generation layer, a key encryption layer, a homomorphic operation layer, a homomorphic decryption layer, and a lightweight optimization layer connected in sequence;

[0039] The key generation layer is used to generate the public key, private key, and evaluation key required for homomorphic encryption, and dynamically optimize the key parameter configuration;

[0040] The key encryption layer is used to introduce a block processing strategy, perform block encoding on the unstructured original CPE data, and encrypt it in parallel;

[0041] The homomorphic operation layer is used to perform addition and multiplication homomorphic operations in the ciphertext state;

[0042] The homomorphic decryption layer is used to decrypt the ciphertext calculation result into a plaintext output, verify the data integrity, and decrypt the original CPE data;

[0043] The lightweight optimization layer constructs a sparse polynomial ring based on the Ring-LWE problem and protects the privacy of the sparse structure through a random masking technique.

[0044] Specifically, the key generation layer generates the public key, private key, and evaluation key based on the Paillier encryption method.

[0045] In a specific embodiment, the specific steps for generating the public key, private key, and evaluation key based on the Paillier encryption method are as follows:

[0046] Select two large prime numbers p and q, calculate n = p × q and λ = lcm(p - 1, q - 1), where lcm represents the least common multiple.

[0047] Select a random integer g ∈ Z n * , that is, select an element from the multiplicative group modulo n. Then, calculate the function f, which is usually related to g and n. For example, it can be f = g n mod n 2 . The calculation result of this function will be used for subsequent decryption operations.

[0048] Set the private key as sk = (λ, f), and the public key as pk = (n, g). These two keys will be used for decryption and encryption operations respectively.

[0049] In another specific embodiment provided by the present invention, the key encryption layer encrypts the plaintext data using the public key to generate ciphertext. The encryption process must ensure the security of the ciphertext during transmission and storage. Even if intercepted by an attacker, no useful information can be extracted from it. At the same time, to meet the characteristics of homomorphic encryption, the encryption process needs to preserve the closure of the ciphertext for specific mathematical operations, that is, the encrypted ciphertext can directly perform certain operations without being decrypted first. The specific encryption process is as follows:

[0050] For the plaintext m ∈ Z to be encrypted n , first randomly select an integer r from Z n * as the random number in the encryption process. The introduction of this random number is to increase the unpredictability of the ciphertext and improve the security of the encryption scheme.

[0051] Calculate the ciphertext c = g m × r n mod n 2 . This formula uses the parameters g and n in the public key to combine the plaintext m and the random number r through modular exponentiation to generate the final ciphertext c. This ciphertext can directly participate in specific mathematical operations in the subsequent homomorphic operation layer.

[0052] In another specific embodiment provided by the present invention, the homomorphic operation layer directly performs specific mathematical operations, such as addition or multiplication, on the ciphertext without touching the plaintext data. This feature enables meaningful calculations to be performed on encrypted data while protecting data privacy, expanding the application scenarios of data processing. The specific implementation of the homomorphic addition operation is as follows:

[0053] 1. Suppose there are two ciphertexts and which are the encryption results corresponding to the plaintexts m1 and m2 respectively.

[0054] 2. Directly perform a multiplication operation on these two ciphertexts to obtain a new ciphertext c sum = c1 × c2 mod n 2 . According to the properties of Paillier encryption, this new ciphertext actually corresponds to the encryption result of the plaintext m1 + m2, that is This shows that the multiplication operation in the ciphertext space is equivalent to the addition operation in the plaintext space, thus realizing homomorphic addition.

[0055] In another specific embodiment provided by the present invention, the homomorphic decryption layer converts the ciphertext after homomorphic operations back to the original plaintext data using the private key. The decryption process must be precise and efficient to ensure the correct plaintext is recovered from the ciphertext, while ensuring the security of the data throughout the process and preventing system vulnerabilities caused by the leakage of the private key. The specific decryption process is as follows:

[0056] For the ciphertext c to be decrypted, first calculate c λ mod n 2 , and then process it using the function L(u) = u - 1 / n to obtain L(c λ mod n 2 ).

