Security operation method and system based on artificial intelligence

Abstract

The invention discloses a security operation method and system based on artificial intelligence, and relates to the technical field of data security, and the method comprises the steps: collecting original data generated by first terminal equipment, and carrying out the encryption processing of the original data through a preset encryption algorithm, and obtaining encrypted data; constructing a weight matrix on the first terminal equipment, and generating a training sample set according to the encrypted data and the weight matrix; the training sample set is transmitted to a second terminal device, the second terminal device performs fitting operation on the training sample set based on the deep neural network model, and a feature vector is output; according to a comparison result of the feature vector and a preset threshold value, a security scoring model is established, and the security scoring model corresponds to a plurality of security levels; and selecting a corresponding operation strategy according to the security level, and executing differential privacy protection processing on the encrypted data. According to the method, decryption processing in a traditional method is avoided by combining the encrypted data with the artificial intelligence model, so that potential safety hazards are reduced, and the calculation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of data security technology, and in particular to a security computing method and system based on artificial intelligence. Background Art

[0002] With the advent of the information age, data security has become an increasingly pressing issue worldwide. Especially with the rapid development of cloud computing, big data, and the Internet of Things, ensuring the security of data during transmission, storage, and processing has become a common challenge across all industries. To address this issue, various data encryption, authentication, and access control technologies have been proposed and widely adopted. For example, traditional symmetric and asymmetric encryption technologies have been widely used to protect data security during transmission, while differential privacy, as an emerging data privacy protection method, has also been gradually applied in big data analysis and artificial intelligence models. In recent years, the development of artificial intelligence technologies, particularly deep learning algorithms, has further promoted research in data processing and security improvements.

[0003] However, existing secure computing methods often face several challenges. First, while traditional encryption algorithms can ensure the confidentiality of data during transmission, they are limited in their processing and subsequent computing capabilities for encrypted data. In particular, in the further analysis and calculation of encrypted data, decryption operations are often required to restore the original data, which increases computational complexity and security risks. Second, while existing AI-based secure computing technologies can process data through deep learning models, most of them ignore the issue of data privacy protection, especially in multi-party computing and data sharing scenarios, where there is a lack of effective privacy protection mechanisms. In addition, existing security assessment models often focus only on a single security dimension, such as password strength and network access control, without comprehensively considering the privacy and computational security of data processing, making it difficult to comprehensively improve data security. More importantly, when faced with encrypted data, existing secure computing methods are often unable to perform effective analysis and processing directly on encrypted data, leading to a contradiction between secure computing and data analysis. Summary of the Invention

[0004] In view of the problems existing in the above-mentioned automated FOTA testing method for lithium battery systems, the present invention is proposed.

[0005] Therefore, the present invention provides a secure computing method based on artificial intelligence, which can solve the problems mentioned in the background technology.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a secure computing method based on artificial intelligence, which includes:

[0008] Collecting original data generated by the first terminal device, and encrypting the original data using a preset encryption algorithm to obtain encrypted data;

[0009] constructing a weight matrix on the first terminal device, and generating a training sample set according to the encrypted data and the weight matrix;

[0010] Transmitting the training sample set to a second terminal device, where the second terminal device performs a fitting operation on the training sample set based on a deep neural network model and outputs a feature vector;

[0011] Establishing a security scoring model based on the comparison result of the feature vector and a preset threshold, wherein the security scoring model corresponds to multiple security levels;

[0012] A corresponding operation strategy is selected according to the security level, and differential privacy protection processing is performed on the encrypted data.

[0013] As a preferred solution of the artificial intelligence-based security computing method of the present invention, the construction of the weight matrix includes:

[0014] Performing feature decomposition on the encrypted data to extract k main feature components;

[0015] Assigning initial weight coefficients to the k main feature components;

[0016] Iteratively optimizing the initial weight coefficients;

[0017] When the iteration converges, calculate the final weight matrix elements;

[0018] The eigendecomposition is expressed as follows:

[0019]

[0020] Among them, T n (x) is the improved Chebyshev polynomial, n is the order of the polynomial, and its value range is [3,10]. x is the input variable, which represents the normalized eigenvalue. i is the index variable, and α is the adaptive coefficient.