[0057] Similarly, calculate g λ mod n 2 and apply the function L to obtain L(g λ mod n 2 ).

[0058] Finally, by calculating the original plaintext m can be recovered. This process utilizes the parameters λ and f in the private key, and through modular arithmetic and the processing of the function L, maps the ciphertext back to the plaintext space.

[0059] In another specific embodiment provided by the present invention, the lightweight optimization layer optimizes the computational and storage resource requirements of the homomorphic encryption model, making it more suitable for running on devices with limited computing power. By simplifying the steps of the encryption and decryption algorithms, the operating efficiency and response speed of the system can be improved without significantly reducing security, thereby expanding the practical application scope of homomorphic encryption technology. When performing consecutive homomorphic addition of multiple ciphertexts, a cumulative calculation method can be adopted to avoid performing a complete modular arithmetic operation for each operation. For example, for multiple ciphertexts c1, c2, …, c k , the addition can be cumulatively calculated in the integer domain first:

[0060]

[0061] Then, when the final result is needed, a unified modular arithmetic operation is performed:

[0062] C final = C sum mod n 2

[0063] This can reduce the number of intermediate modular arithmetic operations and lower the computational complexity. Through the application of these mathematical formulas, the lightweight optimization layer can effectively improve the computational efficiency and resource utilization efficiency of the lightweight homomorphic encryption model, making it more suitable for use in resource-constrained environments.

[0064] Specifically, the scattering invariance neural network recognition module includes a multi-scale scattering transform layer, an invariance feature learning layer, and a feature fusion enhancement layer connected in sequence;

[0065] The multi-scale scattering transform layer performs multi-scale decomposition on the input ciphertext, constructs a feature basis with translational and rotational invariance, and obtains multi-scale scattering coefficients, which provide a basic feature representation for the invariance feature learning layer;

[0066] The invariance feature learning layer learns and optimizes the feature data according to the multi-scale scattering coefficients, and transmits the optimized feature data to the feature fusion and enhancement layer;

[0067] The feature fusion and enhancement layer fuses the optimized feature data at different levels based on the multi-scale scattering coefficients.

[0068] Specifically, the feature fusion and enhancement layer fuses the features extracted by the "multi-scale scattering transform layer" and the "invariance feature learning layer" to form a more comprehensive and robust feature representation. This fused feature representation not only retains the global structure of the data but also has invariance to various transformations, thereby enhancing the expressive and discriminative abilities of the features.

[0069] Specifically, the multi-scale scattering transform layer uses a dynamic wavelet basis set to perform multi-scale decomposition on the input ciphertext, extracts low-frequency features and high-frequency details with geometric invariance; through hierarchical scattering coefficient calculation, constructs a feature expression basis with translational and rotational invariance;

[0070] The invariance feature learning layer introduces a geometric transformation data augmentation strategy to improve the robustness of the model to input perturbations through adversarial training; deploys a feature stability loss function to constrain the geometric invariance of the network output;

[0071] The feature fusion and enhancement layer performs cross-scale fusion on the multi-scale scattering coefficients, strengthens the key feature channels through an attention weighting mechanism; implements feature space orthogonalization processing to eliminate redundant features and improve feature discriminability.

[0072] In another specific embodiment provided by the present invention, the multi-scale scattering transform layer can perform multi-scale and multi-directional feature extraction on the input signal based on the wavelet scattering transform theory. This layer captures the local and global features of the signal by simulating the fluctuation and scattering behavior of the signal at different scales. These features have certain invariance and robustness to noise and geometric deformations of the target, providing a solid foundation for subsequent feature processing and classification.

[0073] Specifically, as Figure 3As shown, the multi-scale scattering transform layer includes a physics model-driven feature extraction module, a data-driven high-dimensional feature compression module, and a pattern recognition module. The CPE data to be recognized is input into the physics model-driven feature extraction module for processes such as convolution and filtering, then undergoes global average pooling through the data-driven high-dimensional feature compression module, and finally is input into the pattern recognition module to output the CPE recognition result.