[0021] As a preferred embodiment of the artificial intelligence-based security computing method of the present invention, the initial weight coefficient is calculated by the following formula:

[0022]

[0023] Among them, w i is the initial weight value of the i-th feature, fi is the value of the i-th eigencomponent, β is the regularization parameter, |F|2 is the L2 norm of the eigenvector F, and k is the total number of eigencomponents;

[0024] The iterative optimization of the initial weight coefficient is expressed by the following formula:

[0025]

[0026] in, is the weight value of the i-th feature at the t-th iteration, t is the number of iterations, γ is the learning rate, λ is the decay coefficient, is the loss function, η is the noise coefficient, and ∈ is a random variable that obeys the standard normal distribution.

[0027] As a preferred solution of the artificial intelligence-based secure computing method of the present invention, wherein: the deep neural network model includes an encoding layer, a feature extraction layer and a mapping layer;

[0028] The autoencoder structure of the encoding layer is shown below:

[0029]

[0030] Among them, h1 is the output feature of the encoding layer, n is the input feature dimension, is the first layer weight matrix, σ is the activation function, x i is the i-th component of the input feature vector, and φ(x) is the improved LeakyReLU function:

[0031] φ(x)=max(0.01x,x)+δsin(ωx)

[0032] Here, δ and ω are the amplitude and frequency parameters, respectively.

[0033] As a preferred solution of the artificial intelligence-based security calculation method of the present invention, the security scoring model is shown as follows:

[0034] S total =φ1S1+φ2S2+φ3S3+ξ(S1S2+S2S3+S3S1)

[0035] Among them, φ1, φ2, φ3 are balance coefficients, and ξ is the interaction strength parameter, which is dynamically updated by the following equation:

[0036]

[0037] Where t is the number of iterations, η φ is the learning rate, which controls the parameter update step size, λ φ is the time attenuation coefficient, is the gradient of the loss function with respect to the weight, S1 is the threat level index, S2 is the vulnerability index, and S3 is the risk exposure index.

[0038] As a preferred solution of the artificial intelligence-based security calculation method of the present invention, the threat level index is shown as follows:

[0039]

[0040] Among them, F out,i is the original value of the i-th feature component, λ1 is the feature interaction strength coefficient, ranging from [0.1, 0.5], k1 is the number of features used for threat calculation, ω i is the dynamic weight coefficient of the i-th feature, which is dynamically calculated through the attention mechanism:

[0041]

[0042] in, is the weight vector of the attention layer, with a dimension of d×1, W b is the feature transformation matrix with dimension d×n, b is the bias vector, and tanh is the hyperbolic tangent activation function.

[0043] As a preferred embodiment of the artificial intelligence-based security calculation method of the present invention, the security level is determined according to a dynamic threshold, with a total of seven security levels;

[0044] The dynamic threshold is calculated as follows:

[0045]

[0046] Among them, L level is the multi-classification loss function:

[0047]

[0048] Among them, p i and q i are the probability distributions of the true level and the predicted level, respectively, ρ is the interval constraint coefficient, is the current value of the i-th threshold parameter at time t, and η is the learning rate parameter.

[0049] As a preferred solution of the artificial intelligence-based security computing method of the present invention, wherein:

[0050] In a second aspect, the present invention further addresses the safety issues existing in AI-based safety computing methods. Embodiments provide an AI-based vehicle fault diagnosis and maintenance assessment system, which includes:

[0051] The collection and encryption module is used to collect the original data generated by the first terminal device and encrypt the original data using a preset encryption algorithm to obtain encrypted data;

[0052] A training module, configured to construct a weight matrix on the first terminal device and generate a training sample set based on the encrypted data and the weight matrix;

[0053] a feature vector construction module, configured to transmit the training sample set to a second terminal device, wherein the second terminal device performs a fitting operation on the training sample set based on a deep neural network model and outputs a feature vector;

[0054] A security grading module, configured to establish a security scoring model based on a comparison result between the feature vector and a preset threshold, wherein the security scoring model corresponds to multiple security levels;

[0055] The differential operation module is used to establish a security scoring model based on the comparison result of the feature vector and the preset threshold, and the security scoring model corresponds to multiple security levels.

[0056] In a third aspect, an embodiment of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the artificial intelligence-based security computing method as described in the first aspect of the present invention is implemented.

[0057] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium on which a computer program is stored, wherein: when the computer program is executed by a processor, it implements any step of the artificial intelligence-based security computing method as described in the first aspect of the present invention.

[0058] The beneficial effects of this invention are: first, by combining encrypted data with artificial intelligence models, the decryption process required in traditional methods is avoided, thereby reducing security risks and improving computational efficiency. Second, it innovatively introduces a multi-level security level mechanism and selects appropriate privacy protection strategies for each level, effectively balancing data availability and privacy protection requirements. Finally, this invention considers multiple complex security factors during the processing process, providing flexible protection solutions for data with different security requirements, thereby significantly improving the overall security of the system and data privacy protection capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort. Among them:

[0060] Figure 1 The figure is a flowchart of a secure computing method based on artificial intelligence. DETAILED DESCRIPTION

[0061] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0062] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0063] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0064] Example 1

[0065] Reference Figure 1 , which is the first embodiment of the present invention, provides a security computing method based on artificial intelligence, comprising the following steps:

[0066] S1: Collecting original data generated by a first terminal device, and encrypting the original data using a preset encryption algorithm to obtain encrypted data;

[0067] Collect original data generated by the first terminal device, the original data including user operation log, system operation status data, network transmission data and device performance data; wherein, the user operation log records the user's click events, input content and operation sequence on the first terminal device, the system operation status data includes CPU usage, memory occupancy and disk read and write rate, the network transmission data includes data packet size, transmission time and network delay, and the device performance data includes battery power, temperature and signal strength; then, call the preset encryption algorithm to encrypt the original data, the preset encryption algorithm adopts asymmetric encryption, uses a 2048-bit RSA key to encrypt the original data in blocks, each data block size is 256 bytes, performs group encryption transformation on the original data, and introduces a random padding scheme in the encryption process to obtain encrypted data in this way; the encrypted data is stored in ciphertext form, and each encrypted data is accompanied by a timestamp and encryption strength identifier.

[0068] S2: constructing a weight matrix on the first terminal device, and generating a training sample set according to the encrypted data and the weight matrix;

[0069] A weight matrix is constructed on the first terminal device. The weight matrix is an n×m-dimensional matrix, where n represents the characteristic dimension of the encrypted data and m represents the number of weight factors. The weight matrix construction process includes:

[0070] First, the encrypted data is subjected to eigendecomposition based on an improved Chebyshev polynomial, where the improved Chebyshev polynomial expression is:

[0071]

[0072] Among them, n is the order of the polynomial, the value range is [3,10], x is the input variable, which represents the normalized eigenvalue, i is the index variable, α is the adaptive coefficient, the value range is [0.1,0.5], and k main eigencomponents are extracted through the polynomial, which is recorded as vector F = [f1,f2,...,f k ];

[0073] Then, the k main feature components are assigned initial weight coefficients, which are calculated using the improved softmax function:

[0074]

[0075] Among them, w i is the initial weight value of the i-th feature, f i is the value of the i-th eigencomponent, β is the regularization parameter, |F|2 is the L2 norm of the eigenvector F, and k is the total number of eigencomponents.

[0076] Secondly, the initial weight coefficient is iteratively optimized by the improved Markov chain Monte Carlo algorithm, and the iterative update formula is:

[0077]

[0078] in, is the weight value of the i-th feature at the t-th iteration, t is the number of iterations, γ is the learning rate, λ is the decay coefficient, is the loss function, η is the noise coefficient, ∈ is a random variable obeying the standard normal distribution, and the number of iterations is set to 1000.

[0079] When the iteration converges, calculate the final weight matrix elements:

[0080]

[0081] Among them, M ijis the element value of the i-th row and j-th column in the weight matrix, w i is the optimized weight value of the i-th feature, f i ,f j ,f l are the i-th, j-th, and l-th feature components, and σ is a bandwidth parameter used to control the similarity decay rate between features. Finally, the encrypted data is divided into a training set and a validation set in a ratio of 80:20.

[0082] S3: Transmitting the training sample set to a second terminal device, where the second terminal device performs a fitting operation on the training sample set based on a deep neural network model and outputs a feature vector;

[0083] The training sample set is transmitted to the second terminal device via a secure channel. The transmission process adopts a secure transmission mechanism based on the quantum key distribution protocol, which includes four stages: quantum state preparation, quantum state transmission, measurement basis selection, and key extraction. In the quantum state preparation stage, the first terminal device encodes the training sample set into a quantum bit sequence, and each quantum bit is randomly selected to be encoded on the basis vector {|0>,|1>} or the diagonal basis vector {|+>,|->}. The quantum state transmission adopts an optical fiber quantum channel, and the mathematical expression of its secure transmission process is:

[0084] After receiving the training sample set, the second terminal device first performs integrity verification and decryption operations, and then builds an improved deep neural network model. The model adopts a multi-layer cascade architecture, including a coding layer, a feature extraction layer, and a mapping layer. The coding layer adopts an improved autoencoder structure, and its calculation expression is:

[0085]

[0086] Among them, h1 is the output feature of the encoding layer, n is the input feature dimension, is the first layer weight matrix, σ is the activation function, x i is the i-th component of the input feature vector, and φ(x) is the improved LeakyReLU function:

[0087] φ(x)=max(0.01x,x)+δsin(ωx)

[0088] The activation function enhances the nonlinear feature extraction capability of the model by introducing a periodic modulation term, where δ and ω are the amplitude and frequency parameters, respectively, which are optimized by the grid search method.

[0089] The feature extraction layer adopts an architecture that combines multi-head attention mechanism and deep residual learning, using adaptive attention weights and dynamic residual connections. Its calculation expression is:

[0090] h2=Attention(h1)+ResBlock(h1)

[0091] in,

[0092]

[0093] Q=h1W Q ,K=h1W K ,V=h1W V

[0094] Among them, M mask is a dynamically generated attention mask matrix used to suppress the interference of irrelevant features, d k is the scaling factor, whose value is dynamically adjusted through the adaptive algorithm; W Q 、W K 、W V It is the transformation matrix of query, key and value, optimized by back-propagation, Q is the query matrix, K is the key matrix, and V is the value matrix.

[0095] The mapping layer designs an innovative dynamic convolution kernel mechanism that can adaptively adjust the convolution kernel parameters to adapt to different feature patterns:

[0096]

[0097] Among them, m is the number of convolution kernels, is the jth feature map, K j (t) is a time-varying convolution kernel whose parameters are dynamically generated by a gated recurrent unit (GRU). * denotes a depth-wise separable convolution operation, which can significantly reduce computational complexity. The final feature vector is generated using a multi-scale feature fusion strategy:

[0098] F out =MaxPool(h3)+α F GlobalAvgPool(h3)

[0099] Among them, MaxPool is the maximum pooling operation, GlobalAvgPool is the global average pooling operation, α F is an adaptive weight coefficient, which is dynamically adjusted through the following mechanism:

[0100]

[0101] Among them, λ is the temperature parameter, which is dynamically adjusted during the training process, and |h3|2 is the L2 norm of the feature vector.

[0102] S4: establishing a security scoring model based on the comparison result of the feature vector and a preset threshold, wherein the security scoring model corresponds to multiple security levels;

[0103] First, based on the original feature vector F out Directly build three core security indicators:

[0104] The threat level index S1 adopts a weighted combination method:

[0105]

[0106] Among them, F out,i is the original value of the i-th feature component, λ1 is the feature interaction strength coefficient, ranging from [0.1, 0.5], k1 is the number of features used for threat calculation, ω i is the dynamic weight coefficient of the i-th feature, which is dynamically calculated through the attention mechanism:

[0107]

[0108] in, is the weight vector of the attention layer, with a dimension of d×1, W b is the feature transformation matrix with dimension d×n, b is the bias vector, and tanh is the hyperbolic tangent activation function.

[0109] The vulnerability index S2 uses nonlinear mapping:

[0110]

[0111] Among them, k2 is the number of features used for vulnerability calculation, α i is the influencing factor of the i-th feature, ranging from [0,1], μ i is the characteristic sensitivity coefficient, which controls the steepness of the nonlinear mapping, and β is the periodic modulation coefficient, which is used to introduce nonlinear fluctuations.

[0112] The risk exposure indicator S3 is expressed in quadratic form:

[0113]

[0114] Among them, Q S is the feature correlation matrix, a semi-positive definite matrix, γ S is the L2 regularization coefficient to prevent overfitting, |F out |2 is the L2 norm of the eigenvector, is the transpose of the eigenvector

[0115] Based on the above three indicators, a comprehensive safety score is constructed:

[0116] S total =φ1S1+φ2S2+φ3S3+ξ(S1S2+S2S3+S3S1)

[0117] Among them, φ1, φ2, φ3 are balance coefficients, and ξ is the interaction strength parameter, which is dynamically updated by the following equation:

[0118]

[0119] Where t is the number of iterations, η φ is the learning rate, which controls the parameter update step size, λ φ is the time attenuation coefficient, is the gradient of the loss function with respect to the weights.