[0074] Specifically, the multi-scale scattering transform layer performs a series of convolution and modulus operations on the input signal to obtain scattering features at different scales. These features can effectively represent the essential characteristics of the signal and remain relatively stable even in the presence of noise interference. The mathematical formula for wavelet scattering transform is:

[0075] S j [f](t) = |f * ψ j |(t)

[0076] where, S j [f](t) represents the scattering feature at the j-th scale, f is the input signal, ψ j is the wavelet function at the j-th scale, * represents the convolution operation, and || represents taking the absolute value. This formula describes how to obtain the scattering features at different scales through the convolution of the wavelet function and the input signal. Further, the multi-scale scattering transform can be expressed as:

[0077]

[0078] where, j1, j2, …, j n represents the combination of different scales. By performing convolution and modulus operations layer by layer, scattering features at multiple scales are obtained, that is, feature data that is robust to local transformations. These features can capture the local fluctuations and structural information of the signal at different scales, providing rich basic data for subsequent feature fusion and classification.

[0079] In another specific embodiment provided by the present invention, the invariance feature learning layer further extracts feature representations that are invariant to noise and target changes through the learning ability of a deep neural network. This layer usually consists of multiple neural network layers, such as convolution layers, pooling layers, and fully connected layers, etc. Through non-linear transformation and parameter learning, the features extracted by the multi-scale scattering transform layer are gradually transformed into higher-level abstract representations.

[0080] These representations can better capture the essential features of the target and have stronger robustness to noise and geometric changes. Specifically, the network automatically adjusts parameters by learning a large number of noisy training samples, so that the extracted features remain consistent under different noise conditions and target pose changes, thereby improving the anti-noise ability and recognition performance of the model. Assuming the input feature is Fin The output feature is F out If the parameters of the convolutional layer are the weight matrix W and the bias vector b, then the convolutional operation can be expressed as:

[0081] F out = σ(W·F in + b)

[0082] where σ is the ReLU activation function. This formula describes how to extract the non - linear representation of features through convolutional operations. During training, the network parameters are optimized by minimizing the loss function, and the corresponding cross - entropy loss can be expressed as:

[0083]

[0084] where, y i is the true label, is the probability output predicted by the network, and N is the number of classes. The gradient is calculated through the backpropagation algorithm, and the Adam optimization algorithm is used to update the model parameters:

[0085]

[0086] where η is the learning rate, and are the gradients of the loss function with respect to the weight and bias respectively. The above formulas together describe the process of gradually optimizing the feature representation through deep learning methods to make it have stronger invariance and robustness of CPE features.

[0087] Specifically, the quantization - aware training module includes a quantization strategy control layer and a quantization training optimization layer. The quantization strategy control layer is used to dynamically analyze the parameter distribution characteristics of each layer of the model, analyze the impact of quantization error on the model accuracy based on gradient propagation, and preferentially protect the accuracy of key layers to achieve the compression of the progressive CPE recognition model with controllable accuracy loss;

[0088] Specifically, the parameters of each layer refer to the weights and activation values of each layer of the scattering - invariance neural network recognition module.

[0089] In another specific embodiment provided by the present invention, as Figure 4 shown, the quantization training optimization layer simulates quantization noise during the training stage, introduces dynamic quantization noise in backpropagation, and enhances the model's adaptability to low - bit - width calculations through pseudo - quantization operations; implements gradient rescaling and calibration techniques to maintain training stability through gradient magnitude normalization and direction correction.

[0090] Further, in the quantization strategy control layer, it is mainly responsible for inserting pseudo-quantization nodes (FakeQuantize Nodes) during the model training process. These nodes simulate the quantization process, enabling the model to gradually adapt to low-precision representations. By statistically analyzing the data ranges of the input and output during training, the quantization strategy control layer can help determine the quantization parameters, thereby maintaining the privacy protection performance of the CPE-identified secure large model as much as possible after quantization. Specifically, during the forward propagation of the model, it will perform range statistics on key data such as activation values and weights, providing the necessary parameter basis for subsequent quantization operations. Assume the floating-point value range is r min , r max , and the quantization target is an n-bit integer (usually n = 8), then the quantization formula is:

[0091]

[0092] where S is the scaling factor, which determines the floating-point value span of each quantization step and realizes quantization by mapping the floating-point value range to the integer range. Z is the zero-point offset, ensuring that the floating-point zero value can be accurately mapped to the integer value, thereby reducing the error caused by quantization to a certain extent. Q is the quantized integer value, and R is the original floating-point value.