[0120] According to the comprehensive security score, a seven-level security classification mechanism is established: Very high security level (Level 7): S total ≥θ1, Features: The system is completely trustworthy, without any security risks, high security level (Level 6): θ2≤S total <θ1, Features: Highly reliable system, few security risks, high security level (Level 5): θ3≤S total <θ2, characteristics: the system is trustworthy, there are a few low-risk hidden dangers, medium security level (Level 4): θ4≤S total <θ3, characteristics: the system is basically trustworthy, there is a medium risk, low security level (Level 3): θ5≤S total <θ4, characteristic: the system has obvious safety risks, low security level (Level 2): θ6≤S total <θ5 characteristic: There is a serious security threat to the system. Recommendation: Initiate emergency response. Very low security level (Level 1): S total <θ6, characteristic: the system is in a high-risk state.

[0121] Threshold parameter θ i Updated by dynamic optimization algorithm:

[0122]

[0123] Among them, L level is the multi-classification loss function:

[0124]

[0125] Among them, p i and q i are the probability distributions of the true level and the predicted level, respectively, ρ is the interval constraint coefficient, and η is the learning rate parameter.

[0126] S5: Select a corresponding operation strategy according to the security level, and perform differential privacy protection processing on the encrypted data.

[0127] For Level 7:

[0128] A lightweight differential privacy strategy is employed, primarily perturbing the data by adding small amounts of Laplace noise. Because this level of security indicates extremely high system security, only the most basic privacy protection is required to ensure high data availability. Specifically, a small amount of random noise is superimposed on the encrypted data. The noise intensity is strictly controlled to a low level using a privacy budget parameter, while maintaining the statistical properties of the data largely unchanged.

[0129] For high security level (Level 6):

[0130] A periodic perturbation term is added to the basic Laplace noise. This strategy maintains high data availability while increasing the difficulty for attackers to predict by introducing periodic variations. The specific operation involves first calculating the basic noise amount and then superimposing a periodic function with adjustable amplitude. This makes the final protection effect time-varying, increasing the data's protection strength.

[0131] For higher security level (Level 5):

[0132] A hybrid mechanism using Laplace and Gaussian noise achieves more flexible privacy protection by dynamically adjusting the mixing ratio of the two noises. This hybrid strategy retains the strong privacy guarantees of the Laplace mechanism while leveraging Gaussian noise to provide better data availability. The system adaptively adjusts the mixing weights based on real-time security monitoring to achieve a good balance between privacy protection and data availability.

[0133] For medium security level (Level 4):

[0134] Implementing an adaptive Gaussian noise protection mechanism, the noise intensity is dynamically adjusted based on data characteristics and time. Specifically, the system first analyzes the characteristic distribution of the encrypted data and then designs appropriate Gaussian noise parameters. The noise intensity varies moderately over time, ensuring adequate privacy protection while maintaining good data availability.

[0135] For lower security level (Level 3):

[0136] A multi-dimensional composite perturbation mechanism is implemented, combining multiple noise types for combined protection. This strategy combines the advantages of various noise types and, through an intelligent weighting mechanism, dynamically adjusts the proportion of each noise type based on data characteristics. The system monitors protection effectiveness in real time and continuously optimizes noise parameters through a feedback mechanism to ensure strong privacy protection.

[0137] For low security level (Level 2):

[0138] Deploy a deep differential privacy protection mechanism, using a deep learning model to perform secondary processing on encrypted data. This strategy first adds basic noise and then transforms features using a trained neural network model. This significantly improves privacy protection while maintaining the necessary usability. The system continuously evaluates protection effectiveness and adjusts model parameters as needed.

[0139] For very low security level (Level 1):

[0140] We implement the strictest differential privacy protection strategy, employing a multi-layer cascaded protection mechanism. First, we add strong composite noise to the data, then transform it using a feature encoder, and finally process it using a specialized protection model. This multi-layered protection ensures strong privacy protection even in the most insecure situations.