[0093] In the quantization training optimization layer, it mainly focuses on how to optimize the quantization parameters and model weights during the training process to minimize the accuracy loss caused by quantization. The optimization strategies at this layer include methods such as adjusting the learning rate and using mixed-precision training. Through a reasonable learning rate scheduling strategy, it can ensure the stable update of the quantization parameters and model weights during training, thereby gradually improving the performance of the model after quantization. In addition, the mixed-precision training method can balance the accuracy and computational efficiency of the model during training. By using different precision representations for data in different layers or different parts, the model can be trained and inferred more efficiently while maintaining a relatively high accuracy. To stabilize the learning of quantization parameters, a phased learning rate scheduling strategy is adopted, and the specific scheduling strategy is as follows:

[0094] Initial stage (0 - 30%): Set the learning rate to a relatively low value, such as η1 = 1×10 -5 . At this time, the model is just starting to adapt to quantization, and the low learning rate helps to stably update the quantization parameters.

[0095] Middle stage (30% - 70%): Gradually increase the learning rate to of the base value, that is to accelerate the model convergence speed.

[0096] Late stage (70% - 100%): Use the cosine decay formula to adjust the learning rate, making the learning rate gradually decrease to help the model fine-tune:

[0097]

[0098] Among them, η base is the base learning rate, t is the current training epoch, and T mid is the end epoch of the middle stage, and T end is the total number of training epochs.

[0099] In addition, in mixed-precision training, the FP16 precision is retained for sensitive layers, and other layers are quantized to balance precision and computational efficiency. The corresponding weight update formula is:

[0100]

[0101] Among them, θ t is the model weight at the t-th epoch, η is the learning rate, is the gradient of the loss function with respect to the weight. During training, quantization parameters (such as the scaling factor S and the zero-point offset Z) also need to be updated. Taking the scaling factor S as an example, its update formula is:

[0102]

[0103] Among them, α is the learning rate, and are the estimated maximum and minimum values of the current batch of data, and n is the number of quantization bits. This formula adjusts the scaling factor by minimizing the quantization error, making the quantized data more accurately represent the original floating-point data.

[0104] This strategy is based on the different sensitivities of different layers to quantization errors. By retaining higher precision for sensitive layers, it protects the key information of the model from being damaged by excessive quantization. At the same time, it quantizes the relatively less sensitive layers, reducing the overall consumption of computing resources and improving the efficiency of training and inference.

[0105] The core innovation of the present invention lies in: ① constructing a collaborative architecture of encrypted computing and intelligent recognition to achieve synchronous optimization of feature parsing and attack defense in the ciphertext state; ② designing a feature extraction network with geometric invariance to suppress adversarial perturbations through multi-scale scattering transformation; ③ developing a lightweight model adapted to encrypted domain computing to break through the resource constraints of edge devices. Compared with traditional solutions, the present invention realizes the coordinated improvement of privacy protection intensity, anti-attack ability, and operation efficiency.

[0106] It should be noted that the implementation details of the present invention can be adjusted according to specific application scenarios, including but not limited to encryption algorithm parameter configuration, neural network layer design, and quantization strategy selection. The combination and optimization of the technical features all fall within the protection scope of the present invention, and any improvement scheme based on the core architecture of the present invention should be included in the scope of the claims.

[0107] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For related parts, reference can be made to the description in the method section.

[0108] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A CPE anti-identification privacy protection system based on a lightweight homomorphic encryption secure large model, characterized in that It includes: a lightweight homomorphic encryption module, a scattering invariance neural network recognition module, and a quantization-aware training module; After processing the acquired original CPE data into ciphertext, the lightweight homomorphic encryption module inputs it into the scattering invariance neural network recognition module; After extracting and recognizing the ciphertext, the scattering invariance neural network recognition module inputs the result as feature data into the quantization-aware training module; The quantization-aware training module performs quantization processing on the feature data, and performs simulated quantization and training optimization according to the ciphertext in the lightweight homomorphic encryption module to adjust the parameters and quantization strategy of the model.

2. The CPE anti-identification privacy protection system based on a lightweight homomorphic encryption secure large model according to claim 1, wherein The lightweight homomorphic encryption module includes a key generation layer, a key encryption layer, a homomorphic operation layer, a homomorphic decryption layer, and a lightweight optimization layer connected in sequence; The key generation layer is used to generate the public key, private key, and evaluation key required for homomorphic encryption, and dynamically optimize the key parameter configuration; The key encryption layer is used to introduce a block processing strategy to perform block coding on the unstructured original CPE data and encrypt it in parallel; The homomorphic operation layer is used to perform addition and multiplication homomorphic operations in the ciphertext state; The homomorphic decryption layer is used to decrypt the ciphertext calculation result into plaintext output, verify the data integrity, and decrypt the original CPE data; The lightweight optimization layer constructs a sparse polynomial ring based on the Ring-LWE problem and protects the privacy of the sparse structure through the random masking technique.

3. The CPE anti-identification privacy protection system based on a lightweight homomorphic encryption secure large model according to claim 2, wherein The key generation layer generates the public key, private key, and evaluation key based on the Paillier encryption method.

4. A CPE anti-identification privacy protection system based on a lightweight homomorphic encryption secure large model according to claim 1, characterized in that, The scattering invariance neural network recognition module includes a multi-scale scattering transform layer, an invariance feature learning layer, and a feature fusion enhancement layer connected in sequence; The multi-scale scattering transform layer performs multi-scale decomposition on the input ciphertext, constructs a feature basis with translation and rotation invariance, and obtains multi-scale scattering coefficients, which provide a basic feature representation for the invariance feature learning layer; The invariance feature learning layer learns and optimizes the feature data according to the multi-scale scattering coefficients, and transmits the optimized feature data to the feature fusion enhancement layer; The feature fusion enhancement layer fuses the optimized feature data at different levels based on the multi-scale scattering coefficients.

5. A CPE anti-identification privacy protection system based on a lightweight homomorphic encryption secure large model according to claim 4, characterized in that, The multi-scale scattering transform layer uses a dynamic wavelet basis set to perform multi-scale decomposition on the input ciphertext, and extracts low-frequency features and high-frequency details with geometric invariance; Through hierarchical scattering coefficient calculation, a feature expression basis with translation and rotation invariance is constructed; The invariance feature learning layer introduces a geometric transformation data augmentation strategy to improve the robustness of the model to input perturbations through adversarial training; Deploy a feature stability loss function to constrain the geometric invariance of the network output; The feature fusion enhancement layer performs cross-scale fusion on the multi-scale scattering coefficients, and strengthens the key feature channels through an attention weighting mechanism; Implement feature space orthogonalization processing to eliminate redundant features and improve feature discriminability.

6. The CPE anti-identification privacy protection system based on a lightweight homomorphic encryption secure large model according to claim 1, characterized in that The quantization-aware training module includes a quantization strategy control layer and a quantization training optimization layer. The quantization strategy control layer is used to dynamically analyze the parameter distribution characteristics of each layer of the model, analyze the impact of quantization error on the model accuracy based on gradient propagation, and prioritize protecting the accuracy of key layers to achieve the compression of the progressive CPE recognition model with controllable accuracy loss. The quantization training optimization layer simulates quantization noise during the training phase, introduces dynamic quantization noise in backpropagation, and enhances the model's adaptability to low-bitwidth calculations through pseudo-quantization operations. Implement gradient rescaling and calibration techniques to maintain training stability through gradient magnitude normalization and direction correction.

Images