[0141] This embodiment also provides a security computing system based on artificial intelligence, including:

[0142] The collection and encryption module is used to collect the original data generated by the first terminal device and encrypt the original data using a preset encryption algorithm to obtain encrypted data;

[0143] A training module, configured to construct a weight matrix on the first terminal device and generate a training sample set based on the encrypted data and the weight matrix;

[0144] a feature vector construction module, configured to transmit the training sample set to a second terminal device, wherein the second terminal device performs a fitting operation on the training sample set based on a deep neural network model and outputs a feature vector;

[0145] A security grading module, configured to establish a security scoring model based on a comparison result between the feature vector and a preset threshold, wherein the security scoring model corresponds to multiple security levels;

[0146] The differential operation module is used to establish a security scoring model based on the comparison result of the feature vector and the preset threshold, and the security scoring model corresponds to multiple security levels.

[0147] This embodiment also provides a computer device suitable for the case of a secure computing method based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the secure computing method based on artificial intelligence proposed in the above embodiment.

[0148] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0149] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the secure computing method based on artificial intelligence proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0150] Example 2

[0151] This is the second embodiment of the present invention. In order to further verify the advanced nature of the present invention, experimental simulation of a security computing method based on artificial intelligence and comparative data with the prior art are provided.

[0152] This example uses a 30-day test environment with a cluster of terminal devices in a large enterprise data center. The test environment includes 100 terminal devices distributed across various business departments, representing representative load characteristics and security risk profiles. Before the experiment begins, a basic environment configuration is performed on all terminal devices, including the installation of the data acquisition module, encryption processing module, and security assessment module.

[0153] During the data collection phase, the system continuously collects raw data generated by the device through a collection agent deployed on the terminal device. The collection frequency is set to once every 5 minutes to ensure the real-time and continuity of the data. The collected data includes user operation logs (on average, each device generates about 5,000 records per day), system status data (CPU usage, memory occupancy, etc., with a sampling interval of 1 minute), network transmission data (including upstream and downstream traffic, data packet size, etc.) and device performance indicators. The following experimental data was obtained according to the steps of Example 1:

[0154] Table 1 Experimental data table

[0155]

[0156] Analysis of experimental data reveals that this security assessment and differential privacy protection system demonstrates significant performance advantages and protection effectiveness. First, the initial risk score (Initial_Risk_Score) shows that the system accurately identifies the security risk levels of different devices, ranging from 0.21 to 0.82, demonstrating strong risk identification capabilities. In particular, the system assigned a high risk score of 0.82 to the high-risk device (Device001). This precise risk quantification provides a reliable basis for subsequent protection strategy selection.

[0157] After differential privacy protection processing, the risk scores of all devices were significantly reduced. Taking Device001 as an example, its risk score dropped from 0.82 to 0.15, a decrease of 81.7%, while maintaining 95% data availability (Data_Utility_Rate). This result verifies that the system has achieved an excellent balance between protection strength and data availability. It is particularly noteworthy that the system adopts differentiated protection strategies for devices of different risk levels, which is reflected in the dynamic adjustment of the privacy budget (Privacy_Budget), which gradually decreases from 0.1 at Level 7 to 0.01 at Level 1, ensuring that the protection strength matches the risk level.

[0158] In terms of processing performance, the system demonstrates excellent real-time performance. Processing time (Processing_Time_ms) increases moderately with increasing protection intensity, but even for Level 1 devices employing the strictest protection strategy, processing time is only 85 milliseconds, fully meeting real-time processing requirements. Furthermore, the system's protection success rate (Success_Rate) generally remains above 90%, with Level 7 devices achieving an exceptionally high success rate of 99.5%, demonstrating the solution's stability and reliability.

[0159] Data analysis also reveals a key characteristic: data availability and protection strength exhibit a reasonable decreasing relationship, but the rate of decay is well controlled within an acceptable range. Even for Level 1 devices using the strictest protection strategy, data availability remains at 78%, significantly outperforming traditional differential privacy schemes (typically in the 50%-60% range). This advantage is primarily due to the adaptive protection mechanism and multi-layer cascade protection strategy employed in this scheme, which effectively mitigates the negative impact of protection strength on data availability.

[0160] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A secure computing method based on artificial intelligence, characterized by: include: Collecting original data generated by the first terminal device, and encrypting the original data using a preset encryption algorithm to obtain encrypted data; constructing a weight matrix on the first terminal device, and generating a training sample set according to the encrypted data and the weight matrix; Transmitting the training sample set to a second terminal device, where the second terminal device performs a fitting operation on the training sample set based on a deep neural network model and outputs a feature vector; Establishing a security scoring model based on the comparison result of the feature vector and a preset threshold, wherein the security scoring model corresponds to multiple security levels; A corresponding operation strategy is selected according to the security level, and differential privacy protection processing is performed on the encrypted data.

2. The artificial intelligence-based secure computing method according to claim 1, wherein: The constructing of the weight matrix comprises: Performing feature decomposition on the encrypted data to extract k main feature components; Assigning initial weight coefficients to the k main feature components; Iteratively optimizing the initial weight coefficients; When the iteration converges, calculate the final weight matrix elements; The eigendecomposition is expressed as follows: Among them, T n (x) is the improved Chebyshev polynomial, n is the order of the polynomial, and its value range is [3,10]. x is the input variable, which represents the normalized eigenvalue. i is the index variable, and α is the adaptive coefficient.

3. The artificial intelligence-based secure computing method according to claim 2, wherein: The initial weight coefficient is calculated by the following formula: Among them, w i is the initial weight value of the i-th feature, f i is the value of the i-th eigencomponent, β is the regularization parameter, |F|2 is the L2 norm of the eigenvector F, and k is the total number of eigencomponents; The iterative optimization of the initial weight coefficient is expressed by the following formula: in, is the weight value of the i-th feature at the t-th iteration, t is the number of iterations, γ is the learning rate, λ is the decay coefficient, is the loss function, η is the noise coefficient, and ∈ is a random variable that obeys the standard normal distribution.

4. The artificial intelligence-based secure computing method according to claim 3, wherein: The deep neural network model includes an encoding layer, a feature extraction layer and a mapping layer; The autoencoder structure of the encoding layer is shown below: Among them, h1 is the output feature of the encoding layer, n is the input feature dimension, is the first layer weight matrix, σ is the activation function, x i is the i-th component of the input feature vector, and φ(x) is the improved LeakyReLU function: φ(x)=max(0.01x,x)+δsin(ωx) Here, δ and ω are the amplitude and frequency parameters, respectively.

5. The artificial intelligence-based secure computing method according to claim 4, wherein: The safety scoring model is shown below: S total =φ1S1+φ2S2+φ3S3+ξ(S1S2+S2S3+S3S1) Among them, φ1, φ2, φ3 are balance coefficients, and ξ is the interaction strength parameter, which is dynamically updated by the following equation: Where t is the number of iterations, η φ is the learning rate, which controls the parameter update step size, λ φ is the time attenuation coefficient, is the gradient of the loss function with respect to the weight, S1 is the threat level index, S2 is the vulnerability index, and S3 is the risk exposure index.

6. The artificial intelligence-based secure computing method according to claim 5, wherein: The threat level index is as follows: Among them, F out,i is the original value of the i-th feature component, λ1 is the feature interaction strength coefficient, ranging from [0.1, 0.5], k1 is the number of features used for threat calculation, ω i is the dynamic weight coefficient of the i-th feature, which is dynamically calculated through the attention mechanism: in, is the weight vector of the attention layer, with a dimension of d×1, W b is the feature transformation matrix with dimension d×n, b is the bias vector, and tanh is the hyperbolic tangent activation function.

7. The artificial intelligence-based secure computing method according to claim 6, wherein: The security level is determined according to a dynamic threshold, with a total of seven security levels; The dynamic threshold is calculated as follows: Among them, L level is the multi-classification loss function: Among them, p i and q i are the probability distributions of the true level and the predicted level, respectively, ρ is the interval constraint coefficient, is the current value of the i-th threshold parameter at time t, and η is the learning rate parameter.

8. An artificial intelligence-based secure computing system, based on the artificial intelligence-based secure computing method according to any one of claims 1 to 7, characterized in that: include: The collection and encryption module is used to collect the original data generated by the first terminal device and encrypt the original data using a preset encryption algorithm to obtain encrypted data; A training module, configured to construct a weight matrix on the first terminal device and generate a training sample set based on the encrypted data and the weight matrix; a feature vector construction module, configured to transmit the training sample set to a second terminal device, wherein the second terminal device performs a fitting operation on the training sample set based on a deep neural network model and outputs a feature vector; A security grading module, configured to establish a security scoring model based on a comparison result between the feature vector and a preset threshold, wherein the security scoring model corresponds to multiple security levels; The differential operation module is used to establish a security scoring model based on the comparison result of the feature vector and the preset threshold, and the security scoring model corresponds to multiple security levels.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the artificial intelligence-based security computing method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the artificial intelligence-based security computing method according to any one of claims 1 to 7 are implemented.

Images